An AI-based method for early warning of marine disasters

Abstract

This invention provides an artificial intelligence-based method for early warning of marine disasters, belonging to the field of marine disaster early warning technology. This invention performs hierarchical normalization of marine dynamic parameters by constructing a four-level time-length sequence and a four-level spatial scale grid. It extracts scale-separated feature vector groups using an ultra-sparse cross-scale coupling matrix and a micro-feature temporal dependency matrix, maps the features to a high-dimensional feature space based on a radial basis function kernel function, and outputs a prediction sequence using a multi-scale coupling model of marine disasters employing a sparse attention mechanism and a continuous depth modeling algorithm. The prediction uncertainty interval is calculated through integrated forecast processing. When the uncertainty exceeds a threshold, the observation density is increased and iterative optimization is performed. This solves the technical problem of insufficient prediction accuracy caused by the inability to simultaneously handle multi-temporal-scale coupling processes in marine disaster early warning.

Description

Technical Field

[0001] This invention belongs to the field of marine disaster early warning technology, and more specifically, relates to a marine disaster early warning method based on artificial intelligence. Background Technology

[0002] Marine disaster early warning technologies typically employ numerical model simulations or statistical regression methods to analyze and predict ocean dynamic parameters. Traditional numerical models calculate the spatiotemporal evolution of parameters such as sea surface temperature, ocean current velocity, and wave height on a fixed spatial grid by solving ocean dynamic equations. Statistical regression methods establish empirical relationships between ocean dynamic parameters and the probability of disaster occurrence based on historical observation data. However, since ocean dynamic processes are inherently complex systems coupled across multiple spatiotemporal scales, small-scale processes transfer energy and matter to larger scales through nonlinear interactions, while the large-scale background field provides boundary conditions and forcing effects on the occurrence and development of small-scale processes. Traditional early warning methods lack effective cross-scale information transfer and feature fusion mechanisms. When dealing with multi-scale coupled processes, they either omit information at key scales or mix features from different scales, resulting in unclear physical meaning of the prediction model. In other words, existing technologies suffer from the technical problem of insufficient prediction accuracy due to the inability to simultaneously handle multi-spatiotemporal coupled processes in marine disaster early warning systems. Summary of the Invention

[0003] In view of this, the present invention provides an artificial intelligence-based marine disaster early warning method, which can solve the technical problem in the prior art that marine disaster early warning cannot simultaneously handle multi-temporal scale coupling processes, resulting in insufficient prediction accuracy.

[0004] This invention is implemented as follows: This invention provides a marine disaster early warning method based on artificial intelligence, comprising the following steps:

[0005] Construct a four-level time length series and a four-level spatial scale grid, and set the time length adjustment factor and the scale adjustment factor;

[0006] Marine observation sensors are deployed at grid nodes to collect marine dynamic parameters, and time and spatial normalization processing is performed.

[0007] By calling the ultra-sparse cross-scale coupling matrix and the micro-feature time-series dependency matrix, the normalized marine dynamic parameters are input into the multi-scale coupling model of marine disasters for cross-scale feature extraction, resulting in a scale-separated feature vector group.

[0008] Calculate the rank of the ultra-sparse cross-scale coupling matrix and perform scale reduction or scale expansion. Map the scale-separated feature vector group to a high-dimensional feature space through the radial basis kernel function.

[0009] The feature vector is input into the multi-scale coupling model of marine disasters. The sparsity parameter of the sparse attention mechanism is determined according to the duration adjustment factor, the scale adjustment factor and the proportion of non-zero elements of the ultra-sparse cross-scale coupling matrix, and the marine disaster prediction sequence is output.

[0010] The marine disaster prediction sequence is integrated and forecasted, and the prediction uncertainty interval is calculated. When the width of the prediction uncertainty interval exceeds a preset threshold, the observation density is increased and the above steps are repeated. When the width of the prediction uncertainty interval is less than the preset threshold, marine disaster early warning information is generated and output.

[0011] Preferably, the establishment of the four-level time length series specifically sets the time length range of the hourly short-term series to 1h to 24h, the daily medium-term series to 24h to 168h, the weekly long-term series to 168h to 720h, and the monthly ultra-long-term series to 720h to 2160h.

[0012] Preferably, time sampling points are set at logarithmic intervals within each time length range. The logarithmic interval is obtained by dividing the upper limit of the time length range by the lower limit, taking the natural logarithm, and then dividing by the number of sampling points minus 1.

[0013] Preferably, the establishment of the four-level spatial scale grid is specifically defined as follows: the grid size of the 100-meter micro-scale grid is set to 100m, the kilometer-scale meso-scale grid is set to 1000m, the ten-kilometer-scale large-scale grid is set to 10000m, and the hundred-kilometer-scale giant-scale grid is set to 100000m.

[0014] Preferably, micro-scale grids at the 100-meter level cover key near-shore protected areas, meso-scale grids at the kilometer level cover the coastal economic belt, large-scale grids at the 10-kilometer level cover the continental shelf area, and mega-scale grids at the 100-kilometer level cover the deep sea and open ocean areas.

[0015] Preferably, the determination of the duration adjustment factor specifically involves setting the base time length as 1 hour for hourly short-term sequences, 24 hours for daily medium-term sequences, 168 hours for weekly long-term sequences, and 720 hours for monthly ultra-long-term sequences.

[0016] Preferably, the target predicted time length is obtained, the ratio of the target predicted time length to the baseline time length of each sequence is calculated as the original time length ratio, the original time length ratio is logarithmically transformed and mapped to the interval [0, 1] through linear transformation to obtain the time length adjustment factor.

[0017] Preferably, the determination of the scale adjustment factor specifically involves setting the reference grid size to 100m for the 100-meter micro-scale grid, 1000m for the kilometer-scale meso-scale grid, 10000m for the 10-kilometer-scale large-scale grid, and 100000m for the 100-kilometer-scale giant-scale grid.

[0018] Preferably, the ratio of the actual grid size to the reference grid size of each level of grid is calculated as the original scale ratio. The original scale ratio is then logarithmically transformed and mapped to the interval [0, 1] through a linear transformation to obtain the scale adjustment factor.

[0019] Preferably, the establishment of the ultra-sparse cross-scale coupling matrix is ​​specifically defined as follows: for grid nodes at adjacent scales, when the spatial distance between two grid nodes is less than 50% of the size of the larger scale grid, the two grid nodes are considered to have a coupling relationship.

[0020] The beneficial effects of this invention are:

[0021] This invention constructs a four-level time-scale sequence, enabling the model to extract corresponding time-scale features for hourly-level sudden events and monthly-level long-term trends. The four-level spatial-scale grid, through a nested structure, provides high-resolution observations in key areas while maintaining computational efficiency over a large area. The ultra-sparse cross-scale coupling matrix retains only the inter-scale coupling relationships with significant physical influence, avoiding redundant information from interfering with the prediction process. The micro-feature temporal dependency matrix captures weak correlation patterns between different time steps, providing a basis for early signal identification of marine disasters. The continuous depth modeling algorithm achieves feature evolution at arbitrary depths through an ordinary differential equation solver, overcoming the gradient vanishing problem in traditional discrete layer networks and enabling the model to capture long-range dependencies in long-term sequences.

[0022] In summary, this invention solves the technical problem mentioned in the background art of insufficient prediction accuracy caused by the inability of marine disaster early warning to simultaneously handle multi-temporal scale coupling processes. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention.

[0024] Figure 2 This is a distribution diagram of kernel width parameters for grids of different spatial scales in the embodiment.

[0025] Figure 3 This is a cloud map showing the distribution of seafloor pressure anomalies at a four-level spatial scale grid 12 hours before the earthquake in the example.

[0026] Figure 4 This is a sequence diagram of marine disaster prediction using a four-level time-length sequence as shown in the example. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0028] like Figure 1The diagram shown is a flowchart of a marine disaster early warning method based on artificial intelligence provided by the present invention. This method includes the following steps:

[0029] S01. Construct a four-level time-length sequence and a four-level spatial scale grid, and set duration adjustment factors and scale adjustment factors for the four-level time-length sequence and the four-level spatial scale grid respectively; wherein, the four-level time-length sequence includes hourly short-term sequence, day-level medium-term sequence, weekly long-term sequence and monthly ultra-long-term sequence, and the four-level spatial scale grid includes hundred-meter micro-scale grid, kilometer-level meso-scale grid, ten-kilometer-level large-scale grid and hundred-kilometer-level giant-scale grid;

[0030] S02. Deploy marine observation sensors at each grid node of the four-level spatial scale grid to collect marine dynamic parameters, including sea surface temperature, ocean current velocity, wave height, salinity, and seabed pressure. Then, perform time normalization processing according to the duration adjustment factor and spatial normalization processing according to the scale adjustment factor on the collected marine dynamic parameters.

