An intelligent seepage monitoring system for hydraulic engineering

Abstract

The application relates to the technical field of safety monitoring of water conservancy projects, and discloses a seepage intelligent monitoring system for water conservancy projects. The system comprises a parameter acquisition unit and a geological detection unit, which synchronously collect hydrological working condition data and geological radar scanning data. A feature generation unit fuses the two types of data to form an optimized multi-dimensional seepage feature set. A model construction unit uses the set to train an evaluation model and extracts a health state index. A real-time processing unit receives real-time operation parameters of the engineering structure, combines the evaluation model, and dynamically calculates a seepage health state index that is adaptive to the working condition. An early warning execution unit judges and triggers early warning according to the dynamic index and a preset risk mapping relationship. The application fuses multi-source heterogeneous data and realizes adaptive adjustment of the evaluation standard to the operation working condition, thereby improving the comprehensiveness of seepage risk evaluation and the accuracy of early warning.

Description

Technical Field

[0001] This invention relates to the field of water conservancy project safety monitoring technology, specifically to an intelligent seepage monitoring system for water conservancy projects. Background Technology

[0002] In the field of water conservancy project safety monitoring, seepage status is a key parameter for assessing the health and stability of structures such as dams and levees. Traditional seepage monitoring technologies mainly rely on single-type sensors deployed at critical locations of the engineering structure. These monitoring methods form the data foundation for seepage safety analysis, and their assessment logic typically depends on fixed thresholds calculated from historical data statistics or theoretical models. When the monitoring data exceeds a preset static threshold, the system triggers an alarm.

[0003] Existing technical solutions have shortcomings. Relying on single-point, single-type sensor data is insufficient to comprehensively reflect the true spatial state of the seepage field and its correlation with the aging and damage within the engineering structure. Seepage anomalies often develop in conjunction with geological defects such as internal cracks, dissolution, and contact zone deterioration. External hydrological response data alone cannot reveal their underlying mechanisms, leading to delayed or misjudgments in risk identification. Assessment methods using static thresholds or fixed models cannot adapt to the complex and variable hydrological conditions in actual water conservancy project operations. Factors such as reservoir water level fluctuations, rainfall, and time-dependent deformation significantly affect the baseline state of the seepage field. Static assessment standards struggle to distinguish between normal operating condition fluctuations and actual hazards, easily generating numerous irrelevant alarms or missed alarms.

[0004] There is a need to develop an intelligent monitoring technology that can integrate internal structural status and external operating condition information, and dynamically adjust evaluation criteria based on real-time operating conditions. This requires addressing how to effectively correlate and integrate time-series monitoring data with spatial detection data to generate features that better characterize the overall health status; at the same time, it requires that the risk assessment model have adaptive capabilities, enabling it to output more targeted assessment results based on real-time operating conditions, thereby improving the accuracy and reliability of early warnings. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent seepage monitoring system for water conservancy projects to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an intelligent seepage monitoring system for water conservancy projects, the system comprising:

[0007] The parameter acquisition unit is used to synchronously collect multi-dimensional seepage monitoring data of hydraulic engineering structures under different hydrological conditions and generate a hydrological-seepage correlation database.

[0008] The geological exploration unit is used to acquire ground-penetrating radar scan data of hydraulic engineering structures under different degrees of aging.

[0009] The feature generation unit, based on the hydrological-seepage correlation database and ground radar scanning data, forms an optimized multidimensional seepage feature set through data fusion technology;

[0010] The model building unit uses the optimized multidimensional seepage feature set to train the seepage health status assessment model and extract seepage health status indicators.

[0011] The risk grading unit defines the criteria for classifying seepage risk and establishes a mapping relationship between seepage health status indicators and seepage risk classification.

[0012] The real-time processing unit receives real-time operating parameters of the hydraulic engineering structure and, in conjunction with the seepage health status assessment model, calculates the seepage health status index that is adaptive to the operating conditions.

[0013] The early warning execution unit determines whether the seepage risk classification meets the early warning conditions based on the adaptive seepage health status indicators and mapping relationships under the operating conditions, and activates the early warning protocol when it does.

[0014] Preferably, the multidimensional seepage monitoring data in the parameter acquisition unit includes water pressure values ​​at measuring points, water flow velocity sequences, and seepage flow rate variation curves;

[0015] The specific process by which the feature generation unit forms an optimized multidimensional seepage feature set based on a hydrological-seepage correlation database and ground-penetrating radar scanning data through data fusion technology is as follows:

[0016] Signal decomposition processing was performed on the hydrological-seepage correlation database to extract time-frequency feature tensors;

[0017] Stratigraphic parameters are inverted from ground-penetrating radar scan data to generate geological feature tensors;

[0018] The time-frequency feature tensor and the geological feature tensor are weighted and combined to generate an optimized multidimensional seepage feature set;

[0019] The step of inverting stratigraphic parameters from ground-penetrating radar scanning data includes:

[0020] The original radar echo signal in the ground-penetrating radar scanning data is acquired, and the original radar echo signal is filtered and denoised to extract effective radar wavefield feature data.

[0021] Based on the propagation law of radar waves in the medium, a forward model is constructed from the formation parameters to the radar wave field characteristic data;

[0022] An inversion algorithm based on gradient descent is used to continuously adjust the assumed formation parameters to minimize the difference between the simulated radar wave field feature data output by the forward model and the actual extracted effective radar wave field feature data. The assumed formation parameters corresponding to the minimized difference are determined as the final formation wave impedance parameters.

