A method and system for rapid identification of dike hazards

Abstract

This invention relates to the field of safety monitoring technology, specifically disclosing a method and system for rapid identification of dike hazards. The method involves deploying a multi-source sensor network to collect multimodal data in real time; extracting Lyapunov index spectrum, correlation dimension, and Kolmogorov entropy from phase space trajectories to form a structural response chaos index; calculating hydrogeological coupling coefficients by combining multi-scale decomposition and mutual information analysis; and fusing these two factors into a three-dimensional hazard feature tensor, which is then input into a pre-trained model based on a deep convolutional neural network and a long short-term memory network to achieve intelligent discrimination of high, medium, and low risk levels. For low-risk conditions, adaptive monitoring instructions are generated, and a risk evolution trend map is output, supporting closed-loop management.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring technology, specifically to a method and system for rapid identification of the dangers of dike hazards. Background Technology

[0002] As a critical barrier in flood control systems, the structural integrity and stability of dikes directly impact watershed safety and the protection of public life and property. Existing dike safety monitoring technologies primarily employ single-source or discretely deployed sensor systems, such as water level gauges, settlement meters, and piezometers. These mainly acquire local, static, or semi-empirical parameters, limiting the monitoring dimensions and hindering comprehensive perception of the multi-physics coupled response of the dike itself and its underlying geological structure under complex hydrological loads. Although some projects have introduced new sensing methods such as distributed fiber optics, microseismic monitoring, and GNSS in recent years, forming preliminary multimodal data acquisition capabilities, most systems remain at the rudimentary stage of "data aggregation + threshold alarm," lacking the ability to model and analyze the inherent nonlinear correlation mechanisms between multi-source heterogeneous data. Especially for hidden dangers caused by internal seepage erosion, progressive soil damage, and clusters of micro-fractures, traditional methods, relying on linear trend extrapolation or empirical criteria, often fail to capture the nonlinear dynamic mutation characteristics before system instability, leading to delayed or misjudgments in early warning. In addition, existing risk assessment models are mostly based on static indicators or simplified mechanical assumptions, failing to fully integrate the chaotic evolution characteristics of structural dynamic response with the multi-scale spatiotemporal coupling effect of environmental factors. As a result, the assessment results lack dynamic adaptability and physical mechanism support, making it difficult to meet the actual needs of high-precision, proactive risk identification.

[0003] The existing technology has the following shortcomings:

[0004] Existing levee hazard identification technologies generally suffer from insufficient modeling of the co-evolution mechanism between the system's nonlinear dynamic behavior and environmental disturbances. In particular, they struggle to quantify the dynamic correlation between structural chaotic response characteristics and hydrogeological coupling effects, thus failing to effectively identify precursors of nonlinear abrupt instability caused by the accumulation of minor disturbances. Specifically, traditional methods typically treat structural deformation response separately from environmental parameters such as pore water pressure and seepage field, performing only statistical correlation analysis or simple superposition judgments. This fails to construct a comprehensive dynamic index system that reflects the overall instability trend of the system. Consequently, when facing hazards with strong nonlinear and critical abrupt characteristics, such as piping initiation, internal void expansion, and the hidden development of slip surfaces, it is difficult to identify predictable dynamic anomalies from normal fluctuations. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for rapid identification of the dangers of dike hazards, so as to solve the problems mentioned above.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for rapid identification of the danger of dike hazards includes the following steps:

[0008] S1: Real-time acquisition of multimodal sensor data of the dike;

[0009] S2: Perform nonlinear dynamic response analysis on multimodal sensing data, construct the chaotic characteristic spectrum of the embankment structure state, and calculate the chaotic index of the structural response.

[0010] S3: Perform multi-scale correlation analysis on the environmental coupling effect in multimodal sensing data, construct the seepage strain field of the geological body, and calculate the hydrogeological coupling coefficient;

[0011] S4: The structural response chaos index and hydrogeological coupling coefficient are fused into a hazard feature tensor. The hazard feature tensor is then input into a pre-trained hazard assessment model to output the risk level of the dike, including high risk level, medium risk level and low risk level.

[0012] S5: Generates continuous monitoring instructions for low-risk levels and provides risk evolution trend maps.

[0013] As a further aspect of the present invention: the nonlinear dynamic response analysis of the multimodal sensing data specifically includes:

[0014] Based on phase space reconstruction theory, dynamic system reconstruction is performed on multimodal sensing data sequences. One-dimensional monitoring data is mapped to a high-dimensional phase space to restore the intrinsic dynamic characteristics of the embankment system. The average divergence rate of adjacent trajectories in the reconstructed phase space is calculated. The state change mode in the dynamic system is identified, and the recursive characteristic changes in the phase space trajectory are detected. Combined with the vibration spectrum characteristics of the embankment structure, a set of characteristic parameters representing the nonlinear dynamic state of the system is extracted.

[0015] As a further aspect of the present invention: the construction of the chaotic feature spectrum of the dike structure state and the calculation of the chaotic index of the structural response specifically include:

[0016] The set of feature parameters is fused using multi-scale entropy to form a chaotic feature spectrum that includes the Lyapunov feature index spectrum, the correlation dimension distribution, and the Kolmogorov entropy value; the deviation between the current state features and the levee stability benchmark state in the feature space is calculated; and finally, the structural response chaos index that comprehensively reflects the degree of instability of the levee structure system is output.

