A method and system for monitoring and early warning of a water conservancy dam safety

Abstract

The application relates to the technical field of water conservancy project safety monitoring, and particularly discloses a water conservancy project dam safety monitoring and early warning method and system, which realizes comprehensive perception and accurate early warning of the health state of a dam structure through composite sensing technology and intelligent analysis algorithms; a three-dimensional monitoring network integrated with interfaces and structures is constructed through the cooperative deployment of micromechanical resonant sensors and distributed optical fiber sensors; a three-dimensional interface peeling characteristic spectrum is constructed based on time-frequency joint analysis technology, and the bonding deterioration state between the sensors and the dam body is accurately identified; a strain field abnormal distribution matrix is established through spatial correlation modeling, accurate positioning of internal damage is realized, a double-channel feature fusion network based on an attention mechanism and a deep neural network evaluator are designed, multi-dimensional correlation analysis of the interface state and structural damage features is carried out, and finally, a three-stage linkage early warning decision tree is used to realize a progressive response from data verification, multi-source verification to emergency linkage.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hydraulic engineering safety monitoring, and particularly relates to a hydraulic engineering dam safety monitoring and early warning method and system. BACKGROUND

[0002] With the continuous expansion of the construction scale of high dams and large reservoirs, dam safety monitoring is facing unprecedented technical challenges. The traditional monitoring system mainly relies on single-point sensors to collect physical quantities such as displacement and seepage pressure, and makes safety judgments through preset thresholds, which is difficult to meet the comprehensive perception needs of modern engineering for the health status of the structure. Especially under complex environmental loads and long-term operation conditions, the damage to the dam structure often presents the characteristics of multi-scale development and multi-factor coupling. The existing technical means have obvious deficiencies in the aspects of early damage identification accuracy, multi-source information fusion depth and intelligent early warning timeliness, and a new generation of intelligent monitoring technology system is urgently needed.

[0003] The prior art has the following disadvantages: In the long-term monitoring process, the interface between the buried sensor and the dam concrete will produce progressive bonding deterioration, resulting in systematic deviation of the monitoring data from the true structure state. This interface peeling phenomenon has the characteristics of strong concealment and slow development, and the conventional monitoring system cannot effectively identify it, often mistaking the signal drift caused by sensor failure as structural deformation, causing serious false alarms. Especially in high-stress areas and high-age dams, this problem is particularly prominent, and has become a deep technical bottleneck affecting the reliability of dam safety monitoring. SUMMARY

[0004] The purpose of the present application is to provide a hydraulic engineering dam safety monitoring and early warning method and system to solve the problems in the above background.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] A hydraulic engineering dam safety monitoring and early warning method, comprising the following steps:

[0007] S1: Real-time acquisition of structure response parameters by a composite sensor array deployed at key positions of the dam, the structure response parameters including: micro-mechanical resonance frequency offset and distributed optical fiber strain gradient;

[0008] S2: Time-frequency joint analysis of the micro-mechanical resonance frequency offset, construction of an interface peeling characteristic spectrum, calculation of an interface bonding deterioration characteristic value for evaluating the mechanical coupling state of the sensor and dam interface;

[0009] S3: Spatial correlation modeling of the distributed optical fiber strain gradient, construction of a strain field abnormal distribution matrix, calculation of a structure damage sensitivity characteristic value for identifying potential damage areas inside the dam;

[0010] S4: The interface adhesion degradation eigenvalue and the structural damage sensitivity eigenvalue are integrated into a comprehensive safety feature vector, which is then input into the pre-trained dam health assessment model for multi-dimensional analysis. The safety status is then graded and assessed based on the analysis results.

[0011] S5: Dynamically adjust the monitoring and early warning strategy based on the assessment results: When the comprehensive security feature value exceeds the first threshold, data credibility verification is triggered; when it exceeds the second threshold, multi-source data fusion verification is initiated; when it exceeds the third threshold, a graded early warning signal is generated and the emergency response mechanism is linked;

[0012] The comprehensive security feature value is represented in the form of a multi-dimensional comprehensive security feature vector.

[0013] As a further solution of the present invention: the construction of the interface peeling characteristic spectrum specifically includes:

[0014] An empirical mode decomposition is performed on the micromechanical resonant frequency offset to obtain intrinsic mode function components containing different time scales; a Hilbert transform is performed on each intrinsic mode function component to extract the instantaneous frequency characteristics; the instantaneous frequency of each component is compared with a preset reference frequency to calculate a multi-scale frequency shift eigenvector; the multi-scale frequency shift eigenvector is aligned in time and space with the ambient temperature monitoring data to construct a three-dimensional interface peeling characteristic spectrum, in which the first dimension represents the time scale, the second dimension represents the frequency offset, and the third dimension represents the temperature compensation coefficient.

[0015] As a further solution of the present invention: the calculation of the interface adhesion degradation characteristic value specifically includes:

[0016] Based on the interface peeling characteristic spectrum, the main frequency band energy distribution of the normal vibration mode is extracted, and its relative entropy value with respect to the reference state is calculated as the first characteristic parameter; then, the degree of distortion of the harmonic component of the tangential vibration mode is analyzed, and its harmonic distortion rate is calculated as the second characteristic parameter; the first and second characteristic parameters are weightedly fused, where the weight coefficient is dynamically adjusted according to the sensor burial depth and concrete age; and through a preset interface state mapping function, the fused characteristic parameters are converted into interface bonding degradation characteristic values, which range from 0 to 1. A larger value indicates a worse interface bonding state.

[0017] As a further solution of the present invention: the specific method for evaluating the mechanical coupling state includes:

[0018] An interface state assessment model based on a deep belief network is established, and the interface adhesion degradation characteristic value is used as the input feature. The hidden layer of the model contains three layers of restricted Boltzmann machines, which are used to extract shallow frequency domain features, mid-level time domain features, and deep nonlinear coupling features, respectively. The output layer of the model uses a softmax function to generate the interface state probability distribution, including four state classifications: intact state, slight peeling, moderate peeling, and severe peeling. When the output probability exceeds the preset threshold, the corresponding state warning signal is triggered, and the abnormal time series characteristics are automatically recorded for subsequent trend analysis.

[0019] As a further solution of the present invention: constructing a strain field abnormal distribution matrix, specifically including:

[0020] The distributed optical fiber strain gradient is spatially gridded, and the dam body is divided into several spatial units. The coefficient of variation of the optical fiber strain gradient in each spatial unit is then calculated to form an initial strain distribution matrix. The strain transfer coefficient of adjacent spatial units is then introduced, and the isolated outliers in the initial strain distribution matrix are corrected using a spatial autocorrelation algorithm. Finally, the corrected matrix is ​​weighted in combination with the stress concentration factor of the finite element model of the dam structure to generate a physically meaningful strain field anomaly distribution matrix, in which the matrix element value reflects the degree of strain anomaly in the corresponding spatial unit.

[0021] As a further solution of the present invention: the specific method for calculating the structural damage sensitivity characteristic value includes:

[0022] Based on the strain field anomaly distribution matrix, high anomaly areas with an error of more than two times the standard deviation are extracted from the matrix; the area ratio and spatial aggregation of the high anomaly areas are calculated as primary features; the deviation angle between the strain gradient direction of the high anomaly areas and the principal stress direction of the dam body is analyzed as secondary features; the primary features and secondary features are input into a pre-trained damage sensitivity assessment network. The damage sensitivity assessment network adopts a three-layer convolutional neural network structure. The first layer extracts local spatial features, the second layer captures long-range spatial correlations, and the third layer outputs structural damage sensitivity eigenvalues. The structural damage sensitivity eigenvalue range is set to 0 to 100, and the larger the value, the higher the damage risk.

[0023] As a further solution of the present invention: the specific method of identifying the potential damage area includes:

[0024] A damage early warning model based on strain field evolution trend is established, and the structural damage sensitivity characteristic values of continuous multiple monitoring periods are formed into a time series; the trend component and the mutation component in the time series are extracted through wavelet transform; when the slope of the trend component exceeds the preset threshold, it is determined that it is a gradual damage area; when the amplitude of the mutation component exceeds three times the noise level, it is determined that it is a sudden damage area; finally, the spatial coordinates of the two types of damage areas are compared with the dam structure drawing to mark the range of the potential damage area that needs to be focused on.

[0025] As a further scheme of the application: the construction process of the dam health assessment model is:

[0026] A double-channel feature fusion network based on attention mechanism is established, wherein the first channel processes the interface bonding degradation characteristic value, and the second channel processes the structural damage sensitivity characteristic value; three feature extraction modules are arranged in each channel, the first layer adopts a sliding window to extract local time sequence features, the second layer captures periodic change rules through autocorrelation analysis, and the third layer generates high-order abstract features using a nonlinear activation function; then the feature vectors output by the double channels are subjected to cross-attention calculation to generate a comprehensive safety feature vector with space-time correlation; finally, a deep neural network evaluator containing five hidden layers is constructed, the first hidden layer performs standardization processing on the input features, the second to fourth hidden layers set differentiated weights corresponding to the physical properties of different parts of the dam, and the fifth hidden layer outputs the safety assessment result through state compression.