[0031] S03. Call the ultra-sparse cross-scale coupling matrix and the micro-feature time-series dependency matrix, input the normalized marine dynamic parameters into the marine disaster multi-scale coupling model for cross-scale feature extraction, and obtain the scale separation feature vector group corresponding to the four-level spatial scale grid.

[0032] S04. Calculate the rank of the ultra-sparse cross-scale coupling matrix. When the rank is less than 30% of the total number of nodes in the fourth-level spatial scale grid, perform scale reduction. When the rank is greater than 85% of the total number of nodes in the fourth-level spatial scale grid, perform scale expansion. Map the scale-separated feature vector group to the high-dimensional feature space through the radial basis kernel function.

[0033] S05. Input the feature vector in the high-dimensional feature space into the marine disaster multi-scale coupling model. The sparsity parameter of the sparse attention mechanism in the marine disaster multi-scale coupling model is determined according to the proportion of non-zero elements of the duration adjustment factor, the scale adjustment factor and the ultra-sparse cross-scale coupling matrix. Output the marine disaster prediction sequence corresponding to the four-level time length sequence.

[0034] S06. Perform integrated forecasting processing on the marine disaster prediction sequence, calculate the prediction uncertainty interval, increase the observation density and re-execute steps S02 to S05 when the width of the prediction uncertainty interval exceeds the preset threshold, and generate and output marine disaster early warning information when the width of the prediction uncertainty interval is less than the preset threshold.

[0035] The steps for establishing the four-level time length sequence specifically include:

[0036] The time range for hourly short-term series is set to 1 hour to 24 hours, the time range for daily medium-term series is set to 24 hours to 168 hours, the time range for weekly long-term series is set to 168 hours to 720 hours, and the time range for monthly ultra-long-term series is set to 720 hours to 2160 hours.

[0037] Time sampling points are set at logarithmic intervals within each time length range of the four-level time length sequence. The logarithmic interval is obtained by dividing the upper limit of the time length range by the lower limit of the time length range, taking the natural logarithm, and then dividing by the number of sampling points minus 1.

[0038] Arrange all time sampling points in the four-level time length sequence in chronological order to form a complete time length sequence.

[0039] The occurrence and development of marine disasters span multiple timescales, from hourly local storms to monthly large-scale ocean current anomalies. Single-timescale prediction models cannot simultaneously capture both short-term emergencies and long-term evolution trends. The four-level time-length series, through hierarchical timescale division, allows the model to employ corresponding feature extraction strategies for dynamic processes at different timescales. Hourly short-term series focus on capturing rapidly changing local disturbances and precursory signals of sudden disasters; day-long series focus on the evolution of ocean mesoscale eddies and frontal systems; weekly long-term series track the propagation characteristics of intraseasonal oscillations and planetary waves; and monthly ultra-long-term series reveal the slow movement of ocean circulation and water mass structure. The system adjusts trends and employs a hierarchical timescale design to avoid feature coupling issues caused by mixing information from all timescales together. The logarithmic interval time sampling point setting ensures that computational redundancy is not generated due to excessive sampling points on longer timescales, while maintaining sufficient time resolution to capture transient changes on shorter timescales. When used in conjunction with the duration adjustment factor, the four-level time length series can automatically adjust the weight of each series according to the timescale of the prediction target. When short-term early warning is needed, the weight of hourly short-term series is increased, and when long-term trend prediction is needed, the weight of monthly ultra-long-term series is increased, thereby achieving adaptive prediction capabilities on the timescale.

[0040] The steps for establishing the fourth-level spatial scale grid specifically include:

[0041] The grid size is set to 100m for the micro-scale grid at the hundred-meter level, 1000m for the meso-scale grid at the kilometer level, 10000m for the large-scale grid at the ten-kilometer level, and 100000m for the giant-scale grid at the hundred-kilometer level.

[0042] The coverage areas of each grid level are determined based on the geographical scope and coastline shape of the warning sea area. The 100-meter-level micro-scale grid covers the key near-shore protection areas, the kilometer-level meso-scale grid covers the coastal economic belt, the 10-kilometer-level large-scale grid covers the continental shelf area, and the 100-kilometer-level mega-scale grid covers the deep sea and open ocean areas.

[0043] Within the coverage area of ​​each grid level, grid cells are divided according to the grid size and grid node coordinates are marked; a nesting relationship between adjacent scale levels is established so that the node coordinates of the finer grid and the node coordinates of the coarser grid are spatially contained or coincident.

[0044] Marine disasters span spatial scales from meter-level wave breaking to kilometer-level ocean basin circulation. Single-resolution grids are insufficient to precisely depict localized small-scale processes, while using high-resolution grids over large areas leads to excessive computational resource consumption. A four-level spatial scale grid system, through a nested multi-scale grid architecture, provides high-resolution micro-scale grids in key areas to capture the impact of complex nearshore topography and local turbulence, while low-resolution giant-scale grids in secondary areas describe the evolution of the large-scale background field. The nested grid relationship facilitates natural information transfer and boundary condition setting between different scales; the micro-scale grid obtains boundary forcing information from the mesoscale grid, and the mesoscale grid... The statistical effects of local processes are fed back to the large-scale grid, forming a two-way coupling channel from bottom to top and from top to bottom. When the four-level spatial scale grid is used in conjunction with the scale adjustment factor, the contribution weight of each level of grid can be dynamically adjusted according to the degree of spatial heterogeneity of the current ocean state. When the ocean state is highly heterogeneous in space, the weight of the micro-scale and meso-scale grids is increased, and when the ocean state is relatively homogeneous, the weight of the large-scale and mega-scale grids is increased. The multi-scale grid system also provides a framework for the parameterization of cross-scale physical processes. The influence of sub-grid scale processes is expressed by the information projection from the high-resolution grid to the low-resolution grid, avoiding the limitations of empirical formulas in traditional parameterization schemes.

[0045] The steps for determining the duration adjustment factor specifically include:

[0046] For the four-level time-length series, a baseline time length of 1 hour is set for the hourly short-term series, 24 hours for the daily medium-term series, 168 hours for the weekly long-term series, and 720 hours for the monthly ultra-long-term series. The target prediction time length for the current prediction task is obtained, and the ratio of the target prediction time length to the baseline time length of each series in the four-level time-length series is calculated as the original duration ratio. A base-10 logarithmic transformation is applied to the original duration ratio to obtain the logarithmic duration ratio. This logarithmic duration ratio is then mapped to the interval [0, 1] through a linear transformation to obtain the duration adjustment factor. The duration adjustment factor is used to balance the prediction weights at different time scales, preventing the rapid accumulation of errors in long-term predictions from causing prediction failure.

[0047] The steps for determining the scale adjustment factor specifically include: setting the reference grid size to 100m for the 100-meter-level micro-scale grid, 1000m for the kilometer-level meso-scale grid, 10000m for the 10-kilometer-level large-scale grid, and 100000m for the 100-kilometer-level mega-scale grid; calculating the ratio of the actual grid size to the reference grid size for each level of the four-level spatial scale grid as the original scale ratio; performing a base-10 logarithmic transformation on the original scale ratio to obtain the logarithmic scale ratio; and mapping the logarithmic scale ratio to the interval [0, 1] through a linear transformation to obtain the scale adjustment factor. The scale adjustment factor is used to coordinate the information transmission intensity between different spatial scales, ensuring that the coupling relationship between microscale turbulence and megascale circulation is correctly expressed.

[0048] The steps for establishing the ultra-sparse cross-scale coupling matrix specifically include:

[0049] Establish spatial adjacency relationships between grids at four spatial scales. For grid nodes at adjacent scales, if the spatial distance between two grid nodes is less than 50% of the size of the larger scale grid, the two grid nodes are considered to have a coupling relationship.

[0050] An initial coupling matrix is ​​established based on the coupling relationship, wherein the number of rows and columns of the initial coupling matrix are equal to the total number of nodes of the fourth-level spatial scale grid;

[0051] The physical influence strength between each pair of coupled nodes is calculated. The physical influence strength is represented by the weighted average of the ocean current velocity gradient, sea surface temperature gradient, and salinity gradient between the coupled nodes. The weight of the ocean current velocity gradient is 0.5, the weight of the sea surface temperature gradient is 0.3, and the weight of the salinity gradient is 0.2. For coupling relationships whose physical influence strength is less than 5% of the average physical influence strength of all coupled node pairs, the corresponding matrix elements in the initial coupling matrix are set to zero to form an ultra-sparse cross-scale coupling matrix.

[0052] The proportion of non-zero elements in an ultra-sparse cross-scale coupling matrix typically falls within the range of [2%, 8%]. Each non-zero element in the matrix represents a significant cross-scale dynamic coupling between corresponding grid node pairs. The ultra-sparse nature of the matrix improves the computational efficiency of cross-scale information transfer while preserving key inter-scale interaction channels.