[0017] As a further aspect of the present invention: the multi-scale correlation analysis of the environmental coupling effect in multimodal sensing data specifically includes:

[0018] The seepage pressure data and strain monitoring data are decomposed into trend, periodic and random components, respectively; the nonlinear correlation strength between hydrological parameters and geological deformation parameters at different time scales is calculated; the coupling characteristic regions of seepage field and strain field in spatial distribution are identified by wavelet coherence analysis; and a multi-scale correlation entropy index is established to quantify the degree of coupling and synergy of hydrogeological system at different temporal and spatial scales.

[0019] As a further aspect of the present invention: the process for obtaining the hydrogeological coupling coefficient is as follows:

[0020] Based on multi-source monitoring data, a three-dimensional spatial distribution cloud map of seepage pressure field and soil strain field is constructed. The pore water pressure monitoring data and distributed strain monitoring data are spatially interpolated and fused. By calculating the dot product of hydraulic gradient vector and strain gradient vector on the seepage path, the historical monitoring data is weighted and processed. Finally, the hydrogeological coupling coefficient is calculated based on the amplitude and duration of the coordinated change characteristics.

[0021] As a further aspect of the present invention: the process of obtaining the hazard feature tensor is as follows:

[0022] A three-dimensional hazard feature tensor containing spatial distribution characteristics is constructed with time series as the third dimension. The first dimension stores the distribution data of the structural response chaos index at different dike sections, and the second dimension stores the distribution data of the hydrogeological coupling coefficient at different depths. Tensor decomposition is used to extract the spatiotemporal coupling modes of hazard features to obtain the core feature tensor. The feature tensors are weighted and fused to finally generate a hazard feature tensor containing spatiotemporal correlation characteristics, which serves as the input to the hazard assessment model.

[0023] As a further aspect of the present invention: the construction process of the risk assessment model is as follows:

[0024] A training sample set is constructed based on historical hazard case data. The input is a hazard feature tensor, and the output is a risk level label labeled by experts. A hybrid architecture of deep convolutional neural network and long short-term memory network is used as the basic model. Spatial features are extracted by multi-scale convolutional kernels, and temporal dependencies are captured by memory network. The two tasks of risk level classification and hazard type identification are optimized simultaneously. A hierarchical loss function is used for model training, and a focal loss function is combined to deal with the sample imbalance problem. Finally, the risk level of the levee hazard is output.

[0025] As a further aspect of the present invention: the step of generating continuous monitoring instructions for low-risk levels and providing a risk evolution trend map specifically includes:

[0026] The evolution of the tensor characteristics of low-risk hazards is analyzed to generate continuous monitoring instructions that include dynamic adjustment schemes for monitoring frequency and key monitoring areas. Combined with prior knowledge of the mechanism of levee instability, a multi-dimensional evolution trend map is constructed that includes risk probability curves in the time dimension and risk propagation paths in the spatial dimension. At the same time, an optimization scheme for priority monitoring based on the trend map is provided to form a closed-loop risk management system.

[0027] A rapid identification system for the danger of dike hazards includes:

[0028] A multi-source sensor data acquisition module, which is used to acquire multimodal sensor data of the dike in real time;

[0029] The structural response chaos analysis module performs nonlinear dynamic response analysis on multimodal sensing data, constructs a chaotic characteristic spectrum of the embankment structure state, and calculates the structural response chaos index.

[0030] The hydrogeological coupling factor extraction module performs multi-scale correlation analysis on the environmental coupling effect in multimodal sensing data, constructs the seepage strain field of the geological body, and calculates the hydrogeological coupling coefficient.

[0031] The risk situation collaborative assessment module integrates the structural response chaos index and the hydrogeological coupling coefficient into a hazard feature tensor. The hazard feature tensor is input into a pre-trained hazard assessment model, and the risk level of the dike is output, including high risk level, medium risk level and low risk level.

[0032] The dynamic response module for risk levels generates continuous monitoring instructions for low-risk levels and provides a risk evolution trend map.

[0033] The beneficial effects of this invention are:

[0034] (1) By acquiring multimodal sensor data in real time and performing nonlinear dynamic response analysis and multi-scale correlation analysis, this invention can comprehensively capture changes in the coupling effect between the structural state of the dike and the environment. This method can not only reveal early signs of the initiation or expansion of damage inside the dike, but also quantitatively assess the impact of changes in hydrogeological conditions on the stability of the dike. Therefore, compared with traditional monitoring methods, this invention can identify potential risks earlier and more accurately, buying valuable time for taking preventive measures.

[0035] (2) This invention integrates the structural response chaos index with the hydrogeological coupling coefficient into a hazard characteristic tensor, and uses a pre-trained hazard assessment model to output the risk level of the dike. It also provides continuous monitoring instructions and a risk evolution trend map, enabling managers to intuitively understand the risk status and development trend of the dike. This data-driven decision support system helps optimize the allocation of monitoring resources and formulate scientific and reasonable emergency plans, thereby improving the management efficiency and emergency response capabilities of the entire dike system. Attached Figure Description

[0036] The invention will now be further described with reference to the accompanying drawings.