[0027] As a further scheme of the application: the dynamic adjustment of the monitoring and early warning strategy specifically includes:

[0028] A three-level linkage early warning decision tree is constructed, the first node of the decision tree corresponds to the first threshold trigger condition, when the comprehensive safety feature value exceeds the first threshold, the system automatically starts the redundant sensor cross-validation process, and the data consistency of at least three different types of sensors is compared to confirm the data reliability; the second node of the decision tree corresponds to the second threshold trigger condition, when the feature value exceeds the second threshold, the system automatically activates the multi-source data fusion engine, the multi-source data fusion engine performs space-time alignment on the structure response parameters, and generates a verification conclusion using a data assimilation algorithm based on physical mechanism; the third node of the decision tree corresponds to the third threshold trigger condition, when the feature value exceeds the third threshold, the system automatically matches the preset emergency response plan library, generates a graded early warning signal according to the specific part of the dam and the damage type, and simultaneously links downstream early warning systems and emergency management platforms through a special communication protocol.

[0029] A water conservancy project dam safety monitoring and early warning system, comprising:

[0030] A data acquisition module, which collects structural response parameters in real time through a composite sensor array deployed at key locations of the dam body. The structural response parameters include micromechanical resonant frequency offset and distributed optical fiber strain gradient;

[0031] An interface state evaluation module, which performs a time-frequency joint analysis on the micromechanical resonant frequency offset, constructs an interface peeling characteristic spectrum, and calculates an interface adhesion degradation characteristic value for evaluating the mechanical coupling state of the sensor-dam interface;

[0032] A damage identification module, which performs spatial correlation modeling on distributed optical fiber strain gradients, constructs a strain field anomaly distribution matrix, and calculates structural damage sensitivity eigenvalues ​​to identify potential damage areas within the dam body;

[0033] A safety assessment module, which fuses the interface adhesion degradation characteristic value and the structural damage sensitivity characteristic value into a comprehensive safety characteristic vector, inputs it into a pre-trained dam health assessment model for multi-dimensional analysis, and performs a safety status classification assessment based on the analysis results;

[0034] An early warning decision module dynamically adjusts the monitoring and early warning strategy according to the evaluation results: when the comprehensive safety feature value exceeds the first threshold, the data credibility check is triggered; when it exceeds the second threshold, the multi-source data fusion verification is started; when it exceeds the third threshold, a graded early warning signal is generated and the emergency response mechanism is linked.

[0035] Beneficial effects of the present invention:

[0036] (1) This invention uses the spatial coordinated deployment of micromechanical resonant sensor arrays and distributed optical fiber sensor networks to construct a multi-scale, multi-dimensional three-dimensional monitoring system: the micromechanical resonant sensor accurately captures the microscopic peeling phenomenon of the sensor-dam interface through its high-frequency vibration response characteristics. Among them, the three-dimensional interface peeling characteristic spectrum analysis technology integrates multi-modal frequency shift characteristics and ambient temperature compensation coefficients; while the distributed optical fiber sensor network, relying on its fully distributed measurement advantages, realizes the continuous spatial analysis of the strain field inside the dam body. The strain field anomaly distribution matrix based on spatial grid processing and finite element model weighting can effectively distinguish between real structural damage and environmental interference. The complementary advantages and cross-validation of the two sensing technologies not only solve the problem of large blind spots in traditional point sensor monitoring, but also realize the full chain monitoring from interface micro-damage to structural macro-defects through the fusion analysis of physical mechanism and data drive.

[0037] (2) The present invention constructs a three-level linkage early warning system based on intelligent decision-making, and realizes the precision and intelligence of dam safety early warning through a multi-dimensional risk assessment mechanism driven by deep learning. The system adopts a closed-loop architecture, in which a dual-channel neural network based on the attention mechanism conducts in-depth correlation analysis on interface degradation characteristics and structural damage characteristics. The output comprehensive safety characteristic value triggers a differentiated response through a dynamically optimized three-level threshold system: the primary early warning stage initiates multi-source data cross-validation including resonant sensors, fiber optic sensors and inclinometers to ensure the reliability of abnormal signals; the intermediate early warning activates the physical mechanism model that integrates environmental loads and structural responses to achieve accurate diagnosis of the cause of abnormalities; the advanced early warning automatically matches the optimal disposal plan based on the digital plan library and realizes linkage with the emergency management system through a dedicated communication protocol. By analyzing the historical early warning accuracy and material aging patterns, adaptive optimization of early warning sensitivity is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The present invention will be further described below with reference to the accompanying drawings.

[0039] Figure 1 This is a flow chart of a water conservancy project dam safety monitoring and early warning method of the present invention;

[0040] Figure 2 It is a flow chart of a water conservancy project dam safety monitoring and early warning system in the present invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0042] See also Figure 1 As shown, the present invention is a water conservancy project dam safety monitoring and early warning method, comprising the following steps:

[0043] S1: A composite sensor array deployed at key locations on the dam body collects structural response parameters in real time, including micromechanical resonant frequency offset and distributed optical fiber strain gradient.

[0044] S2: Perform a time-frequency joint analysis of the micromechanical resonant frequency offset, construct an interface peeling characteristic spectrum, and calculate the interface adhesion degradation characteristic value to evaluate the mechanical coupling state of the sensor and dam interface;

[0045] S3: Model the spatial correlation of distributed optical fiber strain gradients, construct a strain field anomaly distribution matrix, and calculate the structural damage sensitivity eigenvalue to identify potential damage areas inside the dam body;

[0046] S4: The interface adhesion degradation eigenvalue and the structural damage sensitivity eigenvalue are integrated into a comprehensive safety feature vector, which is then input into the pre-trained dam health assessment model for multi-dimensional analysis. The safety status is then graded and assessed based on the analysis results.

[0047] S5: Dynamically adjust the monitoring and early warning strategy based on the assessment results: When the comprehensive security feature value exceeds the first threshold, data credibility verification is triggered; when it exceeds the second threshold, multi-source data fusion verification is initiated; when it exceeds the third threshold, a graded early warning signal is generated and the emergency response mechanism is linked;

[0048] The comprehensive security feature value is represented in the form of a multi-dimensional comprehensive security feature vector.

[0049] In S1, a composite sensor array deployed at key locations on the dam body collects structural response parameters in real time. These parameters include micromechanical resonant frequency offset and distributed optical fiber strain gradient, specifically:

[0050] First, the selection of key areas of the dam body requires comprehensive consideration of the structural characteristics and mechanical properties of the dam body. For concrete gravity dams, key areas include the heel, toe, middle of the dam body, and the area surrounding the corridor; for arch dams, the focus is on the abutment, crown beam, and foundation contact zone. These areas are selected based on the results of finite element stress analysis, with areas with high stress concentration and deformation sensitivity selected as key monitoring areas. In actual construction, the specific location and orientation of each sensor must be determined in combination with the design drawings and site conditions to ensure that the key response characteristics of the dam body can be accurately captured.

[0051] The deployment of the composite sensor array adopts a layered layout strategy. Inside the dam body, a monitoring section is set up every 8 to 10 meters in the height direction, and 3 to 5 monitoring points are arranged in each section to form a three-dimensional monitoring network. Each monitoring point is equipped with an integrated sensor unit, which contains two sensing elements: a micromechanical resonant sensor and a distributed optical fiber sensor. The micromechanical resonant sensor is manufactured using MEMS technology, with a size of 10mm×10mm×5mm. The resonant frequency is designed to be in the range of 1-10MHz and has a temperature self-compensation function. The distributed optical fiber sensor uses special optical fiber with a core diameter of 50μm and an outer diameter of 125μm. The coating has alkali corrosion resistance to ensure long-term and stable operation in a concrete environment.

[0052] The process for collecting the micromechanical resonant frequency offset is as follows: At the beginning of each monitoring cycle, the system sends an excitation signal of a specific frequency to the micromechanical resonant sensor, stimulating the resonator to vibrate through the piezoelectric effect. After amplification and filtering, the vibration signal is then measured by a high-precision frequency counter. Multiple averaging is used during the measurement process to eliminate random errors, achieving a frequency resolution of 0.1 Hz. The frequency offset is calculated by comparing the current measured value with the initial baseline value. To improve reliability, three resonators are positioned in orthogonal directions at each monitoring point to measure the vibration response in different directions.

[0053] Distributed fiber-optic strain gradient acquisition utilizes optical frequency domain reflectometry. The system transmits a frequency-modulated continuous-wave light wave to a sensing fiber embedded in the dam body and measures the strain distribution by detecting the spectral shift of the backscattered Rayleigh light. The measurement spatial resolution is 10 cm, and the strain measurement accuracy reaches ±2 με. A fiber loop arranged in a U-shaped pattern is arranged across each monitoring section to ensure that changes in the three-dimensional strain field are captured. After preprocessing the fiber-optic measurement data, the strain gradient distribution along the fiber length is calculated.

[0054] The data acquisition system's hardware configuration includes: a data acquisition unit installed inside the dam, powered by an industrial-grade embedded processor and featuring a waterproof and dustproof design; a communication system utilizing both optical fiber and wireless transmission channels to ensure reliable data transmission; and a hybrid power supply system using solar cells and batteries to ensure continuous system operation. All data acquisition equipment is grounded for lightning protection and equipped with overvoltage and overcurrent protection circuits.