[0023] Preferably, the feature generation unit performs signal decomposition processing on the hydrological-seepage correlation database and extracts time-frequency feature tensors, including:

[0024] A multi-resolution analysis method was used to divide the water pressure values ​​at the measuring points into frequency bands to obtain the water pressure frequency domain components.

[0025] Spectral analysis of the water flow velocity sequence yields the frequency domain components of the flow velocity.

[0026] Modal decomposition was performed on the seepage flow variation curve to obtain the frequency domain components of the seepage flow.

[0027] A time-frequency characteristic tensor is constructed based on the frequency domain components of water pressure, flow velocity, and seepage flow.

[0028] Preferably, the feature generation unit employs a multi-resolution analysis method, including:

[0029] Discrete wavelet transform is used to process the water pressure values ​​at the measuring points to obtain the water pressure wavelet coefficient sequence. Energy calculation is then performed on the water pressure wavelet coefficient sequence to form a water pressure energy distribution vector.

[0030] Apply a short-time Fourier transform to the water flow velocity sequence to generate a flow velocity spectrum vector;

[0031] Empirical mode decomposition is performed on the seepage flow variation curve to generate the seepage flow intrinsic mode function vector;

[0032] The water pressure energy distribution vector, the flow velocity spectrum vector, and the seepage flow intrinsic mode function vector are integrated into a time-frequency characteristic tensor.

[0033] Preferably, the feature generation unit performs stratigraphic parameter inversion on the ground-penetrating radar scan data to generate geological feature tensors, including:

[0034] Impedance calculations were performed on ground-penetrating radar scan data to derive formation impedance parameters;

[0035] Key reflection coefficients are selected from formation impedance parameters, and geological feature tensors are constructed based on these key reflection coefficients; specifically, this includes:

[0036] Calculate the reflection coefficients between adjacent strata in the formation impedance parameter sequence to form a reflection coefficient sequence, and set a reflection coefficient amplitude threshold in the reflection coefficient sequence;

[0037] Reflection coefficients whose absolute amplitude value is greater than the reflection coefficient amplitude threshold are selected as the key reflection coefficients;

[0038] Extract the depth information of the reflection interface and the geological properties of the formation on both sides of the reflection interface corresponding to each key reflection coefficient;

[0039] The key reflection coefficients, reflection interface depth information, and formation medium property information are vectorized and combined to generate the geological feature tensor.

[0040] Preferably, the weighted combination of the time-frequency feature tensor and the geological feature tensor in the feature generation unit includes:

[0041] The time-frequency feature tensor and the geological feature tensor are normalized to eliminate dimensional differences.

[0042] Weight coefficients are assigned based on feature importance. The normalized time-frequency feature tensor and geological feature tensor are linearly weighted, and the weighted feature tensors are concatenated to form an optimized multidimensional seepage feature set.

[0043] Preferably, the training of the seepage health status assessment model in the model building unit includes:

[0044] A training sample set is selected from the optimized multidimensional seepage feature set, where each sample contains a seepage feature vector and a manually labeled health status label;

[0045] Initialize a linear regression model as the base evaluation model;

[0046] The parameters of the basic assessment model are adjusted through an iterative optimization algorithm to minimize the error between the predicted health status and the labeled values, ultimately resulting in a trained seepage health status assessment model; specifically including:

[0047] Define an error function between the predicted health status and the labeled parameters; calculate the gradient of the error function under the current model parameters, where the gradient is the partial derivative of the error function with respect to the model parameters; update the parameters of the base evaluation model along the opposite direction of the gradient with a preset learning rate; calculate the error using the validation set after each parameter update; stop iterating when the error decrease over multiple consecutive iterations is less than a preset convergence threshold, thus obtaining the trained seepage health status evaluation model.

[0048] Preferably, the seepage health status indicators for adaptive calculation of operating conditions in the real-time processing unit include:

[0049] Real-time monitoring of the operating environment parameters of water conservancy engineering structures, including water level changes and load conditions;

[0050] The feature weights in the seepage health status assessment model are dynamically adjusted based on operating environment parameters.

[0051] Using the adjusted feature weights, the real-time seepage characteristics are weighted and calculated to output an adaptive seepage health status index.

[0052] The dynamic adjustment of feature weights in the seepage health status assessment model based on operating environment parameters includes:

[0053] Real-time reception of water level change data and load condition data;

[0054] Based on preset rules, the water level change data and load condition data are quantified into environmental factors.

[0055] Based on the values ​​of the environmental factors, a preset feature weight adjustment table is searched to obtain the weight adjustment amount corresponding to each seepage feature in the seepage health status assessment model.

[0056] The feature weights of the seepage health status assessment model are corrected using the weight adjustment amount.

[0057] Preferably, the early warning execution unit determines whether the seepage risk classification meets the early warning conditions by including:

[0058] The adaptive seepage health status index is input into the mapping relationship established by the risk classification unit. The corresponding seepage risk classification is queried, and the queried seepage risk classification is compared with the preset safety threshold. If the seepage risk classification exceeds the safety threshold, an early warning protocol is triggered; otherwise, monitoring continues.