[0037] Figure 1 This is a flowchart of the method of the present invention;

[0038] Figure 2 This is a flowchart of the system in this invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figure 1 As shown, this invention provides a method for rapid identification of the danger of dike hazards, comprising the following steps:

[0041] S1: Real-time acquisition of multimodal sensor data of the dike;

[0042] S2: Perform nonlinear dynamic response analysis on multimodal sensing data, construct the chaotic characteristic spectrum of the embankment structure state, and calculate the chaotic index of the structural response.

[0043] S3: Perform multi-scale correlation analysis on the environmental coupling effect in multimodal sensing data, construct the seepage strain field of the geological body, and calculate the hydrogeological coupling coefficient;

[0044] S4: The structural response chaos index and hydrogeological coupling coefficient are fused into a hazard feature tensor. The hazard feature tensor is then input into a pre-trained hazard assessment model to output the risk level of the dike, including high risk level, medium risk level and low risk level.

[0045] S5: Generates continuous monitoring instructions for low-risk levels and provides risk evolution trend maps.

[0046] In S1, multimodal sensor data of the dike is collected in real time, specifically including:

[0047] By deploying a multi-source sensor network in the embankment itself and its underlying geological structure, multi-modal monitoring data reflecting the coupling effect between the embankment structure and the environment can be acquired in real time. The multimodal sensing data mainly includes four categories: The first category is distributed optical fiber sensing data, which uses optical fiber sensors embedded in the interior and surface of the dike to continuously measure the strain distribution and temperature field changes along the optical fiber path, with a spatial resolution of up to the meter level, capable of capturing the temperature change effect caused by uneven deformation and seepage inside the dike; the second category is microseismic monitoring data, which is collected by an array of microseismic sensors arranged on key sections of the dike, recording the elastic wave signals generated by micro-fractures inside the dike foundation soil and rock, used to infer the damage evolution process inside the geological body; the third category is pore water pressure data, which is monitored by piezometers installed at different depths in the dike body and foundation to obtain real-time information on groundwater level fluctuations and seepage field changes, reflecting the impact of hydraulic boundary conditions on dike stability; the fourth category is surface displacement data, which uses ground-based synthetic aperture radar or global navigation satellite system receivers to monitor the two-dimensional or three-dimensional deformation field of the dike surface and surrounding area with millimeter-level accuracy, revealing the overall deformation trend and signs of local instability. All sensing units are connected to the data acquisition platform via industrial IoT nodes. The acquisition terminals employ synchronous triggering or periodic polling mechanisms based on sensor characteristics to ensure data consistency over time. After analog-to-digital conversion, the raw data, along with device identifiers, timestamps, and coordinate information, is formed into standardized data packets and transmitted to the central data processing server via wired or wireless networks. This step enables multi-dimensional, high-frequency perception of the physical state of the dike system, providing a complete and reliable raw data foundation for subsequent hazard identification and analysis.

[0048] The acquisition of multimodal data covers key parameters for safety monitoring of dike projects, forming a three-dimensional sensing system. Distributed fiber optic sensors are typically deployed along the dike axis, focusing on sections with historical risks, complex geological conditions, and areas with strong hydrodynamic forces. Installation methods include direct burial and surface bonding. These sensors monitor both axial tensile and compressive strain and improve strain measurement accuracy through temperature correction. The arrangement of the microseismic monitoring array must consider background noise interference and wave propagation characteristics. Sensors are generally distributed in a grid or linear pattern near the dike toe and potential sliding surfaces, with sampling frequencies set according to the characteristics of the soil and rock mass. Pore water pressure monitoring points are deployed at seepage outlets, potential seepage paths, and groundwater level fluctuation zones. These monitoring points form an observation profile vertically and cover the dike body, foundation, and the affected area on the backwater side horizontally, characterizing the spatiotemporal distribution and hysteresis effects of pore water pressure. Surface displacement monitoring equipment, such as radar interferometers, is fixedly installed at stable reference points along the dike, scanning the dike surface at regular revisit intervals to generate high-precision deformation point cloud data. Meanwhile, satellite positioning receivers operate in continuous reference station mode, providing absolute displacement sequences. All sensor data undergoes preliminary quality checks during acquisition, including signal strength verification, data packet integrity checks, and outlier marking, ensuring a continuous and reliable data stream input to subsequent analysis modules.

[0049] In S2, nonlinear dynamic response analysis is performed on multimodal sensing data to construct a chaotic characteristic spectrum of the levee structure state and calculate the chaotic index of the structural response, specifically including:

[0050] The specific implementation process for nonlinear dynamic response analysis of multimodal sensing data is as follows. First, the dynamic system of the univariate monitoring sequence is reconstructed using phase space reconstruction theory. Taking deep displacement monitoring data of a certain cross-section as an example, its time series is assumed to contain n monitoring values ​​arranged in chronological order. By calculating the delay time parameter and embedding dimension of the sequence, the one-dimensional sequence is mapped to a high-dimensional phase space, resulting in a phase point vector composed of multiple time delay values. The delay time is determined using the autocorrelation function method, and the optimal delay time is taken as the time point when the autocorrelation function of the sequence first decays to approximately 36.8% of its initial value. The embedding dimension is determined using the spurious nearest neighbor method; the minimum dimension corresponding to when the proportion of spurious nearest neighbors in the reconstructed phase space is less than 5% is the appropriate embedding dimension. Through this reconstruction, the one-dimensional observation sequence is transformed into a phase space trajectory that reflects the inherent dynamic characteristics of the system.