[0055] During data collection, the system employs an adaptive sampling strategy. Under normal circumstances, complete data is collected every hour. When an anomaly is detected, the system automatically switches to a high-density data collection mode, collecting data up to once a minute. After preliminary verification, the collected raw data is transmitted to the data center via an encrypted communication protocol. Data verification includes three levels of verification: range checking, continuity checking, and correlation checking, ensuring data reliability and consistency.

[0056] To ensure long-term monitoring accuracy, the system features an automatic calibration function. This process is automatically performed every three months: the micromechanical resonant sensor uses a built-in reference mass for self-calibration, while the distributed fiber optic sensor uses a standard temperature field provided by the temperature control unit for calibration. Calibration data is stored separately for subsequent data correction and compensation.

[0057] Preprocessing monitoring data involves the following steps: first, outlier removal, using a sliding window method to identify and replace measurements that significantly deviate from the normal range; second, temperature compensation, using readings from temperature probes embedded near the sensors to correct for temperature effects on the measurement results; and finally, time alignment, ensuring that data collected by different sensors has the same time base. The preprocessed data is stored in a time series database for subsequent analysis.

[0058] The system also features a data quality assessment function. This function builds a historical performance profile for each sensor and, combined with metrics such as noise level and stability of the current measurement data, assesses data credibility in real time. When the data quality score falls below a set threshold, the system automatically triggers data re-collection or an alarm to ensure the reliability of the monitoring data. The data quality assessment results serve as an important reference for subsequent analysis.

[0059] Through the above-mentioned implementation, the present invention achieves high-precision and high-reliability acquisition of dam structural response parameters. The design of the composite sensor array ensures comprehensive monitoring while enhancing data credibility through cross-verification between multiple sensors. The entire acquisition process is highly automated, meeting the needs of long-term dam monitoring and providing a reliable data foundation for subsequent safety assessments.

[0060] In S2, a time-frequency joint analysis is performed on the micromechanical resonant frequency offset, the interface peeling characteristic spectrum is constructed, and the interface adhesion degradation characteristic value is calculated to evaluate the mechanical coupling state of the sensor and dam interface. Specifically, the following are performed:

[0061] First, during the preprocessing phase for the micromechanical resonant frequency offset, the system performs a quality check on the raw frequency data. This check includes data integrity verification, outlier detection, and noise level assessment. The continuously acquired frequency sequence is smoothed using a sliding window method to eliminate random interference. The system also records ambient temperature, humidity, and other operating parameters for each measurement, providing a basis for subsequent temperature compensation. The preprocessed frequency data is stored in a dedicated buffer for further analysis.

[0062] During the empirical mode decomposition (EMD) phase, the system uses an adaptive decomposition algorithm to process the frequency offset data. This algorithm uses an iterative filtering process to decompose the non-stationary frequency signal into a series of intrinsic mode function components. Each component represents a vibration characteristic at a specific time scale, arranged from high to low frequency. Strict stopping criteria are set during the decomposition process to ensure that each component meets local symmetry and zero-crossing requirements. The system automatically records the center frequency and energy contribution of each component to form a preliminary modal characteristic description.

[0063] In the Hilbert transform processing stage, the system constructs the analytic signal for each IMF component separately. By combining the original signal with a 90-degree phase shift, the instantaneous amplitude and frequency of each component are obtained. This process is implemented using digital filtering methods to ensure the accuracy of the phase characteristics. The system pays special attention to the points of abrupt change and trends in the instantaneous frequency of each component, as these features are often closely related to changes in the interface state. The calculated instantaneous frequency characteristics are accurately matched with the sampling time stamp to form a joint time-frequency distribution.

[0064] During the construction of the multi-scale frequency shift feature vector, the system compares the instantaneous frequency of each component with the reference frequency spectrum established during the initial calibration phase. The comparison uses the dynamic time warping algorithm, which considers the nonlinear deformation of frequency features along the time axis. For each time scale, three characteristic indicators are calculated: the relative frequency shift, the duration of the shift, and the rate of change of the shift. These indicators are arranged in order of time scale from fine to coarse, forming a hierarchical feature vector.

[0065] The system obtains real-time temperature field data and determines the accurate temperature value at each sensor location using a three-dimensional interpolation algorithm. The temperature compensation coefficient is dynamically calculated based on the material thermal expansion characteristic curve, taking into account the difference in thermal expansion coefficients between concrete and sensor materials. The compensated frequency shift feature vector eliminates false signals caused by temperature, providing a more accurate reflection of the interface mechanical state.

[0066] The construction of the three-dimensional interface debonding feature spectrum uses a tensor representation method. The first dimension, the time scale axis, is distributed in a logarithmic manner, covering feature changes from seconds to months; the second dimension, the frequency shift axis, is normalized to unify the dimensions of different sensors; and the third dimension, the temperature compensation coefficient axis, is linearly distributed according to actual measurement values. Each voxel of the feature spectrum stores the feature intensity value under the corresponding conditions, forming a complete interface state representation system.

[0067] In the normal vibration modal analysis stage, the system extracts the energy distribution within a specific frequency band range from the feature spectrum. This frequency band corresponds to the dominant modal of the sensor's normal vibration, usually within the 1-5 MHz range. The energy distribution calculation uses the wavelet packet transform method, dividing the time-frequency plane into multiple analysis windows. The relative entropy value calculation is based on the Kullback-Leibler divergence principle, comparing the difference between the current energy distribution and the reference state. This parameter can sensitively reflect changes in the interface normal stiffness.

[0068] The tangential vibration modal analysis focuses on the changes in harmonic components. The system extracts the amplitude ratio of the third and fifth harmonics through high-order spectrum analysis techniques. The calculation of the total harmonic distortion rate uses the total harmonic distortion formula, but removes the step of normalizing the fundamental energy, preserving the absolute value information. This parameter has specific response characteristics for interface tangential slip and local debonding.

[0069] The feature parameter fusion process utilizes an adaptive weighting strategy. The weight coefficients are dynamically adjusted based on the sensor's location within the dam: For sensors located at the foundation, normal-direction features are given a higher weight; for sensors located in the middle of the dam, features in both directions are balanced. The influence of concrete age is considered using a time-varying function, prioritizing harmonic features in the early stages and energy features in the later stages. The fused composite index is mapped to a range of 0–1 using a sigmoid function, forming a standardized characteristic value for interfacial bond degradation.

[0070] The deep belief network is constructed using a layered training strategy. The first layer, a restricted Boltzmann machine, focuses on learning the statistical laws of frequency domain features. The visible layer units correspond to energy values ​​in different frequency bands, and the hidden layer units are trained using a contrastive divergence algorithm. The second layer, a restricted Boltzmann machine, processes time domain features. Its network structure takes into account the Markov characteristics of time series. The third layer, a restricted Boltzmann machine, exploits the nonlinear coupling relationship between frequency and time domain features.

[0071] A two-stage strategy is employed during model training. The pre-training phase uses a large amount of numerical simulation data to initially determine network parameters through unsupervised learning. The fine-tuning phase utilizes real-world engineering monitoring data and supervised training using a backpropagation algorithm. Early stopping and regularization terms are incorporated into the training process to prevent overfitting. The resulting model is capable of accurately identifying subtle changes in interface state.

[0072] The state classification output is represented probabilistically. A softmax function maps hidden layer features to probability distributions across four state categories. Classification thresholds are dynamically set based on engineering experience: intact state corresponds to a probability value greater than 0.9; mild delamination corresponds to a probability value between 0.7 and 0.9; moderate delamination corresponds to a probability value between 0.4 and 0.7; and severe delamination corresponds to a probability value less than 0.4. The system records the temporal changes in state probabilities and, if a trend of continued deterioration is detected, issues an early warning even if the threshold has not been reached.

[0073] For minor peeling conditions, the system will log the event locally and prompt regular inspections. Moderate peeling triggers a remote alarm to notify technicians. Severe peeling directly triggers the emergency response system. All warning events automatically generate detailed analysis reports, including characteristic spectrum change graphs, state evolution curves, and potential impact assessments.

[0074] The system's self-learning function continuously optimizes evaluation performance. After each manual check and confirmation of the interface status, the system compares the actual results with the predicted results and automatically adjusts the model parameters. Long-term accumulation of engineering cases forms a knowledge base that is used to refine feature extraction algorithms and classification rules. This closed-loop learning mechanism ensures that the system's evaluation accuracy continues to improve over time.

[0075] First, during the preprocessing phase for micromechanical resonant frequency offset, the system performs a quality check on the raw frequency data. This process includes checking data integrity, removing significant outliers, and performing preliminary smoothing. The resonant frequency data for each monitoring point in three orthogonal directions (X, Y, and Z) are stored separately and tagged with the corresponding timestamp and ambient temperature. The system establishes a baseline frequency database for each sensor. This baseline value is the average value determined after 30 days of stable monitoring at the initial installation of the sensor. All subsequent frequency offsets are calculated relative to this baseline value.