[0059] Preferably, the step of inputting the adaptive seepage health status index into the mapping relationship established by the risk grading unit and querying the corresponding seepage risk classification includes:

[0060] Obtain the mapping relationship established by the risk classification unit, wherein the mapping relationship is a correspondence table between seepage health status index values ​​and seepage risk classification;

[0061] In the corresponding table, find the reference index value that is closest to the seepage health status index value that is adaptive to the working condition;

[0062] The seepage risk classification corresponding to the found reference index value is used as the output of the seepage risk classification mapped by the seepage health status index of the working condition adaptive.

[0063] Compared with the prior art, the beneficial effects of the present invention are:

[0064] By integrating temporal hydrological-seepage monitoring data with spatial ground-penetrating radar (GPR) scanning data, an optimized feature set that simultaneously reflects external stimuli and internal structural characteristics is generated. This allows the assessment of seepage status to overcome the limitations of traditional single-dimensional data, establishing a correlation between seepage anomalies and the causes of internal geological defects. The expansion and deepening of feature dimensions enable the more effective capture and identification of early risk signals hidden behind complex data, thereby changing the pattern of relying on isolated, superficial data for judgment.

[0065] The evaluation model is trained based on fused multidimensional features and can dynamically calculate and output health status indicators that match the current operating conditions based on real-time input parameters such as water level and time. This replaces the static method of using fixed thresholds for alarms, allowing the evaluation criteria to automatically adjust according to the actual operating environment of the project. The accuracy of the evaluation results is thus improved, distinguishing between benign changes caused by fluctuations in normal operating conditions and abnormal evolutions representing real hazards, reducing false alarms, and enhancing sensitivity to real risks under complex operating conditions.

[0066] The final risk grading and early warning decisions are based on the aforementioned dynamic, multi-dimensional comprehensive evaluation indicators. Early warning triggers no longer rely on a simple judgment of whether a single data point exceeds a fixed limit, but are based on an intelligent analysis result that has undergone multi-source information verification and operating condition normalization. This fundamentally improves the scientific rigor and reliability of early warning decisions, enabling more objective and proactive indication of risk levels and providing a more accurate and timely basis for taking targeted preventative measures. Attached Figure Description

[0067] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent seepage monitoring system for water conservancy projects described in this invention.

[0068] Figure 2 This is a flowchart for extracting time-frequency feature tensors;

[0069] Figure 3 This is a flowchart of a weighted combination of time-frequency and geological feature tensors;

[0070] Figure 4 This is a curve showing the change in the mean square error of the training and validation sets during the gradient descent iteration process;

[0071] Figure 5 The training error convergence curve of the seepage health status assessment model. Detailed Implementation

[0072] 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.

[0073] Please see Figure 1 This invention provides an intelligent seepage monitoring system for hydraulic engineering projects. The system comprises a parameter acquisition unit, a geological exploration unit, a feature generation unit, a model building unit, a risk classification unit, a real-time processing unit, and an early warning execution unit working collaboratively. The parameter acquisition unit synchronously collects multi-dimensional seepage monitoring data of the hydraulic engineering structure under different hydrological conditions and generates a hydrological-seepage correlation database. The geological exploration unit acquires ground-penetrating radar scan data of the hydraulic engineering structure at different aging levels. Based on the aforementioned hydrological-seepage correlation database and ground-penetrating radar scan data, the feature generation unit forms an optimized multi-dimensional seepage feature set through data fusion technology. The model building unit uses this optimized multi-dimensional seepage feature set to train a seepage health status assessment model and extracts seepage health status indicators from it. The risk classification unit defines seepage risk classification standards and establishes a mapping relationship between seepage health status indicators and seepage risk classifications. The real-time processing unit receives real-time operating parameters of the hydraulic engineering structure and, combined with the trained seepage health status assessment model, calculates condition-adaptive seepage health status indicators. The early warning execution unit determines whether the current seepage risk classification has reached the preset early warning conditions based on the calculated adaptive seepage health status indicators and the mapping relationship established by the risk classification unit, and activates the corresponding early warning protocol when the conditions are met.

[0074] Example 1: In specific implementation, the parameter acquisition unit collects multidimensional seepage monitoring data including water pressure values ​​at measuring points, water flow velocity sequences, and seepage flow rate variation curves. The feature generation unit, based on the hydrological-seepage correlation database and ground-penetrating radar scanning data, forms an optimized multidimensional seepage feature set through data fusion technology. The feature generation unit performs signal decomposition processing on the hydrological-seepage correlation database to extract time-frequency feature tensors. The feature generation unit performs stratigraphic parameter inversion on the ground-penetrating radar scanning data to generate geological feature tensors. The feature generation unit performs weighted combination of the time-frequency feature tensors and the geological feature tensors to generate the optimized multidimensional seepage feature set. In practical implementation, the inversion of stratigraphic parameters from ground-penetrating radar (GPR) scanning data includes acquiring the original radar echo signal from the GPR scanning data, filtering and denoising the original radar echo signal to extract effective radar wavefield feature data, constructing a forward model from stratigraphic parameters to radar wavefield feature data based on the propagation law of radar waves in the medium, and using a gradient descent-based inversion algorithm to continuously adjust the assumed stratigraphic parameters to minimize the difference between the simulated radar wavefield feature data output by the forward model and the actual extracted effective radar wavefield feature data. The assumed stratigraphic parameters corresponding to the minimized difference are determined as the final stratigraphic wave impedance parameters. In some embodiments, the filtering and denoising process is implemented using a digital filter, and the forward model is constructed based on the radar wave equation. It can be understood that the gradient descent-based inversion algorithm uses an error function to measure the difference between the simulated data and the actual data. The error function is defined as the squared Euclidean distance between the simulated radar wavefield feature data and the actual radar wavefield feature data. Represented as:

[0075]

[0076] in: Represents the error function. Represents a formation parameter vector. Indicates the forward model in the th... Simulated radar wavefield characteristic data output from each data point. This indicates that the actual extracted effective radar wavefield feature data is in the first... The value of each data point This represents the total number of data points. The gradient descent algorithm updates the formation parameter vector iteratively. To minimize the error function In practical implementation, the iterative update rule of the gradient descent algorithm is to calculate the gradient of the error function with respect to the formation parameter vector and adjust the formation parameter vector along the negative gradient direction. Optionally, the learning rate parameter controls the step size of each update, and the iteration process continues until the change in the error function value is less than a preset threshold. Taking the processing of ground-penetrating radar scanning data of a water conservancy project dam as an example, after filtering and denoising the acquired raw radar echo signal, 1000 effective radar wavefield feature data points (M=1000) are extracted. The constructed forward model is based on the propagation equation of radar waves in the soil medium, assuming that the formation parameter vector q contains two parameters: wave impedance Z and dielectric constant ε. The error function is set as follows:

[0077]

[0078] in: This represents the j-th simulated radar wavefield feature data output by the forward model. This represents the j-th effective radar wavefield feature data actually extracted.

[0079] Initial assumptions about the formation parameter vector The learning rate is preset to η = 0.001, and the convergence threshold is preset to ε = 0.0001. During the iteration process, the gradient of the error function with respect to q is calculated in the first iteration. Update the parameters along the negative gradient direction to obtain ,calculate and In comparison, this process is repeated. When iterating to the 50th iteration, the error function value decreases from the initial 125.6 to 0.00008, meeting the convergence threshold. The corresponding formation parameter vector at this point... This is the final determined formation impedance parameter. The error between this result and the formation impedance parameter obtained from the field borehole sampling test is less than 3%, which verifies the effectiveness of the algorithm.

[0080] In some embodiments, the forward model can be simplified to a linear model or a full-wavefield simulation method can be used, depending on the geological complexity and computational resources. It is understood that the formation parameter vector includes wave impedance and dielectric constant. In specific implementations, the weighted combination of time-frequency feature tensors and geological feature tensors involves normalizing both tensors and assigning weight coefficients for linear combination. Optionally, the weight coefficients can be set based on feature importance or expert experience, and the linearly combined feature tensors constitute an optimized multidimensional seepage feature set.

[0081] Example 2: See Figure 2In specific implementations, the feature generation unit performs signal decomposition processing on the hydrological-seepage correlation database to extract time-frequency feature tensors. Specifically, this includes using multi-resolution analysis to divide the water pressure values ​​at measuring points into frequency bands to obtain the water pressure frequency domain component, performing spectral analysis on the water flow velocity sequence to obtain the flow velocity frequency domain component, and performing mode decomposition on the seepage flow variation curve to obtain the seepage flow frequency domain component. A time-frequency feature tensor is then constructed based on the water pressure frequency domain component, the flow velocity frequency domain component, and the seepage flow frequency domain component. In specific implementations, the multi-resolution analysis method includes using discrete wavelet transform to process the water pressure values ​​at measuring points to obtain the water pressure wavelet coefficient sequence, and performing energy calculation on the water pressure wavelet coefficient sequence to form a water pressure energy distribution vector. In some embodiments, a short-time Fourier transform is applied to the water flow velocity sequence to generate a flow velocity spectrum vector, and empirical mode decomposition is performed on the seepage flow variation curve to generate a seepage flow intrinsic mode function vector. In specific implementations, the water pressure energy distribution vector, the flow velocity spectrum vector, and the seepage flow intrinsic mode function vector are integrated into a time-frequency feature tensor. It can be understood that the Discrete Wavelet Transform (DWT) performs multi-scale decomposition of the water pressure signal at the measuring point using wavelet basis functions. The calculation of the DWT can be expressed as:

[0082]

[0083] in: Indicated in scale parameter Translation parameters The wavelet coefficients obtained from the calculation are as follows: This represents a discrete sequence of water pressure values ​​from different measuring points. Indicates the mother wavelet function The wavelet basis functions obtained after scaling and translation This represents the index of the signal sequence. Energy calculation on the obtained water pressure wavelet coefficient sequence involves summing the squares of the coefficients at each scale to form a water pressure energy distribution vector. In some embodiments, the mother wavelet function can be either the Daubechies wavelet or the Haar wavelet.

[0084] Taking the water pressure monitoring data at a certain reservoir dam as an example, the collected water pressure signals at the monitoring points... For a discrete sequence (n=0,1,…,999) with a duration of 100 seconds and a sampling frequency of 10Hz, the Daubechies-4 wavelet was selected as the mother wavelet function. The scale parameter a=1,2,3,4 was set, and the translation parameter b was set according to the sampling interval. After calculating the wavelet coefficient sequence at each scale, the sum of squares of the coefficients at each scale was used to obtain the water pressure energy distribution vector. For example, the sum of squares of the wavelet coefficients at scale a=1 is 5.2 Pa², at scale a=2 it is 3.8 Pa², at scale a=3 it is 2.1 Pa², and at scale a=4 it is 1.5 Pa², forming a water pressure energy distribution vector of [5.2,3.8,2.1,1.5]. This vector clearly reflects the energy distribution characteristics of the water pressure value at the measuring point in different frequency bands, providing basic data for the subsequent construction of the time-frequency feature tensor.