[0053] The steps for establishing the micro-feature temporal dependency matrix specifically include:

[0054] For each time step in the four-level time length series, small variation features of ocean dynamic parameters are extracted. These small variation features are represented by the first-order and second-order differences of ocean dynamic parameters between adjacent time steps.

[0055] The correlation of minute changes between different time steps is calculated, and the correlation is obtained by Pearson correlation coefficient.

[0056] A time-series dependency matrix is ​​established, wherein the number of rows and columns of the time-series dependency matrix are equal to the total number of time steps of the four-level time length series, and the matrix elements of the time-series dependency matrix are the Pearson correlation coefficients between corresponding time step pairs;

[0057] For time-dependent matrices where the absolute value of the Pearson correlation coefficient is less than 0.15, set the elements to zero to form micro-feature time-dependent matrices.

[0058] The micro-feature temporal dependency matrix captures the weak correlation patterns of ocean dynamic processes at different time scales. Weak correlations often indicate early signals of ocean disasters. The sparsity processing of the matrix filters out the interference of random noise and retains the temporal dependencies with predictive value.

[0059] The time normalization process specifically includes the following steps:

[0060] The collected ocean dynamic parameters are weighted according to a duration adjustment factor, wherein the weighting method is to multiply the current value of the ocean dynamic parameter by the duration adjustment factor of the corresponding time scale;

[0061] Subtracting the mean of the time series from the weighted ocean dynamic parameters and dividing by the standard deviation of the time series yields the time-normalized ocean dynamic parameters.

[0062] The spatial normalization process specifically includes the following steps:

[0063] The time-normalized ocean dynamic parameters are weighted according to a scale adjustment factor, wherein the weighting method is to multiply the time-normalized ocean dynamic parameters by the scale adjustment factor of the corresponding spatial scale; the weighted ocean dynamic parameters are then subtracted from the mean of the spatial grid and divided by the standard deviation of the spatial grid to obtain the normalized ocean dynamic parameters.

[0064] The specific steps of the scale reduction process include:

[0065] Singular value decomposition is performed on the ultra-sparse cross-scale coupling matrix, and the singular vectors corresponding to the first few largest singular values ​​are extracted as the main scale patterns. The scale-separated feature vector group is projected onto the subspace spanned by the main scale patterns to obtain the dimension-reduced feature vectors. The dimension of the dimension-reduced feature vectors belongs to the interval [40%, 60%] of the original dimension.

[0066] The scale reduction process is triggered when the rank of the ultrasparse cross-scale coupling matrix is ​​low, indicating that the inter-scale coupling relationship of the current ocean state is relatively simple and that information at some scales is redundant. Dimension reduction reduces the computational burden and suppresses the risk of overfitting.

[0067] The steps of the scale-up dimension expansion process specifically include:

[0068] Eigenvalue decomposition is performed on the ultra-sparse cross-scale coupling matrix to identify the weak coupling patterns corresponding to eigenvectors with smaller eigenvalues.

[0069] In the scale-separated feature vector group, an auxiliary feature component corresponding to the weak coupling mode is added. The auxiliary feature component is obtained by extracting high-frequency detail coefficients after performing wavelet transform on the normalized ocean dynamic parameters.

[0070] The dimensions of the expanded feature vectors fall within the range of the original dimensions [140%, 160%].

[0071] The scale expansion process is triggered when the rank of the ultrasparse cross-scale coupling matrix is ​​high, indicating that the inter-scale coupling relationship of the current ocean state is highly active and richer feature representation is needed to accurately characterize the cross-scale dynamic process.

[0072] The calculation steps for the radial basis kernel function specifically include:

[0073] The Euclidean basis function kernel function calculates the Euclidean distance between each feature vector in the scale-separated feature vector group and a preset kernel center vector. This Euclidean distance is divided by the kernel width parameter, negativeed, and then exponentially calculated to obtain the kernel function mapping value. The kernel center vector is selected from the feature vectors of historical marine disaster events using a K-means clustering algorithm. The kernel width parameter is equal to the reciprocal of the scale adjustment factor multiplied by an adjustment coefficient, where the adjustment coefficient is 0.1. The radial basis function maps the linearly inseparable scale-separated feature vector group into a high-dimensional feature space, making it linearly separable. This allows the multi-scale coupling model of marine disasters to learn more complex nonlinear scale coupling relationships. The kernel width parameter adaptively adjusts with changes in the scale adjustment factor, ensuring a reasonable distribution of features at different spatial scales in the high-dimensional feature space.

[0074] The specific structure of the multi-scale coupled marine disaster model is as follows:

[0075] The model consists of an input layer, a multi-scale feature extraction layer, a cross-scale fusion layer, and a prediction output layer;

[0076] The input layer receives the normalized ocean dynamic parameters and embeds them into the initial feature space;

[0077] The multi-scale feature extraction layer contains four parallel feature extraction branches, each corresponding to a scale level in a four-level spatial scale grid. Each feature extraction branch uses a residual connection structure to extract feature representations at the corresponding scale.

[0078] The cross-scale fusion layer calculates the correlation weights between features at different scales through a sparse attention mechanism. The sparse attention mechanism calculates attention weights only for scale pairs corresponding to non-zero elements in the ultra-sparse cross-scale coupling matrix and adjusts the attention distribution in the temporal dimension according to the micro-feature temporal dependency matrix.

[0079] The prediction output layer maps the fused cross-scale features to the prediction target space corresponding to the four-level time length sequence, and outputs the marine disaster prediction sequence.

[0080] The model employs a continuous deep modeling algorithm based on ordinary differential equations, treating the network layers in the multi-scale feature extraction layer and cross-scale fusion layer as discretizations of a continuous-time dynamic system. In forward propagation, an ordinary differential equation solver is used to accurately calculate feature evolution at arbitrary depths and gradient backpropagation.

[0081] The specific steps of the continuous depth modeling algorithm are as follows:

[0082] The hidden states in the network are represented as functions of continuous-time variables, with the initial value of the continuous-time variables set to 0 corresponding to the input layer and the final value of the continuous-time variables set to 1 corresponding to the prediction output layer.

[0083] Define a function whose derivative of the hidden state with respect to continuous-time variables is equal to the network parameters and the current hidden state, and this function is modeled by a neural network;

[0084] The fourth-order Runge-Kutta method is used to solve the ordinary differential equations, and the hidden state of the predicted output layer is obtained by integrating from the initial value of the continuous-time variable to the final value of the continuous-time variable.

[0085] During backpropagation, an adjoint ordinary differential equation is constructed, the solution of which gives the gradient of the loss function with respect to the network parameters.

[0086] Traditional discrete deep networks have a fixed and finite number of layers, while continuous deep modeling algorithms extend the network depth from a discrete number of layers to a continuous time dimension. This allows the model to adaptively adjust the length of feature evolution based on the complexity of the input data. For systems with strong nonlinear and chaotic characteristics in ocean dynamic processes, continuous deep modeling algorithms provide more refined feature evolution paths, avoiding information abrupt changes and gradient discontinuities between discrete layers. The ordinary differential equation solver ensures the smoothness and physical rationality of the feature evolution process. Features at different spatial and temporal scales evolve at different rates in the continuous time dimension, naturally achieving decoupling of multi-scale processes. The accurate gradient calculation accompanying the ordinary differential equation method overcomes the gradient vanishing and gradient explosion problems in traditional backpropagation, enabling the model to capture long-range dependencies in long-term series, which plays a crucial role in suppressing error accumulation in long-term ocean disaster prediction. Continuous deep modeling algorithms also significantly improve the model's memory efficiency because it does not need to store the intermediate activation values ​​of each layer. Instead, the intermediate activation values ​​are recalculated on demand by the ordinary differential equation solver, making it possible to train deeper and more complex cross-scale coupled networks.

[0087] The specific steps for establishing the training dataset for the multi-scale coupled marine disaster model include:

[0088] Collect historical ocean observation data and records of historical ocean disaster events. The time span of the historical ocean observation data shall be no less than 10 years, and the spatial coverage of the historical ocean observation data shall be no less than 50 fourth-level spatial scale grid units.

[0089] Perform quality control and outlier removal on historical ocean observation data, and fill in missing data;

[0090] Based on the occurrence time and spatial location of historical marine disaster events, positive and negative samples in the training dataset are labeled. The positive samples are the time series of marine dynamic parameters within 72 hours before the disaster, and the negative samples are the time series of marine dynamic parameters during the period when no disaster occurred.

[0091] Data augmentation is performed on positive and negative samples, including time shifting, scaling, and adding random perturbations that conform to physical laws.

[0092] The training set, validation set, and test set are divided into three parts in a 7:2:1 ratio.