[0076] Empirical mode decomposition is one of the key technical links of the present invention. The system uses an improved empirical mode decomposition algorithm assisted by adaptive noise to process the frequency offset time series. During the decomposition process, the algorithm automatically determines the required number of decomposition layers, which is 6-8 layers. The intrinsic mode function components obtained by each layer of decomposition have clear physical meanings: high-frequency components reflect the impact of short-term environmental fluctuations, medium-frequency components reflect the dynamic response characteristics of the structure, and low-frequency components represent long-term trend changes. After the decomposition is completed, the system will verify the physical rationality of each component to ensure that the decomposition results are consistent with practical engineering significance.

[0077] During the Hilbert transform processing stage, the system constructs an analytical signal for each intrinsic mode function component. This process accurately extracts the instantaneous frequency and amplitude characteristics of the signal. Specifically, the present invention employs a sliding window Hilbert transform method, where the window length is adaptively adjusted based on component characteristics to balance the requirements of time resolution and frequency resolution. A shorter window is used for high-frequency components to improve time resolution, while a longer window is used for low-frequency components to ensure the accuracy of frequency analysis. The transform results in a time series of instantaneous frequencies for each component, which serves as the basis for subsequent analysis.

[0078] The system compares the instantaneous frequency of each component with the corresponding reference frequency to calculate the relative offset. The reference frequency is the frequency characteristic of each component under normal conditions, obtained through long-term monitoring data. The comparison process not only considers the absolute value of the offset but also analyzes the duration and trend of the offset. For each monitoring point, the system generates a feature vector containing the frequency offset characteristics at different time scales. These features are arranged from high to low frequency to form a complete frequency domain feature representation.

[0079] Temperature compensation is crucial for ensuring monitoring accuracy. The system simultaneously collects temperature data near each sensor and develops a temperature-frequency offset compensation model. This model, based on extensive experimental data, accounts for the thermal expansion characteristics of both concrete and sensor materials. When constructing the 3D interface debonding spectrum, the system converts the original frequency offset into a temperature-compensated equivalent value to eliminate interference caused by temperature fluctuations. The compensation coefficient takes into account the rate of temperature change and historical trends to avoid over- or under-compensation.

[0080] The construction of a three-dimensional interface debonding spectrum involves the following steps: the first dimension, the time scale, corresponds to the intrinsic mode function components of different frequency bands, arranged from fastest to slowest; the second dimension, the frequency offset, represents the actual offset of each component after temperature compensation; and the third dimension, the temperature compensation coefficient, reflects the influence of ambient temperature on the measurement results. This three-dimensional representation comprehensively reflects the time-frequency characteristics of the interface state, providing a rich information foundation for subsequent analysis.

[0081] During the normal vibration modal analysis phase, the system focuses on vibration characteristics perpendicular to the dam surface. The energy distribution of the primary frequency band is calculated using an improved power spectrum estimation method, with spectral resolution enhanced through windowing. The relative entropy value compares the energy distribution difference between the current state and the baseline state, a parameter that is particularly sensitive to early interface delamination. During the calculation process, the system considers the weights of different frequency bands, focusing on the characteristic frequency bands most sensitive to the interface state.

[0082] Tangential vibration modal analysis focuses on vibration characteristics parallel to the dam surface. Harmonic distortion is calculated by analyzing the degree of waveform distortion in the vibration signal. The system establishes a standard sinusoidal reference signal and quantifies the degree of distortion by comparing the measured signal with the reference signal. This parameter is particularly sensitive to interfacial shear slip and can detect even the earliest signs of microslip. The analysis takes into account the sensor's mounting orientation to ensure accurate capture of the tangential vibration characteristics.

[0083] The feature parameter fusion process utilizes a dynamic weighting strategy. The weighting coefficients depend not only on the sensor's depth and the concrete's age, but also on the current environmental conditions and the consistency of historical monitoring data. Deeper sensors are given a higher weight because they are less affected by the surface environment. For older concrete structures, the weighting of tangential vibration is appropriately reduced, as the shear properties are likely to have stabilized. The weighting algorithm is based on statistical analysis of numerous engineering cases to ensure the reliability of the fusion results.

[0084] The interface state mapping function is a nonlinear relationship model established through machine learning methods. During training, a large amount of laboratory simulation data and actual engineering cases are used, which can accurately map the feature parameter space to the state evaluation space. The output value of the function is normalized to a range of 0 to 1, facilitating unified evaluation standards. The mapping process takes into account the particularity of different sensor positions, ensuring spatial consistency of the evaluation results.

[0085] The construction of a deep belief network is a complex process. The first layer of restricted Boltzmann machines is responsible for learning the representation of frequency domain features, and the number of hidden nodes is usually set to 2-3 times the number of input features; the second layer focuses on capturing the evolution law of the time domain, using a network structure with memory function; the third layer learns the nonlinear coupling relationship between different features. The training of each layer uses the contrastive divergence algorithm, and the network parameters are optimized through multiple iterations. After pre-training, actual engineering data is used for fine-tuning to improve the generalization ability of the model.

[0086] The design of the model output layer takes into account the actual engineering needs. The threshold values for the four state classifications are based on statistical analysis of a large amount of measured data: the perfect state corresponds to an output value less than 0.3, the slight peeling is 0.3-0.6, the moderate peeling is 0.6-0.8, and the severe peeling is greater than 0.8. The calculation of output probability uses the softmax function to ensure that the sum of the probabilities of each category is 1. The system records the state probability vector at each evaluation time, forming a complete time-varying state trajectory.

[0087] The triggering mechanism of the early warning signal adopts a multi-condition judgment strategy. In addition to considering the state probability at the current time, it also analyzes the duration and trend of the state change. If the slight peeling state lasts more than 7 days or the moderate peeling state lasts more than 3 days, an early warning will be triggered. For severe peeling state, the system will immediately trigger the highest level of warning and start the emergency response process. The early warning information includes specific location coordinates, evaluation results and suggested measures, facilitating quick response by engineering personnel.

[0088] The recording of abnormal time series features uses compression storage technology. The system automatically identifies the starting point, peak point and ending point of abnormal events, and records key feature parameters. These data are not only used for real-time warning, but also stored in the case library to support subsequent model optimization. The storage format is designed with standardization to ensure long-term readability and analyzability of the data.

[0089] By analyzing the time distribution, spatial correlation and evolution law of abnormal events, the system can predict the potential development trend of faults. The analysis results generate a trend report, including possible cause analysis, development prediction and disposal suggestions. These information has important reference value for long-term safety evaluation and maintenance decision of dams.

[0090] The online updating mechanism of the model ensures accuracy over a long period of use. The system regularly collects new monitoring data and automatically fine-tunes the model parameters. The updating process uses an incremental learning algorithm to avoid the computational burden of retraining. At the same time, the system retains historical model versions to facilitate result comparison and abnormality tracing. The updating frequency is dynamically adjusted based on data quality and usage effectiveness, with a small-scale update every quarter and a comprehensive optimization every year.

[0091] The visualization interface provides an intuitive state display for engineers. The three-dimensional dam model is marked with different colors to indicate the interface state of each monitoring point, supporting multi-time point comparison and viewing. Detailed analysis results can be displayed by clicking on specific monitoring points, including characteristic spectrum, state evolution curve, and evaluation report. The system also provides multiple data export formats for easy data exchange with other analysis software.

[0092] Quality control is implemented throughout the entire analysis process. From data input to result output, each link has a quality check point. Abnormal data is marked and recorded for processing, and key analysis steps save intermediate results for reference. The system regularly generates quality assessment reports to summarize the operation status and data quality of each link, providing a basis for system maintenance and upgrading.

[0093] Compared with traditional methods, the implementation process of the invention has significant advantages. Time-frequency joint analysis can capture both short-term mutations and long-term trends; three-dimensional characteristic spectrum provides more comprehensive state representation; deep belief networks can learn complex nonlinear relationships; and dynamic updating mechanism ensures the accuracy of long-term monitoring.

[0094] The fault-tolerant design of the system ensures stable operation in abnormal situations. When some sensors fail, the system can interpolate and compensate based on the data of adjacent monitoring points; when communication is interrupted, local devices can continue to collect data and store temporarily, and synchronize after communication is restored; when power fails, backup power can maintain key functions for at least 72 hours.

[0095] The user permission management module ensures system security. Different levels of users are granted different operation permissions, and key functions require multiple authentications. All operations are logged, including operator, time, and specific content. Data access uses encrypted transmission, and storage is encrypted to prevent unauthorized access and tampering.

[0096] In S3, spatial correlation modeling of distributed optical fiber strain gradient is performed to construct a strain field anomaly distribution matrix and calculate structural damage sensitivity eigenvalues for identifying potential damage areas inside the dam, including:

[0097] First, during the preprocessing stage of distributed optical fiber strain data, the system performs strict quality control checks on the raw collected strain data. This process includes checking the integrity of the optical fiber link, eliminating measurement noise, and compensating for temperature effects. The measurement data of each sensing fiber is processed through a sliding average filter, and the window length is set to 5 measurement points (corresponding to an actual distance of 50 cm) based on the spatial resolution requirements. At the same time, the system synchronously collects ambient temperature data and performs temperature compensation based on a pre-calibrated temperature-strain relationship model. The compensated strain data is converted into a change relative to the initial reference state to form a strain gradient distribution curve. For long-distance distributed measurements, the system adopts a segmented compensation strategy, dividing the entire optical fiber path into several segments, and establishing a separate temperature compensation model for each segment to improve compensation accuracy.