[0085] Optionally, the short-time Fourier transform performs piecewise Fourier analysis on the water flow velocity sequence using a windowing method, thereby generating a time-varying velocity spectrum vector. In specific implementations, empirical mode decomposition adaptively decomposes the seepage flow variation curve into a series of seepage flow eigenmode function components from high frequency to low frequency; the set of these components constitutes the seepage flow eigenmode function vector. It can be understood that constructing the time-frequency feature tensor involves concatenating and arranging the water pressure energy distribution vector, the velocity spectrum vector, and the seepage flow eigenmode function vector within a tensor structure. Optionally, the time-frequency feature tensor is a three-dimensional array, with its dimensions corresponding to the feature type, frequency components, and time series, respectively.

[0086] Example 3: See Figure 3 In specific implementation, the feature generation unit performs stratigraphic parameter inversion on the ground-penetrating radar (GPR) scan data to generate a geological feature tensor. This process includes calculating the wave impedance of the GPR scan data to derive stratigraphic wave impedance parameters, and selecting key reflection coefficients from the stratigraphic wave impedance parameters to construct the geological feature tensor based on these key reflection coefficients. Specifically, the reflection coefficients between adjacent stratigraphic layers in the stratigraphic wave impedance parameter sequence are calculated to form a reflection coefficient sequence. A reflection coefficient amplitude threshold is set in the reflection coefficient sequence, and reflection coefficients with absolute amplitude values ​​greater than the reflection coefficient amplitude threshold are selected as key reflection coefficients. The reflection interface depth information and stratigraphic medium property information on both sides of the reflection interface corresponding to each key reflection coefficient are extracted. The key reflection coefficients, reflection interface depth information, and stratigraphic medium property information are then vectorized and combined to generate the geological feature tensor. In some embodiments, the formula for calculating the reflection coefficient is:

[0087]

[0088] in: Indicates the first Reflection coefficient at each reflective interface Indicates the first Wave impedance of the strata Indicates the first The wave impedance of each stratum. The reflection coefficient amplitude threshold is a preset constant used to identify interfaces with significant signal strength. In specific implementations, the depth information of the reflection interface comes from the conversion of two-way travel time and wave velocity in the ground-penetrating radar scan data, and the stratum medium property information includes wave impedance values ​​and dielectric constants. It can be understood that the geological feature tensor is a multidimensional array whose row vectors contain the values ​​of key reflection coefficients, their corresponding depths, and the medium property parameters of the upper and lower strata. In some embodiments, the vectorization process combines each key reflection coefficient and its associated information into a feature vector, and the feature vectors corresponding to all key reflection coefficients are stacked along a specific dimension to form the geological feature tensor. The feature generation unit performs a weighted combination of the time-frequency feature tensor and the geological feature tensor. This process includes normalizing the time-frequency feature tensor and the geological feature tensor respectively to eliminate dimensional differences, assigning weight coefficients according to feature importance, linearly weighting the normalized time-frequency feature tensor and the geological feature tensor, and concatenating the weighted feature tensors into an optimized multidimensional seepage feature set. In practice, normalization is achieved using either min-max normalization or Z-score standardization. Optionally, weighting coefficients are determined by a feature importance assessment algorithm. Specifically, the feature importance assessment algorithm utilizes a training sample set from the optimized multidimensional seepage feature set for analysis. This training sample set includes seepage feature vectors and manually labeled health status tags. The algorithm calculates a statistical correlation index, such as the Pearson correlation coefficient or mutual information value, between each seepage feature and the health status tag. Features are ranked according to the calculated statistical correlation index; features with higher correlation indices are considered more important. Then, corresponding weighting coefficients are assigned based on feature importance, with higher-important features receiving larger weighting coefficients and lower-important features receiving smaller weighting coefficients. The specific values ​​of the weighting coefficients are normalized to ensure the sum of all weighting coefficients equals one, guaranteeing a reasonable contribution ratio for each feature during weighted combination. The feature importance assessment algorithm can periodically recalculate the weighting coefficients using an updated training sample set to adapt to changes in data distribution and maintain the timeliness of weight allocation. Linear weighting refers to multiplying each element of the normalized time-frequency feature tensor by its corresponding time-frequency weight coefficient, and multiplying each element of the normalized geological feature tensor by its corresponding geological weight coefficient. In essence, the concatenation operation connects the weighted time-frequency feature tensor and the weighted geological feature tensor along a new feature dimension, thereby forming a higher-dimensional, optimized multidimensional seepage feature set.