[0093] The specific steps for training the multi-scale coupled marine disaster model include:

[0094] Initialize the model parameters using the Xavier initialization method;

[0095] The loss function is defined as the weighted sum of the cross-entropy loss between the predicted sequence and the actual disaster label and the physical consistency constraint loss within the predicted sequence. The physical consistency constraint loss is calculated by checking whether the predicted ocean dynamic parameters satisfy the mass conservation equation, momentum conservation equation, and energy conservation equation.

[0096] The adaptive moment estimation optimization algorithm is used to update the model parameters. The learning rate is set to 0.001, the batch size is set to 32, and the number of training epochs is set to 200.

[0097] The model performance is evaluated using the validation set after each training cycle. If the validation set loss does not decrease for 10 consecutive training cycles, the learning rate is reduced to 50% of its original value. Training is stopped when the validation set loss does not decrease for 20 consecutive training cycles. The generalization performance of the trained model is evaluated using the test set.

[0098] The steps for determining the sparsity parameter specifically include:

[0099] The proportion of non-zero elements in the ultra-sparse cross-scale coupling matrix is ​​calculated as the basic sparsity; the basic sparsity is multiplied by the square root of the duration adjustment factor to obtain the first adjusted sparsity.

[0100] Divide the first adjusted sparsity by the cube root of the scaling factor to obtain the second adjusted sparsity.

[0101] The second sparsity adjustment is restricted to the interval [1%, 15%] as the sparsity parameter of the sparse attention mechanism. The sparsity parameter controls the number of key vectors that each query vector in the sparse attention mechanism focuses on. A smaller sparsity parameter allows the multi-scale coupling model of marine disasters to focus on the most critical cross-scale coupling relationships, improving computational efficiency and reducing the risk of overfitting.

[0102] The steps of the integrated forecasting process specifically include:

[0103] The multi-scale coupled model of marine disasters was trained independently multiple times. Each independent training session used different random initialization parameters and different subsets of training data to obtain several independent prediction models.

[0104] Statistical analysis is performed on the prediction results of all independent prediction models on the same input data. The mean of the prediction results is calculated as the integrated prediction value, and the standard deviation of the prediction results is calculated as a measure of prediction uncertainty.

[0105] The prediction uncertainty interval is constructed based on the distribution characteristics of the prediction results. The upper bound of the prediction uncertainty interval is the integrated prediction value plus 1.96 times the standard deviation, the lower bound of the prediction uncertainty interval is the integrated prediction value minus 1.96 times the standard deviation, and the width of the prediction uncertainty interval is the upper bound of the prediction uncertainty interval minus the lower bound of the prediction uncertainty interval.

[0106] Integrated forecasting processing quantifies the uncertainty of forecasts by combining forecast information from multiple independent forecasting models, providing decision-makers with probabilistic early warnings rather than single deterministic forecasts. This is of great value for risk assessment and emergency response to marine disasters.

[0107] The steps for determining the preset threshold specifically include:

[0108] The distribution of the uncertainty interval for predicting historical marine disaster events was statistically analyzed, and the 75th percentile of the distribution was calculated as the initial threshold.

[0109] The initial threshold is adjusted according to the degree of danger of different disaster types. For high-risk disasters, the initial threshold is reduced to 70% of the original threshold to obtain the preset threshold. For medium-risk disasters, the initial threshold remains unchanged as the preset threshold. For low-risk disasters, the initial threshold is increased to 130% of the original threshold to obtain the preset threshold. The setting of the preset threshold balances the accuracy and timeliness of the early warning. When the width of the prediction uncertainty interval is too large, the data quality is improved by increasing the observation density, thereby improving the reliability of the prediction.

[0110] The steps to increase the observation density specifically include:

[0111] Based on the spatial distribution of the width of the prediction uncertainty interval on a four-level spatial scale grid, the top 20% of grid regions with the largest prediction uncertainty interval width are identified as the densified observation areas.

[0112] Increase the number of mobile observation platforms deployed within the encrypted observation area; these mobile observation platforms include buoys, underwater gliders, and unmanned vessels.

[0113] The observation time interval within the encrypted observation area will be reduced to 50% of the original time.

[0114] We re-execute the time normalization and spatial normalization processes, as well as the cross-scale feature extraction process, using the newly added high-density observation data.

[0115] Increasing observation density is an adaptive data acquisition strategy. By investing more observation resources in areas with a large range of uncertainty in the prediction of multi-scale coupled models of marine disasters, the efficiency and reliability of the overall early warning system are improved.

[0116] The marine disaster early warning information includes disaster type, predicted time interval, affected spatial range, disaster intensity level, and suggested response measures. The disaster type includes storm surge, tsunami, red tide, and sea ice. The predicted time interval is determined based on the time period in the marine disaster prediction sequence that exceeds the disaster threshold in the four-level time length sequence. The affected spatial range is determined based on the grid area in the marine disaster prediction sequence that exceeds the disaster threshold in the four-level spatial scale grid. The disaster intensity level is divided into four levels: mild, moderate, severe, and extremely severe, based on the statistical comparison between the extreme values ​​of marine dynamic parameters in the marine disaster prediction sequence and historical marine disaster events.

[0117] The specific steps of the compressed sensing reconstruction mechanism include:

[0118] By taking advantage of the sparsity assumption of the ultrasparse cross-scale coupling matrix, the complete cross-scale coupling state is regarded as a high-dimensional sparse signal.

[0119] A low-dimensional observation matrix is ​​generated from observation data of a four-level spatial scale grid using a random projection method. The dimension of the low-dimensional observation matrix belongs to the range of the complete cross-scale coupled state dimension [25%, 40%].

[0120] The orthogonal matching pursuit algorithm is used to reconstruct a high-dimensional sparse signal from a low-dimensional observation matrix. The orthogonal matching pursuit algorithm selects the sparse basis function that best matches the observation data through an iterative method.

[0121] The reconstructed high-dimensional sparse signal is used as supplementary information to the ultra-sparse cross-scale coupling matrix to correct and improve the expression of cross-scale coupling relationships. The compressed sensing reconstruction mechanism, when faced with incomplete observation data or gaps in the observation network, utilizes the inherent sparse structure of ocean dynamic processes to recover complete cross-scale coupling information from limited observations, significantly improving the robustness of the early warning system under conditions of limited observation resources.

[0122] The present invention also provides a method for forming a marine disaster early warning system by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions execute the above-described method when the computer is run.

[0123] The specific implementation methods of the above steps are described in detail below.

[0124] The specific implementation of step S01 is as follows: First, a time range of 1 to 24 hours is set for short-term hourly series, a time range of 24 to 168 hours is set for medium-term daily series, a time range of 168 to 720 hours is set for long-term weekly series, and a time range of 720 to 2160 hours is set for ultra-long-term monthly series. Within each time range, the sampling point positions are determined by a logarithmic interval. The logarithmic interval is calculated by dividing the upper limit of the time range by the lower limit, taking the natural logarithm, and then dividing by the number of sampling points minus 1. All sampling points are arranged in chronological order to form a complete time series. This time series design method ensures sufficient resolution on short time scales. High resolution is used to capture abrupt change signals. Logarithmic intervals are used over long time scales to avoid computational redundancy caused by excessively dense sampling points. Subsequently, for different spatial scales, micro-scale grids are set at 100m (hundred-meter level), meso-scale grids at 1000m (kilometer level), large-scale grids at 10000m (ten-kilometer level), and mega-scale grids at 100000m (hundred-kilometer level). The coverage of each grid level is determined based on the geographical boundaries and coastline morphology of the warning sea area. The hundred-meter micro-scale grids are deployed in key near-shore protection areas, the kilometer-scale meso-scale grids cover densely populated coastal economic activity areas, the ten-kilometer-scale large-scale grids cover the continental shelf, and the hundred-kilometer-scale mega-scale grids cover deep-sea and open ocean areas. The system spatially partitions the data according to the corresponding grid size and records the geographic coordinates of the grid nodes. A nested containment relationship is established between adjacent scale levels, ensuring a spatial correspondence between the coordinates of fine and coarse grid nodes. A duration adjustment factor is set for the time series: a baseline time length of 1 hour for short-term hourly series, 24 hours for medium-term daily series, 168 hours for long-term weekly series, and 720 hours for ultra-long-term monthly series. The target prediction time length for the prediction task is obtained, and its ratio to the baseline time length of each series is calculated as the original duration ratio. A base-10 logarithmic transformation is then applied to the original duration ratio to obtain the logarithmic duration. The logarithmic time ratio is transformed to the interval of 0 to 1 through a linear mapping and used as a time adjustment factor. Scale adjustment factors are set for spatial grids: a reference grid size of 100m is set for micro-scale grids at the hundred-meter level, 1000m for meso-scale grids at the kilometer level, 10000m for large-scale grids at the ten-kilometer level, and 100000m for giant-scale grids at the hundred-kilometer level. The ratio of the actual size of each grid level to the reference size is calculated as the original scale ratio. The original scale ratio is then logarithmically transformed to base 10 to obtain the logarithmic scale ratio. The logarithmic scale ratio is then transformed to the interval of 0 to 1 through a linear mapping and used as a scale adjustment factor.