[0098] Spatial gridding is a fundamental step in constructing a strain field model. The system establishes a three-dimensional spatial grid coordinate system based on the structural characteristics of the dam. The grid size is determined by taking into account both the measurement spatial resolution (minimum 10 cm) and computational efficiency requirements, and is typically set to 20 cm × 20 cm × 50 cm (length × width × height). Each grid cell contains data from multiple fiber optic measurement points. The system calculates the strain gradient values ​​for all measurement points within the cell and calculates their average, standard deviation, and coefficient of variation. Special grid cells that span structural joints or material interfaces are marked and processed using special methods. The gridding process fully considers the internal structural characteristics of the dam. For example, specialized grid densification areas are set up for special structural locations such as corridors and drainage pipes.

[0099] The construction of the initial strain distribution matrix adopts a multi-index comprehensive evaluation method. In addition to calculating the coefficient of variation of the strain gradient for each grid cell, the system also considers the following characteristic parameters: the absolute value of the strain gradient, the rate of strain change between adjacent measuring points, and the strain-temperature correlation coefficient. After normalization, these parameters are linearly combined according to predetermined weights to form the initial matrix element values. The row and column indices of the matrix strictly correspond to the spatial grid coordinates, ensuring that the matrix can accurately reflect the spatial distribution characteristics of the strain field within the dam body. For grid cells that lack direct measurement data, the system performs spatial interpolation based on the data of adjacent cells. The interpolation method uses the Kriging algorithm that takes material properties into consideration.

[0100] The system calculates the strain transfer coefficient between each grid cell and its 26 spatial neighbors (directly adjacent cells in three-dimensional space). This calculation not only considers spatial distance but also incorporates engineering parameters such as material uniformity and structural continuity. By constructing a spatial weight matrix, the system can identify and correct isolated outliers in the initial matrix. This correction process utilizes an iterative algorithm, recalculating spatial correlation indices with each iteration until the matrix element values ​​stabilize. This approach effectively distinguishes true localized damage from outliers caused by measurement noise, significantly improving the reliability of the analysis results.

[0101] The finite element model fusion stage achieves an organic combination of measured data and theoretical analysis. The system loads a pre-established parametric finite element model of the dam body, which contains detailed material parameters and boundary condition settings. Under normal operating conditions, the stress concentration factor calculated by the model will be extracted and applied as a weighting factor to the strain field anomaly distribution matrix. The weighting process uses a nonlinear mapping method to ensure that anomalies in high-stress areas are highlighted, while anomalies in low-stress areas are appropriately suppressed. This processing method ensures that the final strain field anomaly distribution matrix not only reflects the measured strain characteristics, but also integrates the analysis results of structural mechanics theory, with a clearer physical meaning.

[0102] Highly anomaly areas are identified using dynamic thresholding technology. The system automatically determines the outlier threshold based on the current monitoring period and statistical analysis of historical data. The traditional fixed threshold of two standard deviations has been improved to an adaptive threshold algorithm that accounts for spatiotemporal correlations: a stricter threshold (e.g., 1.5 standard deviations) is used for areas with stable historical data, while a more relaxed threshold (e.g., 2.5 standard deviations) is used for areas with larger fluctuations. Identified highly anomaly areas undergo morphological processing, including region growing algorithms to connect adjacent anomaly points and remove isolated anomaly areas with excessively small areas, ultimately forming a complete distribution map of the highly anomaly areas.

[0103] The calculation of primary features focuses on engineering practicality. The area ratio feature is calculated by counting the proportion of high-anomaly areas in the entire analysis area. This parameter reflects the extent of the damage. The spatial aggregation feature is characterized by calculating the spatial distribution entropy of high-anomaly areas. Lower entropy values ​​indicate more concentrated anomalies, potentially indicating severe localized damage. The combination of these two primary features can provide a preliminary assessment of the damage type: large areas with low aggregation may indicate uniform degradation, while small areas with high aggregation may indicate localized cracking or damage.

[0104] The analysis of secondary features deeply explores the strain direction information. Based on the results of finite element analysis, the system will establish a database of the principal stress directions of various parts of the dam under normal operating conditions. For each highly abnormal area, the system will calculate the angle between the average strain gradient direction and the theoretical principal stress direction in the area. This deviation angle parameter is particularly sensitive to identifying abnormal stress states. For example, shear failure usually causes a significant deviation between the strain direction and the principal stress direction. The angle calculation uses a three-dimensional space vector analysis method to ensure the accuracy of the results. At the same time, the system will also analyze the degree of discreteness of the strain gradient direction. This parameter can reflect the degree of stress disorder in the damaged area.

[0105] The design of the damage sensitivity assessment network fully considers the needs of engineering applications. The first layer of the convolutional neural network uses a small convolution kernel (3×3×3) to focus on extracting local spatial features. The second layer uses a larger convolution kernel (7×7×7) combined with dilated convolution techniques to effectively expand the receptive field and capture long-range spatial correlations. The third layer replaces the traditional fully connected layer with global average pooling, reducing the number of parameters while enhancing the model's spatial invariance. The network's training data is derived from a large number of historical engineering cases and numerical simulation results, including strain field characteristics under various typical damage modes. The calibration of the structural damage sensitivity eigenvalue is based on engineering risk assessment theory. The system divides the value range from 0 to 100 into five risk levels: 0-20 represents normal status, 20-40 represents slight abnormality, 40-60 represents moderate risk, 60-80 represents high risk, and 80-100 represents extremely high risk. This eigenvalue not only considers current monitoring results but also incorporates damage development trend predictions, forming a comprehensive risk assessment indicator. The calculation process of the eigenvalue is explainable, and the system can generate a detailed evaluation report to illustrate the specific contribution of each influencing factor.

[0106] Time series analysis is implemented using a multi-scale processing approach. The system maintains a sliding time window (default length is 30 monitoring cycles) and performs trend analysis on the structural damage sensitivity eigenvalues ​​within the window. The wavelet transform basis function selection takes into account the characteristics of engineering signals, using the Daubechies wavelet with excellent time-frequency localization properties. Trend components are extracted through low-pass filtering to reflect the long-term evolution of damage; mutation components are obtained by reconstructing high-frequency components to capture short-term abnormal events. During the analysis process, the system automatically optimizes the number of wavelet decomposition layers to accommodate damage development processes with different change rates.

[0107] The determination of progressive damage areas utilizes a multi-evidence fusion strategy. In addition to monitoring the slope of the trend component, the system also considers factors such as the duration and consistency of the slope, coordinated changes in spatially adjacent areas, and correlation with environmental loads. Only when multiple evidence indicators meet pre-set criteria is a final determination of progressive damage made. This conservative strategy effectively reduces false alarm rates and ensures the reliability of early warning information. The determination results are accompanied by a confidence assessment, reflecting the degree of certainty of the judgment.

[0108] Identification of sudden damage areas prioritizes timeliness. The system incorporates a rapid detection channel that immediately initiates a specialized analysis process when a sudden change in the strain signal is detected. Noise levels are estimated using robust statistical methods, based on the fluctuation characteristics of data under recent stable conditions. To enhance detection sensitivity, the system prioritizes sudden changes that occur spatially and exhibit similar characteristics. Once sudden damage is confirmed, the system immediately increases monitoring frequency and initiates emergency response plans.

[0109] The spatial localization of damaged areas utilizes a multi-level verification approach. The system accurately compares the coordinates of the damaged areas obtained from the analysis with the dam structural drawings, taking into account factors such as the location of structural joints, material zoning, and the distribution of construction joints. For damage signals appearing in key locations, the system automatically retrieves the design parameters and construction records for in-depth analysis. The localization results are displayed in a 3D visualization format, supporting multi-angle viewing and cross-sectional analysis, enabling engineers to accurately understand the damage location and its relationship to the structure.

[0110] The early warning information generation process prioritizes practicality. Based on the damage type, extent, and location, the system automatically generates a warning report containing the following content: damage description, possible cause analysis, risk level assessment, and recommended actions. This report utilizes a standardized template to ensure clarity and accuracy. The system also provides detailed supporting data, including raw data from relevant monitoring points, intermediate results from the analysis process, and historical comparative data, for further review by professionals.

[0111] Compared with traditional methods, the implementation process of the present invention has significant advantages: spatial correlation modeling improves the accuracy of damage identification; multi-feature fusion analysis enhances the comprehensiveness of damage assessment; the application of neural networks realizes the intelligent recognition of complex patterns; time series analysis can capture the development trend of damage. The system's adaptive learning mechanism ensures the effectiveness of long-term monitoring. As the operating time accumulates, the system will continuously collect new monitoring data and automatically optimize the analysis model parameters. The learning process uses an online incremental algorithm and will not affect the real-time monitoring function of the system. At the same time, the system will regularly conduct performance evaluations, statistically analyze the accuracy and false alarm rate of various indicators, and provide a basis for subsequent improvements. This self-improvement ability enables the system to adapt to various changes in the long-term operation of the dam.