[0089] Example 4: In specific implementation, the training of the seepage health status assessment model in the model construction unit includes selecting a training sample set from the optimized multi-dimensional seepage feature set, where each sample contains a seepage feature vector and a manually labeled health status label. A linear regression model is initialized as the basic assessment model. The parameters of the basic assessment model are adjusted through an iterative optimization algorithm to minimize the error between the predicted health status and the labeled label, ultimately obtaining the trained seepage health status assessment model. In specific implementation, an error function between the predicted health status and the labeled label is defined. The gradient of the error function relative to the model parameters is calculated under the current model parameters. The parameters of the basic assessment model are updated along the opposite direction of the gradient with a preset learning rate. After each parameter update, the error is calculated using the validation set. When the error decrease over multiple consecutive iterations is less than a preset convergence threshold, the iteration stops, resulting in the trained seepage health status assessment model. It can be understood that the predicted output of the linear regression model is the seepage health status index, which is calculated as a linear combination of the seepage feature vector and the model weight vector plus a bias term. The prediction formula of the linear regression model is:

[0090]

[0091] in: This represents the value of the seepage health status index predicted by the linear regression model. Indicates the corresponding number Model weights for each seepage characteristic, This represents the first element in the input seepage feature vector. The value of each feature, This represents the bias term in a linear regression model. This represents the total number of seepage characteristics. The error function typically uses mean squared error. The iterative optimization algorithm employs gradient descent. The real-time processing unit calculates the adaptive seepage health status index by real-time monitoring of the operating environment parameters of the hydraulic engineering structure, including water level changes and load conditions. Based on these operating environment parameters, the feature weights in the seepage health status assessment model are dynamically adjusted. Using the adjusted feature weights, the real-time seepage characteristics are weighted and calculated to output the adaptive seepage health status index. In specific implementations, dynamically adjusting the feature weights in the seepage health status assessment model based on operating environment parameters includes receiving real-time water level change data and load condition data. Based on preset rules, the water level change data and load condition data are quantified into environmental factors. Based on the values ​​of the environmental factors, a preset feature weight adjustment table is consulted to obtain the weight adjustment amount corresponding to each seepage characteristic in the seepage health status assessment model. The weight adjustment amount is then used to correct the feature weights of the seepage health status assessment model. In some embodiments, the environmental factor is a comprehensive scalar, calculated by weighted summation of the water level change amplitude and load condition level. Optionally, the feature weight adjustment table defines the mapping relationship between different environmental factor value ranges and the adjustment amounts of each seepage feature weight. See Table 1, Feature Weight Adjustment Table.

[0092] Table 1: Feature Weight Adjustment Table

[0093]

[0094] It can be understood that the weight correction operation involves adding the weight adjustment amount obtained from the lookup table to the current feature weights of the seepage health status assessment model. In some embodiments, the water level change data is the difference between the current water level and the reference water level, and the load condition data is the quantified value of the static and dynamic loads on the structure. In specific implementation, using the adjusted feature weights to perform weighted calculations on the real-time seepage features means multiplying the real-time seepage feature vector obtained from real-time acquisition and processing by the feature generation unit with the corrected model weight vector and adding a bias term. The calculation result is the condition-adaptive seepage health status index.

[0095] See Figure 4In the training phase of the seepage health status assessment model, the error changes during the gradient descent iteration process are quantitatively analyzed using the mean squared error curves of the training and validation sets. Specifically, the training set error (blue curve) and validation set error (red curve) show a gradual decreasing trend with increasing iteration count: the initial error decrease rate is relatively fast, reflecting the rapid convergence of model parameters towards the optimal solution; as the number of iterations approaches 20, the error decrease rate slows down, gradually approaching the preset convergence threshold (purple dashed line). When the number of iterations exceeds 30, both the training and validation set errors stabilize around the convergence threshold, indicating that the model parameters have reached a relatively optimal state. During parameter configuration, the convergence threshold is set to 0.0, the learning rate uses a preset fixed value, and the small fluctuations in the validation set error reflect the change in the model's generalization ability to unseen data.

[0096] Example 5: In a specific implementation, determining whether the seepage risk classification meets the warning conditions in the early warning execution unit includes inputting the condition-adaptive seepage health status index into the mapping relationship established by the risk grading unit, querying the corresponding seepage risk classification, comparing the queried seepage risk classification with a preset safety threshold, and triggering an early warning protocol if the seepage risk classification exceeds the safety threshold; otherwise, monitoring continues. In a specific implementation, inputting the condition-adaptive seepage health status index into the mapping relationship established by the risk grading unit to query the corresponding seepage risk classification includes obtaining the mapping relationship established by the risk grading unit. The mapping relationship is a correspondence table between seepage health status index values ​​and seepage risk classifications. In the correspondence table, the reference index value closest to the condition-adaptive seepage health status index value is found, and the seepage risk classification corresponding to the found reference index value is output as the seepage risk classification mapped by the condition-adaptive seepage health status index. In some embodiments, finding the reference index value closest to the condition-adaptive seepage health status index value is achieved by calculating the absolute difference between the index values, the mathematical expression of which is:

[0097]