[0125] The specific implementation of step S02 is as follows: An array of marine observation sensors is deployed at each grid node of a four-level spatial scale grid. Sensor types include temperature probes for collecting sea surface temperature data, current meters for collecting ocean current velocity data, wave height meters for collecting wave height data, salinity meters for collecting salinity data, and pressure sensors for collecting seabed pressure data. Each sensor collects data according to the sampling intervals set in the four-level time length sequence. The collected marine dynamic parameters are first time-normalized, and the collected data are multiplied by a duration adjustment factor corresponding to the corresponding time scale to achieve weighted processing. The time-scale normalization process involves subtracting the mean of the parameter's time series and dividing by the standard deviation of the time series to achieve time-scale normalization. This normalization eliminates dimensional differences between data at different time scales and standardizes the data distribution. After time normalization, spatial normalization is performed by multiplying the time-normalized data by the scale adjustment factor of the corresponding spatial scale to achieve spatial weighting. Then, the mean of the parameter on the spatial grid is subtracted and divided by the standard deviation of the spatial grid to achieve spatial-scale normalization. This dual normalization process ensures that ocean dynamic parameters at different temporal and spatial scales are comparable and consistent before being entered into the model.

[0126] The specific implementation of step S03 is as follows: When establishing the ultra-sparse cross-scale coupling matrix, the adjacency relationship between the four-level spatial scale grids is first determined. For any two grid nodes at adjacent scales, the spatial distance between the two nodes is calculated and compared with 50% of the larger scale grid size. When the distance is less than this threshold, the two nodes are considered to have a coupling relationship. An initial coupling matrix is ​​constructed based on the coupling relationship, with the number of rows and columns of the matrix equal to the total number of nodes in the four-level grid. The physical influence intensity between each pair of coupled nodes is calculated. This intensity is represented by a weighted average of the ocean current velocity gradient, sea surface temperature gradient, and salinity gradient between the nodes, with weighting coefficients respectively... The matrix elements with physical influence strength less than 5% of the average influence strength of all coupled nodes are set to zero to form an ultra-sparse cross-scale coupling matrix, with the values ​​being 0.5, 0.3, and 0.2. The proportion of non-zero elements in this matrix is ​​typically between 2% and 8%. Subsequently, a micro-feature temporal dependency matrix is ​​constructed. For each time step of the fourth-level time series, the small variation features of ocean dynamic parameters are extracted. These small variations are represented by calculating the first and second differences of parameters between adjacent time steps. The Pearson correlation coefficients of the small variation features between different time steps are calculated, and a temporal dependency matrix is ​​constructed. The number of rows and columns in this matrix is ​​equal to the total number of time steps in the fourth-level time series. The matrix elements represent the correlation coefficients of corresponding time step pairs. Matrix elements with absolute correlation coefficients less than 0.15 are set to zero to form a micro-feature temporal dependency matrix. This matrix captures weak correlation patterns of ocean dynamic processes at different time scales and filters out random noise. After calling the ultra-sparse cross-scale coupling matrix and the micro-feature temporal dependency matrix, the normalized ocean dynamic parameters are input into the multi-scale coupling model of ocean disasters. The model includes an input layer, a multi-scale feature extraction layer, a cross-scale fusion layer, and a prediction output layer. The input layer embeds the normalized parameters into the initial feature space. The multi-scale feature extraction layer contains four parallel branches corresponding to four levels. At each scale level of the spatial scale grid, each branch uses a residual connection structure to extract feature representations of the corresponding scale. The residual connection alleviates the gradient vanishing problem of deep networks through skip connections. The cross-scale fusion layer calculates the correlation weights between features of different scales through a sparse attention mechanism. The sparse attention only calculates the weights for scale pairs corresponding to non-zero elements in the ultra-sparse cross-scale coupling matrix, and adjusts the attention distribution of the temporal dimension according to the micro-feature temporal dependency matrix. This sparse mechanism significantly reduces computational complexity and focuses on key coupling relationships. Through cross-scale feature extraction, scale-separated feature vector groups corresponding to the four-level spatial scale grid are obtained.

[0127] The specific implementation of step S04 is as follows: The rank of the ultra-sparse cross-scale coupling matrix is ​​calculated. The rank reflects the number of linearly independent rows or columns in the matrix. When the rank is less than 30% of the total number of nodes in the fourth-level spatial scale grid, scale reduction is triggered. Singular value decomposition is performed on the ultra-sparse cross-scale coupling matrix, and the singular vectors corresponding to the first few largest singular values ​​are extracted as the main scale patterns. The scale-separated feature vector group is projected onto the subspace spanned by the main scale patterns to obtain the dimensionality-reduced feature vectors. The dimensionality after reduction is 40% to 60% of the original dimension. This dimensionality reduction operation reduces the computational burden and suppresses overfitting when the inter-scale coupling relationship is relatively simple. When the rank is greater than 85% of the total number of nodes in the fourth-level spatial scale grid, scale expansion is triggered. Eigenvalue decomposition is performed on the ultra-sparse cross-scale coupling matrix, and the weak coupling patterns corresponding to the feature vectors with smaller eigenvalues ​​are identified. The weak coupling patterns are then added to the scale-separated feature vector group. Auxiliary feature components are obtained by extracting high-frequency detail coefficients after performing wavelet transform on normalized ocean dynamic parameters. Wavelet transform can analyze the local features of the signal in both the time and frequency domains. After dimensional expansion, the dimension is 140% to 160% of the original dimension. This dimensional expansion operation provides richer feature representations when the inter-scale coupling relationship is highly active. Subsequently, the scale-separated feature vector group is mapped to a high-dimensional feature space through the radial basis function kernel function. The Euclidean distance between each feature vector and the preset kernel center vector is calculated. The kernel center vector is selected from the feature vectors of historical marine disaster events through the K-means clustering algorithm. The Euclidean distance is divided by the kernel width parameter, the negative value is taken, and an exponential operation is performed to obtain the kernel function mapping value. The kernel width parameter is equal to the reciprocal of the scale adjustment factor multiplied by the adjustment coefficient 0.1. The radial basis function kernel function maps linearly inseparable feature vectors to a high-dimensional space to make them linearly separable, enabling the model to learn complex nonlinear scale coupling relationships.

[0128] The specific implementation of step S05 is as follows: Feature vectors from the high-dimensional feature space are input into a multi-scale coupled model for marine disaster prediction. The model employs a continuous deep modeling algorithm based on neural ordinary differential equations (ODEs). The hidden state in the network is represented as a function of a continuous-time variable. The initial value of the continuous-time variable is set to 0, corresponding to the input layer, and the final value is set to 1, corresponding to the prediction output layer. The derivative of the hidden state with respect to the continuous-time variable is defined as a function of the network parameters and the current hidden state. This function is modeled through a neural network, and the ODEs are solved using the fourth-order Runge-Kutta method. The hidden state of the prediction output layer is obtained by integrating from the initial value to the final value. The fourth-order Runge-Kutta method is a classic algorithm for numerically solving ODEs. It improves the solution accuracy through multi-step prediction and correction. During backpropagation, an adjoint ODE is constructed. The solution of the adjoint equation gives the gradient of the loss function with respect to the network parameters. This method avoids the gradient problem in traditional backpropagation. To address the vanishing degree and gradient explosion problems, continuous deep modeling algorithms extend network depth from discrete layers to a continuous time dimension, enabling the model to adaptively adjust the feature evolution time based on input complexity. The sparsity parameter of the sparse attention mechanism in the model is determined by a combination of the proportion of non-zero elements in the ultra-sparse cross-scale coupling matrix, the duration adjustment factor, and the scale adjustment factor. First, the proportion of non-zero elements in the ultra-sparse cross-scale coupling matrix is ​​calculated as the basic sparsity. The basic sparsity is multiplied by the square root of the duration adjustment factor to obtain the first adjusted sparsity. The first adjusted sparsity is divided by the cube root of the scale adjustment factor to obtain the second adjusted sparsity. The second adjusted sparsity is restricted to the range of 1% to 15% as the final sparsity parameter. The sparsity parameter controls the number of key vectors that each query vector focuses on. A smaller sparsity allows the model to focus on key cross-scale coupling relationships. The model outputs a marine disaster prediction sequence corresponding to a four-level time length sequence.