[0051] In the reconstructed phase space, the average divergence rate of adjacent trajectories, i.e., the maximum Lyapunov exponent, is calculated. Specifically, a small-data-volume method is used: a reference point is selected in the phase space, its nearest neighbor is found, the initial distance between these two points is calculated, and then the distance between these two phase points is tracked over time. By performing a linear regression of the logarithm of the distances between all phase points relative to time, the slope of the resulting regression line is the estimated value of the maximum Lyapunov exponent. Simultaneously, a recursive graph analysis method is used to identify state change patterns: a recursive matrix is ​​calculated, where each element takes the value 1 or 0, depending on whether the distance between the corresponding two phase points is less than a preset threshold. By analyzing the diagonal structure and the distribution pattern of recursive points in the recursive graph, changes in recursive characteristics appearing in the phase space trajectory are detected. Combined with the vibration spectrum characteristics acquired by distributed fiber optic sensing, the nonlinear characteristics of the system are analyzed from a frequency domain perspective, ultimately extracting a set of characteristic parameters including the maximum Lyapunov exponent, recursion rate, determinism, and laminarity.

[0052] The specific process for constructing the chaotic feature spectrum of the dike structure state is as follows. The feature parameter set obtained in the preceding steps undergoes multi-scale entropy fusion processing. First, the time series of each feature parameter is multi-scaled: given a scale factor, the original sequence is divided into multiple non-overlapping windows of equal length, and the arithmetic mean of the data within each window is calculated to form a coarse-grained sequence. Then, the sample entropy value of the coarse-grained sequence at each scale is calculated. The sample entropy is calculated as follows: for a sequence of a given length, the pattern dimension and similarity tolerance are set, and the number of all m-dimensional pattern vector pairs satisfying the distance condition is counted. Then, the number of m+1-dimensional pattern vector pairs satisfying the distance condition is counted. The sample entropy is the negative natural logarithm of the ratio of these two numbers. Through this multi-scale entropy analysis, the complexity measure of each feature parameter at different time scales is obtained.

[0053] Based on this, a chaotic feature spectrum comprising three core indicators is constructed: First, the Lyapunov characteristic index spectrum, which, in addition to the maximum Lyapunov index, calculates the entire Lyapunov index spectrum using matrix factorization, reflecting the average divergence or convergence rate in each direction of the phase space; second, the correlation dimension distribution, calculated using a distance statistical algorithm: in the reconstructed phase space, the proportion of phase pairs with a distance less than a given radius to the total number of phase pairs is calculated, and the correlation dimension estimate is obtained by the slope of the linear region of this proportion relative to the radius in the double logarithmic graph; third, the Kolmogorov entropy, estimated by calculating the correlation integral ratio under different embedding dimensions, reflecting the system information loss rate. These indicators are then weighted and fused with multi-scale entropy values ​​according to their importance in levee stability assessment to form a complete chaotic feature spectrum.

[0054] The calculation process of the structural response chaos index includes the following steps: First, establish the baseline stability state of the embankment: Select monitoring data of the embankment during normal operation and a stable hydrogeological condition stage, and calculate the baseline values ​​of each parameter in its chaotic characteristic spectrum, including the average value and covariance matrix of each parameter. For the current state, calculate its characteristic parameter vector. Use weighted Mahalanobis distance to calculate the deviation between the current state and the baseline state. This distance calculation considers the weight of each characteristic parameter, which is determined according to the degree of influence of the parameter on the stability of the embankment, and is calculated through the pairwise comparison matrix method. Considering the differences in the material properties and structural types of the embankment project, material correction coefficients and structural correction coefficients are introduced to correct the distance values. The material correction coefficient is determined according to the nonlinear characteristics of the embankment filling material, with a value of 1.0-1.2 for cohesive soil and 0.8-1.0 for sandy soil; the structural correction coefficient is determined according to the cross-sectional type of the embankment, with a value of 1.0 for homogeneous embankments, 0.9 for inclined wall embankments, and 0.85 for core wall embankments. Finally, the corrected distance is mapped to the 0-1 interval using a sigmoid function to obtain the structural response chaos index. The closer the index value is to 1, the higher the degree of instability of the levee structure system, providing a quantitative basis for subsequent risk level determination.