[0112] The visualization analysis platform provides an intuitive operation interface for engineers. The system supports multiple data display methods: damage distribution cloud map on the three-dimensional dam model, time history curve of key parameters, spatial profile analysis view, etc. All graphics support interactive operation, which can freely adjust the viewing angle, zoom ratio and display parameters. The analysis results can be exported in multiple standard formats, facilitating data exchange and integrated analysis with other professional software.

[0113] Compared with the prior art, the embodiment firstly applies spatial correlation modeling and neural network technology to dam damage identification; proposes the concept of strain field abnormal distribution matrix; develops a multi-level analysis process specifically for dam damage assessment; realizes a complete intelligent solution from strain monitoring to damage warning. The precision and efficiency of dam safety monitoring have been significantly improved, providing a new technical means for engineering safety management.

[0114] In S4, the interface bonding degradation characteristic value and the structure damage sensitivity characteristic value are fused into a comprehensive safety characteristic vector, which is input into the pre-trained dam health assessment model for multi-dimensional analysis, and the safety state is classified and evaluated according to the analysis results, which specifically includes:

[0115] Firstly, in the data preprocessing stage before the characteristic value fusion, the system will normalize the interface bonding degradation characteristic value and the structure damage sensitivity characteristic value. The normalization method uses the improved RobustScaler algorithm, which can effectively eliminate the influence of different dimensions based on the statistical distribution characteristics of the characteristic values, while retaining the effective information of outliers. The preprocessing process also includes time alignment processing to ensure that the two characteristic value sequences have the same time reference. For time points with missing data, the system uses a multiple interpolation method considering seasonality and trend to fill in the missing data, ensuring data continuity. Each characteristic value sequence will also be smoothed, using a sliding average method with an adaptive window length, which dynamically adjusts the window length according to the characteristic value change rate, balancing the needs of noise suppression and feature preservation.

[0116] The first channel of the dual-channel feature fusion network specifically processes eigenvalues ​​of interfacial adhesion degradation. The first-layer feature extraction module in this channel utilizes a variable-length sliding window technique. The window length automatically adjusts based on the frequency of eigenvalue changes, ranging from 3 to 21 time points. Multiple statistics, including mean, variance, skewness, and kurtosis, are calculated for the eigenvalues ​​within each window to form a preliminary time series feature vector. The second-layer module uses autocorrelation analysis to uncover periodic patterns and calculate autocorrelation coefficients at different time lags, focusing on analyzing the strength of correlation within key periods such as 24 hours (daily cycle) and 168 hours (weekly cycle). The third-layer module uses a combined activation function, first performing a nonlinear transformation using the Sigmoid function, then enhancing the feature representation using the Swish function, ultimately outputting a 128-dimensional high-order abstract feature vector.

[0117] The second channel is specifically optimized for processing structural damage sensitivity eigenvalues. The first layer of this channel also uses sliding window technology, but with a larger window length to capture long-term trends in damage development. The second layer's autocorrelation analysis focuses on long-term patterns, with a particular focus on cyclical features of 720 hours (monthly cycles) and 2160 hours (seasonal cycles). The third layer uses the GELU activation function, which is more suitable for processing damage features with asymmetric distribution characteristics. A residual connection structure is also implemented within the channel to prevent feature vanishing in deep networks. The network structures of the two channels are symmetrical, but the parameters are independent, ensuring that each channel focuses on extracting different types of features.

[0118] Attention calculations use a scaled dot product approach, but incorporate positional encoding information to preserve the temporal order of eigenvalues. This paper employs asymmetric attention weighting to prioritize queries against interface feature channels, based on engineering experience that reveals that interface issues often precede structural damage. The output of the attention layer undergoes layer normalization and feedforward network processing to form a 256-dimensional comprehensive security feature vector, which incorporates both temporal and spatial correlation information.

[0119] The construction of the deep neural network evaluator takes into account the physical characteristics of the dam structure. When the first hidden layer performs feature standardization, it distinguishes different types of features and adopts different standardization strategies: interface features are standardized using Z-score, and damage features are standardized using Min-Max. The second to fourth hidden layers set differentiated weights according to the parts of the dam body: the base is given a higher weight coefficient to reflect its critical impact on overall safety; the weight of the top of the dam is appropriately reduced because it is more sensitive to temperature changes. Each hidden layer is equipped with a Dropout regularization layer, and the dropout rate is dynamically adjusted according to the training stage. The initial value is 0.3 and gradually decreases to 0.1 as training progresses. The network is designed with a step-by-step decrease in the number of hidden units, which are 512, 256, 128, 64, and 32 neurons, respectively. This structure not only ensures feature extraction capabilities but also prevents overfitting.

[0120] The state compression output layer uses a special combination of activation functions. The 32-dimensional features are first mapped to the range [-1, 1] using the Tanh function, and then converted to a probability distribution using the Softmax function. The output includes probabilities for five safety levels: safe, basically safe, mildly dangerous, moderately dangerous, and severely dangerous. The system records the complete probability distribution vector, not just the category with the highest probability, to retain more information for subsequent analysis. If the difference in probability between categories falls below a threshold, the result is marked as low confidence, indicating that manual review is required.

[0121] The model training process is divided into three phases. The pre-training phase uses large-scale numerical simulation data generated based on finite element models of typical dam types, covering a variety of possible damage scenarios and interface issues. The fine-tuning phase uses actual engineering monitoring data, focusing on optimizing the parameters of the network's later layers. The online learning phase continuously absorbs new monitoring data and uses the momentum SGD algorithm for small-batch parameter updates. The training process uses an early stopping strategy and model checkpointing techniques to ensure optimal generalization performance. The loss function is designed to be weighted cross entropy, which imposes a greater penalty on misjudgments of high-risk levels.

[0122] The numerical simulation data generated ensures engineering realism. The simulation model includes detailed material parameters and boundary conditions, simulating various degradation mechanisms such as concrete aging, steel corrosion, and foundation settlement. Over 1,000 sets of time-history data are generated for each damage condition, covering the entire process from intact to severely damaged. Measurement noise and environmental interference are accounted for during data generation, ensuring that the simulation data closely resembles actual monitoring conditions. Latin hypercube sampling is used for simulation parameters to ensure full coverage of the parameter space.

[0123] The processing of actual project data prioritizes privacy protection and data security. All monitoring data is anonymized before input into the model, removing project identifiers. Data storage utilizes encryption technology, and access requires multiple authentication methods. The system regularly verifies data integrity to prevent accidental corruption or malicious tampering. Data use strictly complies with relevant laws, regulations, and industry standards.

[0124] The model's performance is evaluated using multiple metrics. Classification accuracy primarily assesses the accuracy of safety level judgments; recall focuses on the ability to identify dangerous conditions; false alarm rate controls the frequency of false alarms; and response time ensures that real-time performance requirements are met. The evaluation process utilizes a cross-validation approach, dividing the dataset into training, validation, and test sets. The system regularly generates performance evaluation reports to guide continuous model optimization.

[0125] The system's adaptive capabilities ensure long-term effectiveness. Model parameters are continuously optimized as new monitoring data accumulates, adapting to changing characteristics of the dam during aging. Furthermore, the system regularly benchmarks against the latest engineering specifications to ensure that assessment standards remain current. When a dam undergoes major repairs or renovations, the system initiates a dedicated learning process to rapidly adapt to changes in the structural state.

[0126] The visual analysis interface provides intuitive decision support for engineers. The system displays the temporal evolution of safety assessment results, spatial distribution cloud maps, and contribution analysis of key parameters. All charts support interactive operation and customization. The automatic assessment report generation function outputs industry-standard document formats for easy archiving and reporting.

[0127] In S5, the monitoring and early warning strategy is dynamically adjusted based on the evaluation results: when the comprehensive security feature value exceeds the first threshold, data credibility verification is triggered; when it exceeds the second threshold, multi-source data fusion verification is initiated; when it exceeds the third threshold, a graded early warning signal is generated and the emergency response mechanism is linked. Specifically,

[0128] First, during the construction phase of the three-level coordinated early warning decision tree, the system establishes an initial threshold system based on project safety standards and historical monitoring data. The first threshold is set as the boundary value of the safe state, typically corresponding to the 95th percentile of the comprehensive safety characteristic value; the second threshold reflects a clearly abnormal state, taking the 99th percentile; and the third threshold corresponds to a dangerous state, taking 80% of the historical maximum value. These thresholds are initially set based on the engineering experience of similar dams and are calibrated during the trial operation phase. Each node in the decision tree is equipped with detailed conditional judgment logic and action instructions, forming a complete early warning response process. The system maintains a separate decision tree instance for each monitoring point, ensuring that the early warning strategy can adapt to the characteristics of different locations.

[0129] When the comprehensive safety characteristic value exceeds the first threshold, the system initiates a redundant sensor cross-validation process, which includes multiple verification steps. First, the system checks the data consistency of different sensor types (such as micromechanical resonant sensors, fiber optic sensors, inclinometers, etc.) at the same monitoring point, calculating the correlation coefficient and difference between the sensor data. Next, it analyzes the coordinated changes in adjacent monitoring points to verify the spatial continuity of the abnormal signal. Finally, it compares the current data with historical data patterns under the same operating conditions to assess its rationality over time. The verification process uses a voting mechanism, and when more than two-thirds of the verification indicators pass, the data is deemed credible. The system maintains a complete verification log, including the specific results and judgment basis of each verification step, to support subsequent analysis.