[0129] The specific implementation of step S06 is as follows: Integrated forecasting processing is performed on the marine disaster prediction sequence. The multi-scale coupled model for marine disasters is trained independently multiple times. Each training iteration uses different random initialization parameters and different subsets of training data to obtain several independent prediction models. Statistical analysis is performed on the prediction results of all independent models on the same input data. The mean of the prediction results is calculated as the integrated prediction value, and the standard deviation of the prediction results is calculated as a measure of prediction uncertainty. A prediction uncertainty interval is constructed based on the distribution characteristics of the prediction results. The upper bound of the interval is the integrated prediction value plus 1.96 times the standard deviation, the lower bound is the integrated prediction value minus 1.96 times the standard deviation, and the interval width is the upper bound minus the lower bound. This integration method quantifies prediction uncertainty by integrating information from multiple independent models. When determining a preset threshold, the distribution of the prediction uncertainty interval width of historical marine disaster events is statistically analyzed, and the 75th quantile of the distribution is calculated as the initial threshold. The initial threshold is adjusted according to the hazard level of different disaster types. For high-risk disasters, the initial threshold is reduced to 70% of the initial threshold as the preset threshold; for medium-risk disasters, the initial threshold remains unchanged; and for low-risk disasters, the initial threshold is increased to 130% of the initial threshold as the preset threshold. When the prediction... When the uncertainty interval width exceeds a preset threshold, the observation density is increased. Based on the spatial distribution of the predicted uncertainty interval width on the four-level spatial scale grid, the top 20% of grid areas with the largest predicted uncertainty interval width are identified as dense observation areas. The number of mobile observation platforms deployed in the dense observation areas is increased. Mobile observation platforms include buoys, underwater gliders, and unmanned vessels. The observation time interval in the dense observation areas is shortened to 50% of the original. Steps S02 to S05 are re-executed using the newly added high-density observation data. When the predicted uncertainty interval width is less than the preset threshold, marine disaster early warning information is generated and output. The early warning information includes the disaster type, the predicted time interval, the affected spatial range, the disaster intensity level, and suggested countermeasures. The disaster types include storm surge, tsunami, red tide, and sea ice. The predicted time interval is determined based on the time period in the marine disaster prediction sequence in the four-level time length sequence that exceeds the disaster threshold. The affected spatial range is determined based on the grid area in the marine disaster prediction sequence in the four-level spatial scale grid that exceeds the disaster threshold. The disaster intensity level is divided into four levels: mild, moderate, severe, and extremely severe, based on the comparison between the extreme values ​​of marine dynamic parameters in the marine disaster prediction sequence and the statistical analysis of historical marine disaster events.

[0130] Specifically, the principle of this invention is as follows: This invention can solve the technical problem that marine disaster early warning cannot simultaneously handle multi-temporal and spatial scale coupled processes. The fundamental reason is that the solution establishes an explicit multi-scale representation framework and cross-scale information transmission channel. Traditional methods mix information from different temporal and spatial scales in a single feature space for processing, resulting in the coupling relationship between scales being implicit in complex nonlinear mappings that are difficult to identify and control. In contrast, this invention separates multi-scale information at the input stage through a four-level time length sequence and a four-level spatial scale grid, and extracts feature representations independently at each scale level, ensuring that the physical characteristics of the dynamic processes at each scale are fully characterized. The establishment of the ultra-sparse cross-scale coupling matrix is ​​based on the calculation of the physical influence intensity between grid nodes. This matrix explicitly defines which scale pairs have significant coupling effects. The sparse attention mechanism selectively calculates the cross-scale correlation weights based on this matrix. This design ensures that the model only learns physically meaningful inter-scale interactions, avoiding spurious correlations in traditional fully connected attention mechanisms. The micro-feature temporal dependency matrix quantifies the correlation of small changes in ocean dynamic parameters between different time steps through the Pearson correlation coefficient, providing an interpretable basis for information transmission across time scales. The duration adjustment factor and scale adjustment factor dynamically adjust the contribution weights of each scale according to the target time length of the prediction task and the spatial heterogeneity of the ocean state, realizing the adaptive capability of the prediction model. The continuous depth modeling algorithm expands the network depth from discrete layers to a continuous time dimension, enabling features at different scales to evolve at different rates over continuous time. This design aligns with the physical fact that different scales in ocean dynamic processes have different characteristic times. The rapid evolution of small-scale turbulence corresponds to a short feature evolution time, while the slow adjustment of large-scale circulation corresponds to a long feature evolution time. The ordinary differential equation solver ensures the smoothness of the feature evolution path and avoids abrupt information changes between discrete layers. The accurate gradient calculation of the accompanying ordinary differential equation method solves the error accumulation problem in long-term series prediction. When the width of the prediction uncertainty interval exceeds a preset threshold, the system triggers an increased observation density mechanism, investing more observation resources in areas with high prediction uncertainty, thus forming a closed-loop feedback between data acquisition and the prediction model. Therefore, the technical solution of this invention is logically self-consistent, and the various technical features support each other to jointly solve the prediction problem of multi-temporal-scale coupled processes.

[0131] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0132] The specific implementation of step S01 is as follows: First, the logarithmic interval sampling points are set for the four-level time length series. The formula for calculating the logarithmic interval value is expressed as follows:

[0133] ;

[0134] In the formula, These are logarithmic interval values, in dimensionless units. This represents the upper limit of the time range, in hours (h), for short-term series on the hourly scale. h, for the mid-term sequence of the Heaven level h, for weekly long-term sequences h, for monthly ultra-long-term sequences h; This is the lower limit of the time range, in hours, for short-term series on the hourly scale. h, for the mid-term sequence of the Heaven level h, for weekly long-term sequences h, for monthly ultra-long-term sequences h; This refers to the number of sampling points, expressed in units of 10 to 20; It is the natural logarithm function.

[0135] The formula for calculating the duration adjustment factor is as follows:

[0136] ;

[0137] In the formula, For the first Duration adjustment factor for time series, These correspond to hourly short-term sequences, daily medium-term sequences, weekly long-term sequences, and monthly ultra-long-term sequences, respectively, with units that are dimensionless. The target prediction time is in hours. For the first The base time length of the time series is in hours (h), where h, h, h, h; It is a logarithmic function with base 10; The minimum log-time ratio among all time series, in dimensionless units, is calculated... hour The minimum value is obtained; The maximum value of the logarithmic duration ratio among all time series, in dimensionless units, is calculated. hour The maximum value is obtained.

[0138] The formula for calculating the scaling factor is as follows:

[0139] ;

[0140] In the formula, For the first Scale adjustment factor for spatial grids. These correspond to micro-scale grids at the hundred-meter level, meso-scale grids at the kilometer level, large-scale grids at the ten-kilometer level, and giant-scale grids at the hundred-kilometer level, respectively, with units being dimensionless. For the first The actual grid size of the sub-grid, in meters; For the first The reference grid size for the level-1 grid, in meters, is where... m, m, m, m; The minimum logarithmic scale ratio across all spatial grids, in dimensionless units, is calculated... hour The minimum value is obtained; The maximum value of the logarithmic scale ratio across all spatial grids, in dimensionless units, is calculated. hour The maximum value is obtained.

[0141] The specific implementation of step S02 is as follows: the calculation formula for time normalization processing is expressed as follows:

[0142] ;

[0143] In the formula, These are the time-normalized ocean dynamic parameters. For parameter type index, These correspond to sea surface temperature, ocean current velocity, wave height, salinity, and seabed pressure, respectively. For time-scale indexing This is an index for time sampling points, in dimensionless units; For the first The duration adjustment factor for the time series, in dimensionless units; For the first The raw ocean dynamic parameters collected at each time sampling point are in units determined according to the parameter type: sea surface temperature is in °C, ocean current velocity is in m / s, wave height is in m, salinity is in PSU, and seabed pressure is in Pa. For the first The parameter type in the first Mean values ​​of ocean dynamic parameters for time series, in units of same; For the first The parameter type in the first Standard deviation of ocean dynamic parameters for time series, in units of same.

[0144] in, The calculation formula is expressed as follows:

[0145] ;

[0146] In the formula, For the first The total number of sampling points for the time series, expressed in units of points; The meaning is the same as described above.

[0147] The calculation formula is expressed as follows:

[0148] .

[0149] The calculation formula for spatial normalization is expressed as follows:

[0150] ;

[0151] In the formula, These are the normalized ocean dynamic parameters. For parameter type index, For time-scale indexing For spatial scale level indexing, For grid node indexes, the units are dimensionless; For the first The scale adjustment factor of the spatial grid, in dimensionless units; For the first Ocean dynamic parameters after time normalization of each grid node, in dimensionless units; For the first The parameter type in the first Mean values ​​of ocean dynamic parameters in a spatial grid, in dimensionless units; For the first The parameter type in the first The standard deviation of ocean dynamic parameters in a spatial grid is given in dimensionless form.

[0152] in, The calculation formula is expressed as follows:

[0153] ;

[0154] In the formula, For the first The total number of nodes in the hierarchical spatial grid, expressed in units of nodes; The meaning is the same as described above.

[0155] The calculation formula is expressed as follows:

[0156] .