[0098] in: This represents the index of the reference index value that minimizes the distance in the correspondence table. This represents the seepage health status index value obtained through real-time calculation and adaptive operation. This indicates the first element in the correspondence table. Reference permeation health status index value for row storage, This operation retrieves the index of the independent variable that minimizes the function value. It can be understood that the search process involves traversing all reference indicator values ​​in the corresponding table and calculating the real-time indicator value. With each reference indicator value The absolute difference is calculated, and the reference index value with the smallest absolute difference is selected. In specific implementation, the correspondence table is a static lookup table that stores discrete reference seepage health status index values ​​and their predefined, one-to-one corresponding seepage risk classifications. Optionally, seepage risk classifications can be represented by integer levels, such as 1 representing low risk, 2 representing medium risk, and 3 representing high risk. Taking the real-time seepage monitoring of a certain levee project as an example, the mapping relationship established by the risk grading unit is a correspondence table between seepage health status index values ​​and risk classifications, where the reference index values ​​and corresponding risk classifications are: [0.0-2.0) corresponds to low risk, [2.0-4.0) corresponds to medium risk, [4.0-6.0) corresponds to relatively high risk, and [6.0-10.0] corresponds to high risk, with a preset safety threshold of level 3. The real-time processing unit calculates the seepage health status index values ​​that are adaptive to the working conditions. The reference index values ​​in the corresponding relationship table are traversed according to the lookup formula. , The absolute differences are calculated as follows: , ,get The corresponding reference indicator value of 4.0 belongs to risk category level 3. Since this risk category equals the preset safety threshold, it meets the warning conditions. The warning execution unit triggers the warning protocol, sends a warning message to the monitoring center, and records the event log. If the indicator value is subsequently calculated in real-time... If the corresponding risk classification is found to be Level 2 and does not exceed the safety threshold, then monitoring will continue.

[0099] In some embodiments, comparing the queried seepage risk classification with a preset safety threshold directly compares the numerical level of the risk classification. It can be understood that the preset safety threshold is a set upper limit for a risk level; for example, if the safety threshold is set to risk level 2, and the queried seepage risk classification is level 3, level 3 exceeds level 2, thus meeting the warning conditions. In specific implementations, triggering the warning protocol means initiating a preset alarm action sequence, which may include illuminating alarm lights, sending alarm information to the monitoring center, and recording alarm event logs. Optionally, continued monitoring means that the warning execution unit does not execute any alarm actions, but instead waits for and processes the next adaptive seepage health status indicator transmitted by the real-time processing unit, and repeats the above query, comparison, and judgment process.

[0100] See Figure 5 In the training process of the seepage health status assessment model, the analysis of error convergence characteristics uses the changes in the mean squared error (MSE) of the training and validation sets as the core observation indicator. Specifically, the training set error (green curve) and the validation set error (red curve) show a gradual decreasing trend with the increase of the number of training iterations, their magnitude decreasing from the initial 10... 4The error decays from level ¹ to around level 10²². Error convergence is determined based on a preset convergence threshold (0.05, blue dashed line). As training iterations progress to the later stages, both types of errors approach this threshold. During training, fluctuations in the training set error reflect the dynamic process of model parameter adjustment: rapid error decreases in the early stages correspond to the model's initial fit to the training samples; local fluctuations in the middle stages are normal oscillations during gradient updates in parameter optimization; and the later stabilization of the error indicates that the model parameters are close to the optimal solution. The trends of the validation set error and the training set error are consistent, with no significant deviation, indicating that the model has not experienced overfitting or underfitting. In terms of parameter configuration, the number of training iterations is set to 50. The convergence criterion is that the error decrease over multiple consecutive iterations is less than the convergence threshold (0.05). As shown in the figure, in the later stages of iteration (approximately after 40 iterations), the error has stabilized near the threshold, meeting the convergence requirements for model training.

[0101] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A seepage intelligent monitoring system for hydraulic engineering, characterized in that, The system consists of the following units: A parameter acquisition unit for synchronously collecting multi-dimensional seepage monitoring data of the water conservancy structure under different hydrological conditions and generating a hydrology-seepage correlation database; A geological detection unit for obtaining geological radar scanning data of the water conservancy structure under different aging degrees; A feature generation unit for forming an optimized multi-dimensional seepage feature set through data fusion technology based on the hydrology-seepage correlation database and the geological radar scanning data; A model construction unit for training a seepage health state evaluation model and extracting seepage health state indicators using the optimized multi-dimensional seepage feature set; A risk classification unit for defining seepage risk classification standards and establishing a mapping relationship between the seepage health state indicators and the seepage risk classification; A real-time processing unit for receiving real-time operation parameters of the water conservancy structure, combining the seepage health state evaluation model, and calculating seepage health state indicators adaptive to the working conditions; An early warning execution unit for determining whether the seepage risk classification meets the early warning conditions according to the seepage health state indicators adaptive to the working conditions and the mapping relationship, and activating an early warning protocol when the conditions are met; The multi-dimensional seepage monitoring data in the parameter acquisition unit includes water pressure values, water flow velocity sequences, and seepage flow variation curves of the measuring points; The specific process of forming the optimized multi-dimensional seepage feature set through data fusion technology based on the hydrology-seepage correlation database and the geological radar scanning data in the feature generation unit is as follows: Signal decomposition processing is performed on the hydrology-seepage correlation database to extract time-frequency feature tensors; Stratum parameter inversion is performed on the geological radar scanning data to generate geological feature tensors; The time-frequency feature tensors and the geological feature tensors are combined by weighting to generate the optimized multi-dimensional seepage feature set; The stratum parameter inversion of the geological radar scanning data includes: Original radar echo signals in the geological radar scanning data are obtained, and the original radar echo signals are filtered and denoised to extract effective radar wave field feature data; According to the propagation law of radar waves in the medium, a forward modeling model from stratum parameters to radar wave field feature data is constructed; An inversion algorithm based on gradient descent is adopted to continuously adjust the assumed stratum parameters, so that the difference between the simulated radar wave field feature data output by the forward modeling model and the actually extracted effective radar wave field feature data is minimized. When the difference is minimized, the corresponding assumed stratum parameters are determined as the final stratum wave impedance parameters.