[0055] In S3, multi-scale correlation analysis is performed on the environmental coupling effects in multimodal sensing data to construct the seepage strain field of the geological body and calculate the hydrogeological coupling coefficient, specifically including:

[0056] The specific implementation process of multi-scale correlation analysis of environmental coupling effects in multimodal sensing data is as follows. First, the seepage pressure data and strain monitoring data are decomposed into time series components. Using a method based on local characteristic scale decomposition, the seepage pressure time series data for each monitoring point is decomposed into three components: a trend component, a periodic component, and a random component. The trend component reflects the long-term variation of seepage pressure and is extracted using a moving average method. The moving average window length is set to 30 days based on hydrogeological conditions. The periodic component includes components affected by periodic factors such as rainfall and tides, and the main periodic components are identified through spectral analysis. The random component characterizes noise fluctuations and abnormal changes in the monitoring data. Similarly, the strain monitoring data undergoes the same decomposition process to obtain the corresponding three components. After component decomposition, the nonlinear correlation strength between hydrological parameters and geological deformation parameters at different time scales is calculated. A mutual information calculation method based on probability density estimation is used to calculate the mutual information values ​​of the seepage pressure and strain components in the trend, periodic, and random components, respectively. The specific calculation process is as follows: First, the data sequences of the two components are binned, and the appropriate number of bins is determined based on the data distribution characteristics; then, the number of data points in each bin is counted, and the joint probability distribution and marginal probability distribution are calculated; finally, the correlation strength between the two components is calculated according to the definition of mutual information. Using this method, the coupling strength index between the seepage field and the deformation field at different time scales can be obtained.

[0057] Based on time-domain correlation analysis, wavelet coherence analysis is used to identify the spatially distributed coupling characteristic regions of the seepage field and strain field. Morlet wavelets are selected as the basis function to perform two-dimensional wavelet transforms on the spatially distributed seepage pressure and strain field data. Specifically, the monitoring area is divided into regular grid points, and continuous wavelet transforms are performed on the seepage pressure and strain time series at each grid point to obtain wavelet coefficients. Then, the wavelet coherence spectra of the two fields are calculated: first, the cross-power spectrum of the wavelet coefficients of the seepage pressure and strain fields is calculated; then, their respective self-power spectra are calculated; finally, the cross-coherence coefficient is calculated. A threshold of 0.7 is set for the cross-coherence coefficient. When the cross-coherence coefficient of a region exceeds this threshold and the phase relationship is stable, that region is considered a characteristic region of seepage-strain coupling. By analyzing the distribution of cross-coherence coefficients at different spatial scales, coupling regions at multiple scales from local to global are identified. Simultaneously, wavelet phase difference analysis is used to determine the sequence and time-delay relationship of seepage and strain changes, providing a basis for judging the coupling mechanism.

[0058] A multi-scale correlation entropy index is established to quantify the coupling and synergy of hydrogeological systems at different spatiotemporal scales. First, spatiotemporal scale parameters are defined: time scales are divided into short-term (1-7 days), medium-term (7-30 days), and long-term (over 30 days); spatial scales are divided into local (single monitoring point), regional (monitoring point group), and global (entire dike section). For each spatiotemporal scale combination, the joint distribution entropy and conditional distribution entropy of the seepage pressure field and strain field are calculated. The specific calculation steps are as follows: the data of the two fields are standardized, and an appropriate number of partitions is determined based on the data distribution range; the probability of the two field data falling within each partition combination is calculated, and the joint distribution entropy is calculated; the conditional probability of the strain field data given the seepage pressure field data is calculated, and the conditional distribution entropy is calculated; finally, the mutual information entropy of the two fields is calculated as the correlation entropy value at that scale. By comparing the correlation entropy values ​​at different spatiotemporal scales, the coupling and synergy of the hydrogeological system can be quantified; a higher entropy value indicates a more complex system coupling and a lower degree of synergy.

[0059] The process of obtaining the hydrogeological coupling coefficient includes the following steps: First, a three-dimensional spatial distribution cloud map of the seepage pressure field and the soil-rock strain field is constructed based on multi-source monitoring data. Spatial interpolation is first performed on the pore water pressure monitoring data using the Kriging interpolation method, considering the spatial correlation and variability of the monitoring points to generate a continuous three-dimensional pore water pressure distribution field. During the interpolation process, a suitable variogram model is determined based on the density and distribution characteristics of the monitoring points to ensure the rationality of the interpolation results. Similarly, spatial interpolation is performed on the distributed strain monitoring data to generate a three-dimensional strain distribution field. Both fields are discretized using a unified spatial grid, with the grid size determined according to the monitoring point density, typically set to 1 / 2 to 1 / 3 of the average spacing between monitoring points.

[0060] The synergistic relationship between seepage and soil deformation is quantified by calculating the dot product of the hydraulic gradient vector and the strain gradient vector along the seepage path. First, the hydraulic gradient vector is calculated on a three-dimensional grid: for each grid point, the hydraulic gradient components in three directions are calculated using the central difference method based on the pore water pressure values ​​of its adjacent grid points. The strain gradient vector is calculated in the same way. Then, the dot product of the two vectors is calculated, reflecting the degree of synergy between seepage and deformation. A positive value indicates that the two change in the same direction, resulting in enhanced synergy; a negative value indicates that they change in opposite directions, resulting in mutual inhibition. Historical monitoring data is weighted using an exponential decay weighting method, with recent data having a larger weight and historical data weight decreasing over time. The decay coefficient is determined based on the memory characteristics of the embankment material, typically set to 0.95, meaning that the weight of each day's data decreases to 95% of the previous day's weight.