[0157] The specific implementation of step S03 is as follows: the formula for calculating the physical influence intensity of the ultra-sparse cross-scale coupling matrix is ​​expressed as follows:

[0158] ;

[0159] In the formula, For the first The grid node and the first The intensity of physical influence between grid nodes, in dimensionless units; For the first The grid node and the first The ocean current velocity gradient magnitude between grid nodes, in units of 1 / s, is obtained through... Calculated, where For the first Ocean current velocity at each grid node, in m / s. For the first Ocean current velocity at each grid node, in m / s. For the first The grid node and the first The spatial distance between grid nodes, in meters; The default value is to reference the ocean current velocity gradient. / s; For the first The grid node and the first The sea surface temperature gradient modulus between grid nodes, in °C / m, is obtained through... Calculated, where For the first The sea surface temperature of each grid node, in °C. For the first Sea surface temperature of each grid node, in °C; To reference the sea surface temperature gradient, the default is... ℃ / m; For the first The grid node and the first The salinity gradient magnitude between grid nodes, in units of PSU / m, is obtained through... Calculated, where For the first Salinity of each grid node, in PSU. For the first The salinity of each grid node, in PSU.

[0160] The formula for calculating the Pearson correlation coefficient in the micro-feature temporal dependency matrix is ​​as follows:

[0161] ;

[0162] In the formula, For time steps With time step The Pearson correlation coefficient between them, in dimensionless units; For time steps First The minute variation characteristics of a marine dynamic parameter, in dimensionless units; For time steps First The minute variation characteristics of a marine dynamic parameter, in dimensionless units; For time step The mean value of all minute variations in ocean dynamic parameters, in dimensionless units, is obtained through... Calculated; For time step The mean value of all minute variations in ocean dynamic parameters, in dimensionless units, is obtained through... Calculated; This represents the total number of marine dynamic parameters, expressed in units of individual parameters. In this scheme... .

[0163] in, The calculation formula is expressed as follows:

[0164] ;

[0165] In the formula, For time step First These are normalized ocean dynamic parameters, in dimensionless units. For time step First These are normalized ocean dynamic parameters, in dimensionless units. For time step First These are normalized ocean dynamic parameters, in dimensionless units. The time step is expressed in hours (h). The default time step is 1 hour. This is the first-order difference weighting coefficient, with an empirical value of 0.6; The second-order difference weighting coefficient has an empirical value of 0.4.

[0166] The specific implementation of step S04 is as follows: the formula for calculating the radial basis kernel function is expressed as follows:

[0167] ;

[0168] In the formula, For the first The eigenvector and the eigenvector The kernel function mapping value of each kernel center vector, in dimensionless units; For the first Each scale-separated feature vector is dimensionless and is a column vector containing multiple feature components. For the first A kernel center vector, with dimensionless units, is a representative vector selected from the feature vectors of historical marine disaster events using the K-means clustering algorithm; The Euclidean distance between the eigenvectors and the kernel center vector is dimensionless. Calculated, where For the first The th eigenvector of the th feature vector One portion, For the first The first kernel center vector One portion, The dimension of the feature vector; This is the adjustment coefficient, with an empirical value of 0.1; The kernel width parameter is determined by the formula. The calculated value is dimensionless. It is a natural exponential function.

[0169] The specific implementation of step S05 is as follows: the formula for calculating the sparsity parameter is expressed as follows:

[0170] ;

[0171] In the formula, represents the sparsity parameter of the sparse attention mechanism, in dimensionless form; The proportion of non-zero elements in the ultrasparse multi-scale coupling matrix, in dimensionless units, is expressed through... Calculated, where The number of non-zero elements in the ultrasparse multi-scale coupling matrix, expressed in units of 1. The total number of elements in the ultrasparse cross-scale coupling matrix, expressed in units of 1, typically 1. The value ranges from 0.02 to 0.08; This is a duration adjustment factor, with dimensionless units. This is a scale adjustment factor, with a dimensionless unit. For the truncation function, Limited to the range inside, when Time return ,when Time return ,when Time return .

[0172] The specific implementation of step S06 is as follows: the formula for calculating the width of the uncertainty interval is expressed as follows:

[0173] ;

[0174] In the formula, To predict the width of the uncertainty interval, the unit is the same as the prediction target; 3 represents the standard deviation of the prediction results, with the same units as the prediction target; 3.92 is the coefficient, derived from... We obtain that 1.96 is the standard normal distribution quantile at the 95% confidence level.

[0175] in, The calculation formula is expressed as follows:

[0176] ;

[0177] In the formula, The total number of independent prediction models, expressed in units of 5 to 10; For the first The prediction results of each independent prediction model are in the same unit as the prediction target. To integrate the predicted values, the units are the same as the prediction target.

[0178] The calculation formula is expressed as follows:

[0179] .

[0180] The formula for calculating the preset threshold is as follows:

[0181] ;

[0182] In the formula, The preset threshold is used, and the unit is the same as the width of the prediction uncertainty interval. The initial threshold is calculated by the 75th percentile of the distribution of the prediction uncertainty interval width of historical marine disaster events, with the same unit as the prediction uncertainty interval width; 0.7 is the adjustment coefficient for high-risk disasters; and 1.3 is the adjustment coefficient for low-risk disasters.

[0183] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0184] To verify the effectiveness of this invention, the technicians conducted a retrospective analysis of the Western Pacific tsunami disaster that occurred in March 2011. This tsunami, triggered by an undersea earthquake, caused severe coastal damage. The technicians used the marine disaster early warning method of this invention to conduct a retrospective predictive study of this historical event.

[0185] The engineers first constructed a four-level time-length series and a four-level spatial scale grid. The time range for the hourly short-term series was set to 1 to 24 hours; the day-level medium-term series to 24 to 168 hours; the week-level long-term series to 168 to 720 hours; and the month-level ultra-long-term series to 720 to 2160 hours. Regarding the four-level spatial scale grid, the 100-meter micro-scale grid has a size of 100 meters, covering nearshore areas within 5 km of the coastline; the kilometer-level meso-scale grid has a size of 1000 meters, covering the coastal economic zone from 5 to 50 km of the coastline; the ten-kilometer-level large-scale grid has a size of 10,000 meters, covering the continental shelf area from 50 to 500 km of the coastline; and the hundred-kilometer-level mega-scale grid has a size of 100,000 meters, covering the deep-sea area beyond 500 km of the coastline. The total number of nodes in the four-level spatial scale grid is 18,750, including 2,500 micro-scale grid nodes at the hundred-meter level, 5,000 meso-scale grid nodes at the kilometer level, 6,250 large-scale grid nodes at the ten-kilometer level, and 5,000 mega-scale grid nodes at the hundred-kilometer level.

[0186] Technicians deployed ocean observation sensors, including temperature sensors, current meters, wave height meters, salinity meters, and seabed pressure sensors, at various nodes of a four-level spatial scale grid. The ocean dynamic parameters collected from 72 hours before the earthquake to the moment of earthquake occurrence are shown in Table 1.

[0187] Table 1. Mean values ​​of ocean dynamic parameters at different time points before the earthquake.

[0188]

[0189] Technicians performed time and spatial normalization on the collected ocean dynamic parameters. Since the target prediction time for the current mission is 48 hours, the calculated duration adjustment factor is 0.12 for hourly short-term sequences, 0.68 for day-level medium-term sequences, 0.15 for weekly long-term sequences, and 0.05 for monthly ultra-long-term sequences. For four spatial scale grids, the calculated scale adjustment factor is 0.85 for hundred-meter micro-scale grids, 0.62 for kilometer-level meso-scale grids, 0.38 for ten-kilometer-level large-scale grids, and 0.15 for hundred-kilometer-level giant-scale grids.

[0190] Technicians established an ultra-sparse cross-scale coupling matrix. Based on the spatial adjacency relationships between the four-level spatial scale grids, the physical influence strength between each pair of coupled nodes was calculated. The weighting coefficient for ocean current velocity gradient was 0.5, for sea surface temperature gradient it was 0.3, and for salinity gradient it was 0.2. Coupling relationships with physical influence strength less than 5% of the average were zeroed out, resulting in an ultra-sparse cross-scale coupling matrix with a non-zero element ratio of 4.7% and a rank of 5438. Since the rank was less than 30% of the total number of nodes in the four-level spatial scale grid, dimensionality reduction was performed. Singular value decomposition was performed on the ultra-sparse cross-scale coupling matrix, and the singular vectors corresponding to the top 8650 largest singular values ​​were extracted as the main scale patterns. The dimensionality of the reduced feature vectors was 46% of the original dimension.