[0130] The multi-source data fusion verification after the second threshold is triggered is a multi-stage analysis process. The system first performs spatiotemporal data alignment, unifying the structural response data and environmental load data (such as water pressure, temperature, and seismic motion) to the same time base and spatial coordinate system. This alignment takes into account differences in data acquisition frequency and transmission delays, and uses cubic spline interpolation to ensure time synchronization accuracy. A physics-based data assimilation algorithm is then applied to optimally fuse the observed data with the finite element model predictions. This assimilation process employs the ensemble Kalman filter method, characterizing uncertainty through multiple sets of model parameters. The verification conclusion not only confirms the anomaly but also assesses possible causes and development trends, providing a more comprehensive basis for subsequent decision-making.

[0131] The emergency response mechanism triggered by the third threshold implements hierarchical and classified disposal. The system will match the most appropriate response plan from the plan library based on the importance of the specific part of the dam, the dangerousness of the damage type, and the current environmental conditions. Each plan in the plan library contains detailed implementation steps, division of responsibilities, and resource requirements. The early warning signals are divided into four levels: blue, yellow, orange, and red, corresponding to attention, warning, alarm, and emergency status respectively. After the signal is generated, the system will push it to downstream systems and relevant responsible personnel in real time through dedicated communication protocols (such as MQTT) to ensure the timeliness and reliability of information transmission. The communication process uses encryption and redundant transmission mechanisms to ensure smooth communication in extreme situations.

[0132] The system continuously collects verification results and subsequent feedback from each warning, calculating its accuracy and false alarm rate. Based on these performance indicators and incorporating material aging factors reflected in the dam's operational life, threshold settings are regularly optimized. The adjustment algorithm employs a conservative strategy, gradually optimizing while ensuring safety: thresholds are appropriately relaxed for high-accuracy warning types and increased for more frequent warnings. Safety constraints are incorporated into the adjustment process to prevent a single, significant change from causing system instability. A detailed change log and justification are generated for each adjustment for management review.

[0133] The data type selection in the redundant sensor cross-validation process is targeted. Based on the nature of the anomaly signature, the system automatically selects the three most relevant sensor types for verification. For suspected interface issues, micromechanical resonant sensors, fiber Bragg grating sensors, and resistance strain gauges are prioritized; for suspected structural damage, distributed optical fiber, inclinometers, and acoustic emission sensors are used. The calculation of verification metrics takes into account differences in sensor characteristics, such as stricter tolerances for higher-precision fiber optic sensors and more relaxed requirements for environmentally sensitive resistance strain gauges. The system maintains a reliability score for each sensor and dynamically adjusts its weight during the verification process.

[0134] The multi-source data fusion engine features intelligent environmental load processing. The system establishes an environmental-structural response correlation model to identify the normal response patterns of the dam under different environmental conditions. When performing data fusion, the significance of the current environmental loads is first assessed, distinguishing between normal and extreme load conditions. For extreme load conditions, a specialized analysis mode is activated to account for special factors such as material nonlinearity. During the data assimilation process, the adjustment range of model parameters is limited by physical constraints to prevent parameter combinations that are inconsistent with engineering practices. The assimilation results are accompanied by uncertainty estimates, reflecting the reliability of the conclusions.

[0135] The emergency response plan library consists of basic templates and interchangeable components, allowing for rapid assembly and generation of customized plans based on specific circumstances. Plans cover multiple dimensions, including enhanced monitoring (such as increasing measurement frequency and initiating special inspections), engineering measures (such as lowering reservoir water levels and implementing temporary reinforcements), and management responses (such as evacuation and expert consultation). The system calculates the suitability score of each plan based on real-time conditions, recommends the optimal option, and provides comparisons of alternatives. During plan execution, the system tracks implementation results and supports dynamic adjustments.

[0136] The grading criteria for early warning signals take into account multiple factors. A blue alert indicates an unconfirmed anomaly requiring increased attention; a yellow alert indicates a confirmed anomaly but manageable risk; an orange alert indicates a high-risk situation requiring immediate action; and a red alert indicates an imminent dam failure risk requiring emergency action. The grading process considers the following indicators: anomaly magnitude, development rate, spatial extent, structural significance, and environmental conditions. The system generates a detailed explanation of the grading basis to avoid inconsistencies caused by subjective judgment.

[0137] The implementation of the communication protocol prioritizes reliability and real-time performance. The system utilizes a publish-subscribe model, allowing warning information to be pushed simultaneously to multiple terminals. Communication content utilizes a standardized data format, including structured information such as timestamp, location coordinates, warning level, detailed description, and recommended actions. For red alerts, the system activates communication assurance mode, automatically switching to a backup channel, increasing transmission priority, and enabling a receipt confirmation mechanism to ensure the delivery of critical information.

[0138] The system's implementation is independent of a specific hardware platform and can be deployed in the cloud for centralized management or in edge computing for rapid local response. For particularly large water conservancy projects, a distributed architecture can be employed, deploying multiple decision-making nodes in different areas and enabling integrated early warning through upper-level coordination. This flexible deployment approach accommodates projects of all sizes.

[0139] The entire early warning decision-making process emphasizes traceability. The system fully records data from threshold triggering to response implementation, including raw monitoring values, intermediate analysis results, decision-making basis, and execution status. These records are stored in an immutable manner, supporting retrospective analysis. Based on this complete data record, the system can regularly generate operational reports, compile statistics on early warning effectiveness indicators, and guide continuous system optimization.

[0140] See also Figure 2 As shown, a water conservancy project dam safety monitoring and early warning system includes:

[0141] A data acquisition module, which collects structural response parameters in real time through a composite sensor array deployed at key locations of the dam body. The structural response parameters include micromechanical resonant frequency offset and distributed optical fiber strain gradient;

[0142] An interface state evaluation module, which performs a time-frequency joint analysis on the micromechanical resonant frequency offset, constructs an interface peeling characteristic spectrum, and calculates an interface adhesion degradation characteristic value for evaluating the mechanical coupling state of the sensor-dam interface;

[0143] A damage identification module, which performs spatial correlation modeling on distributed optical fiber strain gradients, constructs a strain field anomaly distribution matrix, and calculates structural damage sensitivity eigenvalues ​​to identify potential damage areas within the dam body;

[0144] A safety assessment module, which fuses the interface adhesion degradation characteristic value and the structural damage sensitivity characteristic value into a comprehensive safety characteristic vector, inputs it into a pre-trained dam health assessment model for multi-dimensional analysis, and performs a safety status classification assessment based on the analysis results;

[0145] An early warning decision module, which dynamically adjusts the monitoring and early warning strategy according to the evaluation result: triggers data credibility verification when the comprehensive safety characteristic value exceeds the first threshold value, starts multi-source data fusion verification when the comprehensive safety characteristic value exceeds the second threshold value, and generates a hierarchical early warning signal and links the emergency response mechanism when the comprehensive safety characteristic value exceeds the third threshold value.

[0146] The working principle of the present application is as follows: the present application realizes real-time collection of micro-mechanical resonance frequency shift and distributed optical fiber strain gradient by deploying a composite sensing array, constructs an interface peeling characteristic spectrum and calculates an interface bonding deterioration characteristic value by using time-frequency joint analysis, obtains a strain field anomaly distribution matrix and a structure damage sensitivity characteristic value by combining spatial correlation modeling, fuses a comprehensive safety characteristic vector after generation, inputs a dam health evaluation model for multi-dimensional analysis, and finally dynamically adjusts the monitoring and early warning strategy according to the evaluation result. This method realizes intelligent monitoring of the whole process from data collection, feature extraction, state evaluation to early warning decision, and significantly improves the accuracy and reliability of dam safety monitoring. The present application first deploys a composite sensing array containing micro-mechanical resonance sensors and distributed optical fiber sensors at key positions of the dam body to collect structural response parameters in real time. By performing empirical mode decomposition and Hilbert transform on the micro-mechanical resonance frequency shift, a three-dimensional interface peeling characteristic spectrum is constructed and an interface bonding deterioration characteristic value is calculated to accurately evaluate the mechanical coupling state of the sensor and the dam interface. At the same time, spatial gridding processing and autocorrelation analysis are performed on the distributed optical fiber strain gradient to construct a strain field anomaly distribution matrix and calculate a structure damage sensitivity characteristic value, which effectively identifies potential damage areas inside the dam body. After fusing the two types of characteristic values, they are input into a double-channel feature fusion network based on an attention mechanism and a deep neural network evaluator to realize multi-dimensional accurate evaluation of the dam health state. Finally, through a three-level linked early warning decision tree, data verification, multi-source verification or emergency response are dynamically triggered according to the comprehensive safety characteristic value to form a complete dam safety monitoring and early warning closed loop.

[0147] The above describes one embodiment of the present application in detail, but the content described is only the preferred embodiment of the present application and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made within the scope of the present application should still belong to the patent coverage of the present application.