[0061] Finally, the hydrogeological coupling coefficient is calculated based on the amplitude and duration of the coordinated change characteristics. First, the standard deviation of the dot product operation is calculated as a measure of the amplitude of the coordinated change. Then, the duration of continuous coordinated change is calculated, defined as the length of time during which the dot product values ​​maintain the same sign and the absolute value exceeds a threshold. The amplitude and duration indices are normalized using a min-max normalization method to map them to the range of 0-1. The hydrogeological coupling coefficient is obtained by multiplying the amplitude index by a weight of 0.6 and the duration index by a weight of 0.4. The hydrogeological coupling coefficient has a value range of 0-1; a larger value indicates a stronger coupling effect between seepage and deformation, and a higher risk of system instability.

[0062] In S4, the structural response chaos index and hydrogeological coupling coefficient are fused into a hazard feature tensor. This hazard feature tensor is then input into a pre-trained hazard assessment model, which outputs the risk level of the dike, including high-risk, medium-risk, and low-risk levels. Specifically, these levels include:

[0063] The construction of the hazard feature tensor first defines a three-dimensional data structure based on the spatial structure of the dike project and the monitoring deployment plan. The first dimension of the hazard feature tensor represents different cross-sections of the dike, arranged in order of actual mileage markers, with each cross-section storing the calculated structural response chaos index. The second dimension represents the vertical depth direction, storing the hydrogeological coupling coefficients at different depths based on geological stratification and sensor deployment. The third dimension is the time series, continuously storing feature data at different time steps according to the monitoring cycle. In this way, the dike state at a single point in time is expressed as a two-dimensional feature matrix, while the state changes over continuous time periods constitute a three-dimensional tensor.

[0064] After obtaining the original three-dimensional feature tensor, tensor decomposition technology is used to extract its inherent spatiotemporal coupling modes. Specifically, the Tucker decomposition algorithm is selected for dimensionality reduction and feature extraction of the original tensor. This decomposition divides the original tensor into a product of a core tensor and three factor matrices. The core tensor contains the main coupling relationships between features, while the three factor matrices correspond to the feature patterns in the cross-sectional, depth, and time dimensions, respectively. By retaining the highest-energy components in the core tensor and discarding lower-energy noise components, the risk features are purified and compressed, resulting in a core feature tensor that reflects the essential risk characteristics of the dike system. To further enhance the expressive power of key features, a spatiotemporal attention mechanism is introduced to weightedly fuse the core feature tensor. This mechanism adaptively weights feature values ​​at different cross-sections, depths, and time points using a trained attention weight matrix, highlighting the feature contributions of areas and time periods with significant changes and high correlation to risks. This ultimately generates an enhanced risk feature tensor, which serves as the input to the subsequent risk assessment model.

[0065] The construction of the hazard assessment model is a supervised learning process based on historical data. First, a high-quality training sample set needs to be built, collecting historical hazard case data. This includes hazard feature tensors calculated under various working conditions as input features, and risk level labels marked by domain experts based on actual hazard occurrences as output targets. Risk levels are divided into three levels: high risk corresponds to working conditions with obvious signs of instability or localized damage, requiring immediate early warning; medium risk corresponds to working conditions with abnormal changes but not yet posing a direct danger, requiring increased monitoring frequency; and low risk corresponds to working conditions within the normal fluctuation range. The sample set must cover typical cases from different hydrological seasons, different geological conditions, and different engineering locations to ensure the model's generalization ability.

[0066] The model employs a hybrid architecture combining deep convolutional neural networks (CNNs) and long short-term memory (LSTM) networks. The CNN portion handles the spatial dimension information of the feature tensors, using multiple convolutional kernels of different scales to perform convolution operations in both the cross-sectional and depth dimensions, extracting local spatial features and capturing spatial hierarchical patterns. The LTM portion processes the temporal sequence, learning the long-term dependencies of features over time through its gating mechanism to capture the evolutionary trends of risk situations. The outputs of the two networks are fused at a high-level feature layer and jointly input into a fully connected classification layer. The model adopts a multi-task learning framework, simultaneously optimizing two related tasks: risk level classification and hazard type identification. By sharing a low-level feature extraction layer and setting separate task-specific output layers, the model's feature representation ability and generalization performance are improved.

[0067] The model training employs a strategy combining hierarchical loss and focus loss. The hierarchical loss treats risk level classification as an ordered classification problem, penalizing deviations between predicted and true risk levels based on the degree of risk, rather than simply differences in class labels. The focus loss addresses the imbalance in the number of samples at different risk levels by reducing the loss contribution of easily classified samples, thus focusing the model training on harder-to-classify samples and improving the ability to identify minority classes (such as high-risk samples). During training, backpropagation is used to optimize model parameters, adaptive moment estimation is used to adjust the learning rate, and early stopping is employed to prevent overfitting, ultimately resulting in a risk assessment model with good generalization ability. This model can receive real-time generated risk feature tensors and output the corresponding risk level probability distribution, providing direct evidence for subsequent early warning decisions.