[0191] Technicians established a micro-feature temporal dependency matrix. First and second differences of ocean dynamic parameters were extracted for each time step in the four-level time-length sequence, and Pearson correlation coefficients were calculated between different time steps. Matrix elements with absolute Pearson correlation coefficients less than 0.15 were set to zero. The resulting micro-feature temporal dependency matrix showed a significant but weak correlation between the period 12 hours to 0 hours before the earthquake and the period 48 hours to 36 hours before the earthquake, with a Pearson correlation coefficient of 0.23, indicating that ocean dynamic processes exhibited similar minor variations in these two periods.

[0192] like Figure 2 As shown, the technicians mapped the scale-separated feature vector group to a high-dimensional feature space using a radial basis function kernel. The kernel center vectors were selected from the feature vectors of historical tsunami events using a K-means clustering algorithm, resulting in 120 kernel center vectors. The kernel width parameter was calculated based on a scale adjustment factor: 1.18 for a 100-meter micro-scale grid, 1.61 for a kilometer-scale meso-scale grid, 2.63 for a 10-kilometer-scale large-scale grid, and 6.67 for a 100-kilometer-scale giant-scale grid. Figure 2 The paper clearly demonstrates the correspondence between kernel width parameters and scale adjustment factors at four scale levels. As the spatial scale increases from hundreds of meters to hundreds of kilometers, the kernel width parameter shows an exponential growth trend, while the scale adjustment factor shows a decreasing trend, reflecting the differentiated processing strategies in the feature mapping process at different spatial scales.

[0193] Technicians input feature vectors from a high-dimensional feature space into a multi-scale coupled model of marine disasters. The model employs a continuous deep modeling algorithm based on neural ordinary differential equations, treating the network layers in the multi-scale feature extraction layer and the cross-scale fusion layer as discretizations of a continuous-time dynamic system. The sparsity parameter of the sparse attention mechanism is calculated based on the proportion of non-zero elements in the ultra-sparse cross-scale coupling matrix, the duration adjustment factor, and the scale adjustment factor. The base sparsity is 4.7%, the first adjusted sparsity is 1.63%, the second adjusted sparsity is 1.81%, and the final sparsity parameter after limiting it to an interval is 1.81%.

[0194] like Figure 3 As shown, technicians created a four-level spatial scale grid map of the distribution of seafloor pressure anomalies 12 hours before the earthquake. The map uses a red-blue-green gradient to display the seafloor pressure anomaly values ​​at different grid nodes. Red areas indicate seafloor pressure significantly higher than normal levels, blue areas indicate seafloor pressure lower than normal levels, and green areas indicate seafloor pressure close to normal levels. A distinct red high-pressure anomaly band is visible above the epicenter region, with pressure anomalies reaching [value missing]. A ring-shaped pressure gradient distribution formed within a 200km radius of the epicenter. A 100-meter-scale micro-grid captured the fine structure of local pressure fluctuations in the nearshore region, exhibiting fine texture features near the coastline; a kilometer-scale meso-scale grid showed the pressure propagation path along the coastline, presenting a wavy distribution radiating from the epicenter to the surrounding area; a 10-kilometer-scale large-scale grid revealed the pressure response characteristics of the continental shelf region, manifested as large-scale gradient color patches; and a 100-kilometer-scale mega-grid described the pressure background field distribution in the deep sea region, showing the overall macroscopic trend of change. The epicenter location marked in the cloud image precisely coincides with the center of the high-pressure anomaly region, and the ring-shaped pressure gradient band is clearly visible, verifying the effective capture capability of the ultra-sparse multi-scale coupling matrix for multi-scale dynamic processes.

[0195] like Figure 4 As shown, the model outputs marine disaster prediction sequences corresponding to four time-length series. The hourly short-term series prediction shows that seafloor pressure began to rise abnormally 8 hours before the earthquake, with a rate of increase reaching [missing information]. The curve shows a steep upward trend near the time of the earthquake; the day-level medium-term sequence prediction shows that the ocean current velocity oscillates more strongly 36 hours before the earthquake, with an oscillation period of 4.5 hours, and the curve shows regular fluctuations accompanied by an overall upward trend; the week-level long-term sequence prediction shows that the sea surface temperature gradient above the epicenter region shows an abnormal distribution pattern 5 days before the earthquake, and the curve begins to slowly climb 120 hours before the earthquake; the month-level ultra-long-term sequence prediction shows that the ocean circulation structure in this sea area shows a slow adjustment trend one month before the earthquake, and the curve shows gentle long-period fluctuations. Figure 4The four curves of different colors represent prediction sequences at four time scales. Vertical dashed lines and text labels are added at two key time points: 8 hours and 36 hours before the earthquake, highlighting important early warning signals of abnormal seafloor pressure and enhanced ocean current velocity oscillations. The gray shaded area from 24 hours to 0 hours before the earthquake indicates a time window with high prediction uncertainty, suggesting that technicians need to increase the density of observations during this period.

[0196] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0197] Table 2. Variable Explanation Table (Part 1)

[0198]

[0199] Table 3. Variable Explanation Table (Part Two)

[0200]

[0201] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A marine disaster early warning method based on artificial intelligence, characterized in that, Includes the following steps: Construct a four-level time length series and a four-level spatial scale grid, and set the time length adjustment factor and the scale adjustment factor; Marine observation sensors are deployed at grid nodes to collect marine dynamic parameters, and time and spatial normalization processing is performed. By calling the ultra-sparse cross-scale coupling matrix and the micro-feature time-series dependency matrix, the normalized marine dynamic parameters are input into the multi-scale coupling model of marine disasters for cross-scale feature extraction, resulting in a scale-separated feature vector group. Calculate the rank of the ultra-sparse cross-scale coupling matrix and perform scale reduction or scale expansion. Map the scale-separated feature vector group to a high-dimensional feature space through the radial basis kernel function. The feature vector is input into the multi-scale coupling model of marine disasters. The sparsity parameter of the sparse attention mechanism is determined according to the duration adjustment factor, the scale adjustment factor and the proportion of non-zero elements of the ultra-sparse cross-scale coupling matrix, and the marine disaster prediction sequence is output. The marine disaster prediction sequence is integrated and forecasted, and the prediction uncertainty interval is calculated. When the width of the prediction uncertainty interval exceeds a preset threshold, the observation density is increased and the above steps are repeated. When the width of the prediction uncertainty interval is less than the preset threshold, marine disaster early warning information is generated and output.

2. The method according to claim 1, characterized in that, The establishment of the four-level time length series specifically involves setting the time length range of the hourly short-term series to 1h to 24h, the daily medium-term series to 24h to 168h, the weekly long-term series to 168h to 720h, and the monthly ultra-long-term series to 720h to 2160h.

3. The method according to claim 2, characterized in that, Time sampling points are set at logarithmic intervals within each time length range. The logarithmic interval is obtained by dividing the upper limit of the time length range by the lower limit, taking the natural logarithm, and then dividing by the number of sampling points minus 1.

4. The method according to claim 3, characterized in that, The establishment of the four-level spatial scale grid is specifically defined as follows: the grid size of the 100-meter micro-scale grid is set to 100m, the kilometer-scale meso-scale grid is set to 1000m, the 10-kilometer-scale large-scale grid is set to 10000m, and the 100-kilometer-scale giant-scale grid is set to 100000m.

5. The method according to claim 4, characterized in that, 100-meter-scale micro-grids cover key near-shore protected areas, 1-kilometer-scale meso-scale grids cover coastal economic zones, 10-kilometer-scale large-scale grids cover continental shelf areas, and 100-kilometer-scale mega-scale grids cover deep-sea and open ocean areas.

6. The method according to claim 5, characterized in that, The determination of the duration adjustment factor specifically involves setting the baseline time length as 1 hour for hourly short-term sequences, 24 hours for daily medium-term sequences, 168 hours for weekly long-term sequences, and 720 hours for monthly ultra-long-term sequences.

7. The method according to claim 6, characterized in that, Obtain the target predicted time length, calculate the ratio of the target predicted time length to the baseline time length of each sequence as the original time length ratio, perform a logarithmic transformation on the original time length ratio and map it to the interval [0, 1] through a linear transformation to obtain the time length adjustment factor.

8. The method according to claim 7, characterized in that, The determination of the scale adjustment factor specifically involves setting the baseline grid size to 100m for the 100-meter micro-scale grid, 1000m for the kilometer-scale meso-scale grid, 10000m for the 10-kilometer-scale large-scale grid, and 100000m for the 100-kilometer-scale giant-scale grid.

9. The method according to claim 8, characterized in that, The ratio of the actual grid size to the reference grid size at each level is calculated as the original scale ratio. The original scale ratio is then logarithmically transformed and mapped to the interval [0, 1] through a linear transformation to obtain the scale adjustment factor.

10. The method according to claim 9, characterized in that, The establishment of the ultra-sparse cross-scale coupling matrix is ​​specifically defined as follows: for grid nodes at adjacent scales, if the spatial distance between two grid nodes is less than 50% of the size of the larger scale grid, the two grid nodes are considered to have a coupling relationship.

Images