2. The intelligent seepage monitoring system for hydraulic engineering according to claim 1, wherein, The signal decomposition processing of the hydrology-seepage correlation database to extract time-frequency feature tensors in the feature generation unit includes: A multi-resolution analysis method is used to divide the water pressure values into frequency bands to obtain water pressure frequency domain components; Spectral analysis is performed on the water flow velocity sequences to obtain flow velocity frequency domain components; Modal decomposition is performed on the seepage flow variation curves to obtain seepage flow frequency domain components; Based on the water pressure frequency domain components, the flow velocity frequency domain components, and the seepage flow frequency domain components, a time-frequency feature tensor is constructed.

3. The intelligent seepage monitoring system for hydraulic engineering according to claim 2, characterized in that, The multi-resolution analysis method used in the feature generation unit includes: Discrete wavelet transform is used to process the water pressure values to obtain water pressure wavelet coefficient sequences, and energy calculation is performed on the water pressure wavelet coefficient sequences to form a water pressure energy distribution vector; applying short-time Fourier transform to the water flow velocity sequence to generate a flow velocity spectrum vector; performing empirical mode decomposition on the seepage flow variation curve to generate a seepage flow intrinsic mode function vector; integrating the water pressure energy distribution vector, the flow velocity spectrum vector and the seepage flow intrinsic mode function vector into a time-frequency feature tensor.

4. The intelligent seepage monitoring system for hydraulic engineering according to claim 3, characterized in that, The feature generation unit includes: calculating the wave impedance of the geological radar scanning data to derive the formation wave impedance parameters; selecting key reflection coefficients from the formation wave impedance parameters, and constructing a geological feature tensor based on the key reflection coefficients; specifically including: calculating the reflection coefficients between adjacent formations in the formation wave impedance parameter sequence to form a reflection coefficient sequence, and setting a reflection coefficient amplitude threshold in the reflection coefficient sequence; selecting reflection coefficients with an absolute value greater than the reflection coefficient amplitude threshold as the key reflection coefficients; extracting the reflection interface depth information and formation medium attribute information on both sides of each key reflection coefficient; vectorizing the key reflection coefficients, reflection interface depth information and formation medium attribute information to generate the geological feature tensor.

5. The intelligent seepage monitoring system for hydraulic engineering according to claim 4, characterized in that, The feature generation unit includes: normalizing the time-frequency feature tensor and the geological feature tensor to eliminate dimensional differences; assigning weight coefficients according to feature importance, linearly weighting the normalized time-frequency feature tensor and the geological feature tensor, and concatenating the weighted feature tensors into an optimized multi-dimensional seepage flow feature set.

6. The intelligent seepage monitoring system for hydraulic engineering of claim 1, wherein, The model construction unit includes: selecting a training sample set from the optimized multi-dimensional seepage flow feature set, wherein each sample contains a seepage flow feature vector and an artificially labeled health status label; initializing a linear regression model as a basic evaluation model; adjusting the parameters of the basic evaluation model through an iterative optimization algorithm to minimize the error between the predicted health status and the labeled label, and finally obtaining a trained seepage flow health status evaluation model; specifically including: defining an error function between the predicted health status and the labeled label; calculating the gradient of the error function under the current model parameters, the gradient being the partial derivative of the error function with respect to the model parameters, updating the parameters of the basic evaluation model in the opposite direction of the gradient with a preset learning rate, after each parameter update, using the validation set to calculate the error, when the error reduction of continuous multiple iterations is less than a preset convergence threshold, stopping iteration, and obtaining a trained seepage flow health status evaluation model.

7. The intelligent seepage monitoring system for hydraulic engineering according to claim 6, characterized in that, The real-time processing unit includes: real-time monitoring of the operating environment parameters of the water conservancy project structure, including water level changes and load conditions; dynamically adjusting the feature weights in the seepage flow health status evaluation model according to the operating environment parameters; using the adjusted feature weights to weight and calculate the real-time seepage flow features, and outputting the working condition adaptive seepage flow health status indicator; The feature weights in the seepage flow health status evaluation model are dynamically adjusted according to the operating environment parameters, including: Real-time receive water level change data and load condition data; Quantify the water level change data and load condition data into environmental factors based on preset rules; According to the numerical value of the environmental factor, find the preset characteristic weight adjustment table to obtain the weight adjustment amount corresponding to each seepage characteristic in the seepage health state evaluation model; Use the weight adjustment amount to correct the characteristic weight of the seepage health state evaluation model.

8. The intelligent seepage monitoring system for hydraulic engineering of claim 1, wherein, The determination of whether the seepage risk classification reaches the early warning condition in the early warning execution unit includes: Input the working condition adaptive seepage health state index into the mapping relationship established by the risk grading unit to query the corresponding seepage risk classification, compare the queried seepage risk classification with the preset safety threshold, if the seepage risk classification exceeds the safety threshold, trigger the early warning protocol, otherwise continue to monitor.

9. The seepage intelligent monitoring system according to claim 8, wherein, The input of the working condition adaptive seepage health state index into the mapping relationship established by the risk grading unit to query the corresponding seepage risk classification includes: Obtain the mapping relationship established by the risk grading unit, which is a corresponding relationship table of seepage health state index value and seepage risk classification; In the corresponding relationship table, find the reference index value closest to the working condition adaptive seepage health state index value; The seepage risk classification corresponding to the found reference index value is output as the seepage risk classification mapped by the working condition adaptive seepage health state index.

Images