[0068] In S5, continuous monitoring instructions are generated for low-risk levels, and risk evolution trend maps are provided, specifically including:

[0069] For work conditions assessed as low-risk, the system generates targeted continuous monitoring instructions. These instructions are based on the analysis of recent hazard characteristic tensor time-series data, employing time-series prediction algorithms to identify the patterns and trends in the changes of each characteristic parameter. Specifically, stability is assessed by calculating the moving average and rate of change of each element in the characteristic tensor over time. If the parameter changes are stable and within normal fluctuation ranges, the monitoring frequency is appropriately reduced, such as adjusting the acquisition interval for specific points from once per minute to once every ten minutes. If certain parameters are identified as showing a continuous, slow upward trend even though they have not exceeded thresholds, the monitoring areas corresponding to these parameters are subject to focused monitoring, and these key areas requiring continued monitoring or even enhanced monitoring are clearly marked in the instructions. The monitoring instructions are output in a structured data format, including the monitoring point number, adjusted acquisition frequency, execution time window, and other specific, actionable instructions, and are automatically sent to the on-site acquisition equipment for execution via a data interface.

[0070] While generating monitoring instructions, the system constructs a multi-dimensional risk evolution trend map to support medium- and long-term decision-making. The map's generation combines data-driven and mechanism-driven approaches: First, based on historical and current hazard characteristic tensor data, a deep spatiotemporal neural network model is used to make multi-step predictions for the future, outputting risk probability change curves for the overall dike and key components over a future period. These curves, with time on the horizontal axis and risk occurrence probability on the vertical axis, visually demonstrate the risk development trend. Second, prior knowledge of the physical mechanisms of dike instability is incorporated, such as the expansion path of seepage failure or the formation mechanism of the slip surface, simulating the spatial development path of potential risks. The most likely propagation direction and impact range of the risk are marked on the map using arrows or contour lines, forming a spatial risk propagation path map. The final risk evolution trend map is a comprehensive visualization product integrating time-dimensional prediction curves and spatial-dimensional propagation paths.

[0071] Based on the aforementioned trend graphs, the system further provides suggestions for monitoring optimization and contingency plan activation priorities, forming a closed-loop management system. By analyzing the upward slope of the risk probability curve and key nodes of the spatial propagation path in the trend graphs, the system identifies shortcomings in the current monitoring deployment plan, such as which areas have predicted increased risk but insufficient monitoring coverage. This leads to the generation of optimized monitoring point deployment plans or suggestions for adjusting the existing monitoring frequency. Simultaneously, based on the predicted risk arrival time and impact range, and in accordance with the activation conditions in the emergency plan database, the system calculates an activation priority index for different plans. This index is weighted by factors such as the predicted risk level, expected occurrence time, and the importance of the affected area, and corresponding response measures are suggested. All analysis results and post-instruction effect data are re-collected and fed back into the hazard assessment model and trend prediction model for self-learning and optimization, thereby achieving closed-loop risk management of monitoring, assessment, prediction, early warning, and feedback, continuously improving the system's early warning accuracy and decision support capabilities.

[0072] Please see Figure 2 As shown, a rapid identification system for the risk of dike hazard includes:

[0073] A multi-source sensor data acquisition module, which is used to acquire multimodal sensor data of the dike in real time;

[0074] The structural response chaos analysis module performs nonlinear dynamic response analysis on multimodal sensing data, constructs a chaotic characteristic spectrum of the embankment structure state, and calculates the structural response chaos index.

[0075] The hydrogeological coupling factor extraction module performs multi-scale correlation analysis on the environmental coupling effect in multimodal sensing data, constructs the seepage strain field of the geological body, and calculates the hydrogeological coupling coefficient.

[0076] The risk situation collaborative assessment module integrates the structural response chaos index and the hydrogeological coupling coefficient into a hazard feature tensor. The hazard feature tensor is input into a pre-trained hazard assessment model, and the risk level of the dike is output, including high risk level, medium risk level and low risk level.

[0077] The dynamic response module for risk levels generates continuous monitoring instructions for low-risk levels and provides a risk evolution trend map.

[0078] The working principle of this invention is as follows: Real-time acquisition of multimodal sensor data from the dike (such as distributed fiber optic sensor data, microseismic monitoring data, pore water pressure data, and surface displacement data); nonlinear dynamic response analysis to construct a chaotic characteristic spectrum of the dike structure state and calculate the structural response chaos index; multi-scale correlation analysis of environmental coupling effects to construct the seepage strain field of the geological body and calculate the hydrogeological coupling coefficient; fusion of these two into a hazard characteristic tensor input to a pre-trained hazard assessment model to output a risk level; and finally, generation of continuous monitoring instructions and provision of risk evolution trend maps for low-risk levels. Through this series of steps, multi-dimensional, high-frequency perception of the physical state of the dike system and rapid identification and assessment of hazards are achieved, providing scientific basis and technical support, and effectively improving dike safety management and emergency response capabilities.