Claims

1. A water conservancy project dam safety monitoring and early warning method, characterized in that: The following steps are involved: S1: A composite sensor array deployed at key locations on the dam body collects structural response parameters in real time, including micromechanical resonant frequency offset and distributed optical fiber strain gradient. S2: Perform a time-frequency joint analysis of the micromechanical resonant frequency offset, construct an interface peeling characteristic spectrum, and calculate the interface adhesion degradation characteristic value to evaluate the mechanical coupling state of the sensor and dam interface; The calculation of the interface adhesion degradation characteristic value specifically includes: Based on the interface peeling characteristic spectrum, the main frequency band energy distribution of the normal vibration mode is extracted, and its relative entropy value with respect to the reference state is calculated as the first characteristic parameter; then, the degree of distortion of the harmonic component of the tangential vibration mode is analyzed, and its harmonic distortion rate is calculated as the second characteristic parameter; the first and second characteristic parameters are weightedly fused, where the weight coefficient is dynamically adjusted according to the sensor embedment depth and the concrete age; and the fused characteristic parameters are converted into the interface bonding degradation characteristic value through a preset interface state mapping function. The interface bonding degradation characteristic value ranges from 0 to 1, and a larger value indicates a worse interface bonding state. S3: Model the spatial correlation of distributed optical fiber strain gradients, construct a strain field anomaly distribution matrix, and calculate the structural damage sensitivity eigenvalue to identify potential damage areas inside the dam body; The specific method for calculating the structural damage sensitivity characteristic value includes: Based on the strain field anomaly distribution matrix, high anomaly areas exceeding two standard deviations are extracted from the matrix; the area proportion and spatial aggregation of the high anomaly areas are calculated as primary features; the deviation angle between the strain gradient direction of the high anomaly areas and the principal stress direction of the dam body is analyzed as a secondary feature; the primary features and secondary features are input into a pre-trained damage sensitivity assessment network, which adopts a three-layer convolutional neural network structure. The first layer extracts local spatial features, the second layer captures long-range spatial correlations, and the third layer outputs structural damage sensitivity eigenvalues. The structural damage sensitivity eigenvalue range is set to 0 to 100, and the larger the value, the greater the damage risk. S4: The interface adhesion degradation eigenvalue and the structural damage sensitivity eigenvalue are integrated into a comprehensive safety feature vector, which is then input into the pre-trained dam health assessment model for multi-dimensional analysis. The safety status is then graded and assessed based on the analysis results. S5: Dynamically adjust the monitoring and early warning strategy based on the assessment results: When the comprehensive security feature value exceeds the first threshold, data credibility verification is triggered; when it exceeds the second threshold, multi-source data fusion verification is initiated; when it exceeds the third threshold, a graded early warning signal is generated and the emergency response mechanism is linked; The comprehensive security feature value is represented in the form of a multi-dimensional comprehensive security feature vector.

2. A water conservancy project dam safety monitoring and early warning method according to claim 1, characterized in that: The construction of the interface peeling characteristic spectrum specifically includes: An empirical mode decomposition is performed on the micromechanical resonant frequency offset to obtain intrinsic mode function components containing different time scales; a Hilbert transform is performed on each intrinsic mode function component to extract the instantaneous frequency characteristics; the instantaneous frequency of each component is compared with a preset reference frequency to calculate a multi-scale frequency shift eigenvector; the multi-scale frequency shift eigenvector is aligned in time and space with the ambient temperature monitoring data to construct a three-dimensional interface peeling characteristic spectrum, in which the first dimension represents the time scale, the second dimension represents the frequency offset, and the third dimension represents the temperature compensation coefficient.

3. A water conservancy project dam safety monitoring and early warning method according to claim 1, characterized in that: Specific methods for evaluating the mechanical coupling state include: An interface state assessment model based on a deep belief network is established, and the interface adhesion degradation characteristic value is used as the input feature. The hidden layer of the model contains three layers of restricted Boltzmann machines, which are used to extract shallow frequency domain features, mid-level time domain features, and deep nonlinear coupling features, respectively. The output layer of the model uses a softmax function to generate the interface state probability distribution, including four state classifications: intact state, slight peeling, moderate peeling, and severe peeling. When the output probability exceeds the preset threshold, the corresponding state warning signal is triggered, and the abnormal time series characteristics are automatically recorded for subsequent trend analysis.

4. A water conservancy project dam safety monitoring and early warning method according to claim 1, characterized in that: The constructing of the strain field abnormal distribution matrix specifically includes: The distributed optical fiber strain gradient is spatially gridded, and the dam body is divided into several spatial units. The coefficient of variation of the optical fiber strain gradient in each spatial unit is then calculated to form an initial strain distribution matrix. The strain transfer coefficient of adjacent spatial units is then introduced, and the isolated outliers in the initial strain distribution matrix are corrected using a spatial autocorrelation algorithm. Finally, the corrected matrix is ​​weighted in combination with the stress concentration factor of the finite element model of the dam structure to generate a physically meaningful strain field anomaly distribution matrix, in which the matrix element value reflects the degree of strain anomaly in the corresponding spatial unit.

5. A water conservancy project dam safety monitoring and early warning method according to claim 1, characterized in that: Specific methods for identifying areas of potential damage include: A damage warning model based on the strain field evolution trend is established, and the structural damage sensitivity characteristic values ​​of multiple consecutive monitoring cycles are used to form a time series; the trend component and mutation component in the time series are extracted through wavelet transform; when the slope of the trend component exceeds the preset threshold, it is determined to be a progressive damage area; when the amplitude of the mutation component exceeds three times the noise level, it is determined to be a sudden damage area; finally, the spatial coordinates of the two types of damage areas are compared with the dam structure drawings, and the scope of potential damage areas that need special attention is marked.

6. A water conservancy project dam safety monitoring and early warning method according to claim 1, characterized in that: The construction process of the dam health assessment model is as follows: A dual-channel feature fusion network based on the attention mechanism is established, in which the first channel processes the interface adhesion degradation eigenvalues, and the second channel processes the structural damage sensitivity eigenvalues. A three-layer feature extraction module is set up within each channel. The first layer uses a sliding window to extract local temporal features, the second layer captures periodic changes through autocorrelation analysis, and the third layer uses a nonlinear activation function to generate high-order abstract features. The feature vectors output by the dual channels are then subjected to cross-attention calculation to generate a comprehensive safety feature vector with spatial-temporal correlation. Finally, a deep neural network evaluator consisting of five hidden layers is constructed. The first hidden layer standardizes the input features, the second to fourth hidden layers set differentiated weights corresponding to the physical characteristics of different parts of the dam body, and the fifth hidden layer outputs the safety assessment results through state compression.

7. A water conservancy project dam safety monitoring and early warning method according to claim 1, characterized in that: The dynamic adjustment of monitoring and early warning strategies specifically includes: A three-level linkage early warning decision tree is constructed. The first-level node of the decision tree corresponds to the first threshold trigger condition. When the comprehensive safety characteristic value exceeds the first threshold, the system automatically starts the redundant sensor cross-validation process, and confirms the data credibility by comparing the consistency of the monitoring data of at least three different types of sensors; the second-level node of the decision tree corresponds to the second threshold trigger condition. When the characteristic value exceeds the second threshold, the system automatically activates the multi-source data fusion engine. The multi-source data fusion engine aligns the structural response parameters in time and space, and uses a data assimilation algorithm based on physical mechanisms to generate verification conclusions; the third-level node of the decision tree corresponds to the third threshold trigger condition. When the characteristic value exceeds the third threshold, the system automatically matches the preset emergency response plan library, and generates a graded early warning signal according to the specific part of the dam body and the damage type. At the same time, it links the downstream early warning system and emergency management platform through the communication protocol.

8. A water conservancy project dam safety monitoring and early warning system, characterized in that: A method for monitoring and early warning of dam safety in a water conservancy project according to any one of claims 1 to 7, comprising: A data acquisition module, which collects structural response parameters in real time through a composite sensor array deployed at key locations of the dam body. The structural response parameters include micromechanical resonant frequency offset and distributed optical fiber strain gradient; An interface state evaluation module, which performs a time-frequency joint analysis on the micromechanical resonant frequency offset, constructs an interface peeling characteristic spectrum, and calculates an interface adhesion degradation characteristic value for evaluating the mechanical coupling state of the sensor-dam interface; A damage identification module, which performs spatial correlation modeling on distributed optical fiber strain gradients, constructs a strain field anomaly distribution matrix, and calculates structural damage sensitivity eigenvalues ​​to identify potential damage areas within the dam body; A safety assessment module, which fuses the interface adhesion degradation characteristic value and the structural damage sensitivity characteristic value into a comprehensive safety characteristic vector, inputs it into a pre-trained dam health assessment model for multi-dimensional analysis, and performs a safety status classification assessment based on the analysis results; An early warning decision module dynamically adjusts the monitoring and early warning strategy according to the evaluation results: when the comprehensive safety feature value exceeds the first threshold, the data credibility check is triggered; when it exceeds the second threshold, the multi-source data fusion verification is started; when it exceeds the third threshold, a graded early warning signal is generated and the emergency response mechanism is linked.

Images