[0079] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for quickly identifying the danger of embankment danger, characterized in that, The method comprises the following steps: S1: collecting multi-modal sensing data of the embankment in real time; S2: performing nonlinear dynamic response analysis on the multi-modal sensing data, constructing a chaotic characteristic spectrum of the embankment structure state, and calculating a structure response chaos degree index; S3: performing multi-scale correlation analysis on the environmental coupling effect in the multi-modal sensing data, constructing a seepage strain field of the geological body, and calculating a hydrogeological coupling variation coefficient, specifically comprising: decomposing seepage pressure data and strain monitoring data into trend items, periodic items and random item components respectively; calculating the nonlinear correlation strength between hydrological parameters and geological deformation parameters at different time scales; identifying the coupling characteristic region of the seepage strain field of the geological body in the spatial distribution through wavelet coherence analysis; and establishing a multi-scale correlation entropy index to quantify the coupling and coordination degree of the hydrogeological system at different time and space scales; based on the multi-source monitoring data, constructing a three-dimensional spatial distribution cloud map of the seepage strain field of the geological body, spatially interpolating and fusing the pore water pressure monitoring data and the distributed strain monitoring data; performing weighted processing on the historical monitoring data by calculating the dot product of the hydraulic gradient vector and the strain gradient vector on the seepage path; and finally calculating the hydrogeological coupling variation coefficient according to the amplitude and duration of the coordinated change characteristics; S4: fusing the structure response chaos degree index and the hydrogeological coupling variation coefficient into a danger feature tensor, inputting the danger feature tensor into a pre-trained danger assessment model, and outputting the risk level of the embankment, including high risk level, medium risk level and low risk level; S5: generating a continuous monitoring instruction for the low risk level and providing a risk evolution trend map.

2. The method for quickly identifying the danger of embankment according to claim 1, characterized in that, The nonlinear dynamic response analysis on the multi-modal sensing data specifically comprises: reconstructing the dynamic system based on the phase space reconstruction theory to map one-dimensional monitoring data to high-dimensional phase space, recover the inherent dynamic characteristics of the embankment system, calculate the average divergence rate of adjacent trajectories in the reconstructed phase space, identify the state mutation mode in the dynamic system, detect the recursive characteristic change in the phase space trajectory, and extract a feature parameter set representing the nonlinear dynamic state of the system in combination with the vibration frequency spectrum characteristics of the embankment structure.

3. The method according to claim 2, wherein, According to the feature parameter set, a chaotic characteristic spectrum of the embankment structure state is constructed, and a structure response chaos degree index is calculated, specifically comprising: performing multi-scale entropy fusion on the feature parameter set to form a chaotic characteristic spectrum containing Lyapunov characteristic exponent spectrum, correlation dimension distribution and Kolmogorov entropy value; calculating the deviation degree of the current state feature and the embankment stable benchmark state in the feature space; and finally outputting a structure response chaos degree index comprehensively reflecting the instability degree of the embankment structure system.

4. The method for quickly identifying the danger of embankment according to claim 1, characterized in that, The acquisition process of the danger feature tensor is as follows: a three-dimensional danger feature tensor containing spatial distribution characteristics is constructed with time series as the third dimension, wherein the first dimension stores the distribution data of the structure response chaos degree index at different embankment sections, and the second dimension stores the distribution data of the hydrogeological coupling variation coefficient at different depths; a time-space coupling mode of the danger feature is extracted by tensor decomposition to obtain a core feature tensor; The feature tensors are weighted and fused to finally generate a risk situation feature tensor containing spatio-temporal correlation characteristics, which is used as an input of the risk situation assessment model.

5. The method for quickly identifying the danger of embankment according to claim 1, characterized in that, The construction process of the risk situation assessment model is as follows: A training sample set is constructed based on historical risk situation case data, the input is a risk situation feature tensor, and the output is a risk level label annotated by an expert; a deep convolutional neural network and a long short-term memory network hybrid architecture are used as a basic model, spatial features are extracted through multi-scale convolution kernels, and time-dependent relationships are captured through a memory network; Two tasks of risk level classification and risk situation type identification are optimized synchronously; a hierarchical loss function is used for model training, and a focal loss function is used to deal with the sample imbalance problem; and finally, the risk level of the embankment risk situation is output.

6. The method for quickly identifying the danger of embankment according to claim 1, characterized in that, The generation of the continuous monitoring instruction for the low risk level and the provision of the risk evolution trend graph are as follows: The evolution law of the risk situation feature tensor of the low risk level is analyzed, a continuous monitoring instruction containing a monitoring frequency dynamic adjustment scheme and a key monitoring area direction is generated, a multi-dimensional evolution trend graph containing a time dimension risk probability curve and a space dimension risk propagation path is constructed based on the prior knowledge of the embankment instability mechanism, a preplan starting priority monitoring optimization scheme based on the trend graph is provided, and a closed-loop risk control strategy is formed.

7. A quick identification system for danger of embankment danger, characterized in that, A method for quickly identifying the danger of an embankment risk situation according to any one of claims 1-6, comprising: A multi-source sensing data acquisition module, which is used for real-time acquisition of multi-modal sensing data of the embankment; A structure response chaos degree analysis module, which performs nonlinear dynamic response analysis on the multi-modal sensing data, constructs a structure state chaos characteristic spectrum of the embankment, and calculates a structure response chaos degree index; A hydrogeological coupling factor extraction module, which performs multi-scale correlation analysis on the environmental coupling effect in the multi-modal sensing data, constructs a geological body seepage strain field, and calculates a hydrogeological coupling coefficient; A risk situation collaborative analysis module, which fuses the structure response chaos degree index and the hydrogeological coupling coefficient into a risk situation feature tensor, inputs the risk situation feature tensor into a pre-trained risk situation assessment model, and outputs a risk level of the embankment, including a high risk level, a medium risk level and a low risk level; A risk situation level dynamic response module, which generates a continuous monitoring instruction for the low risk level and provides a risk evolution trend graph.

Images