Safety early warning system based on dam crack identification

Abstract

The invention discloses a safety early warning system based on dam crack identification. The system comprises an intelligent crack identification layer, a multi-source monitoring layer, a fusion analysis layer and an intelligent control layer. The crack intelligent identification layer extracts crack parameters and calculates a crack density index CDI by using adaptive graying, entropy optimization segmentation and a U-Net network through an image acquisition and processing module; the multi-source monitoring layer monitors displacement and water level and calculates displacement offset Sd and a water storage abnormal index Wa; the fusion analysis layer fuses multi-source parameters based on a dynamic stability coefficient DSC model, outputs comprehensive stability indexes and divides early warning grades; and the intelligent control layer executes response measures such as targeted drainage, full-dam drainage or emergency broadcasting according to the early warning level. The system has a parameter self-learning capability, can dynamically optimize model parameters, and realizes early recognition, comprehensive evaluation and accurate prevention and control of dam safety.

Description

Technical Field

[0001] This invention relates to a safety early warning system based on dam crack identification, belonging to the field of safety early warning technology. Background Technology

[0002] As a crucial water conservancy infrastructure, the structural safety of dams directly impacts the safety of life and property in downstream areas and the stable development of the regional economy. Therefore, real-time and accurate monitoring and early warning of dam health, particularly surface cracks, is a core issue in the field of engineering safety. However, current dam safety monitoring technologies and systems have revealed several pressing technical bottlenecks in practice, mainly in the following aspects: First, crack detection suffers from significant delays and subjectivity. Current technology still largely relies on regular manual inspections. This method is not only inefficient and costly, but also makes it difficult to continuously and in real-time detect small cracks on the dam surface, especially those appearing early (e.g., cracks less than 0.2 mm wide). Manual interpretation is easily affected by factors such as the inspector's experience, the site environment, and fatigue levels, resulting in a high rate of subjective misjudgment, estimated at 15% to 30%. This leads to many potential safety hazards going undetected and unaddressed in a timely manner.

[0003] Second, existing early warning systems have limitations due to their single-parameter nature. Most current safety monitoring systems typically set fixed alarm thresholds for single physical parameters such as displacement, seepage flow, and water level. This isolated early warning model ignores the fact that dam structural instability is the result of the coupling and combined effects of multiple factors (such as crack propagation, water level changes, and dam displacement). For example, even if the water level has not exceeded the preset safety threshold, rapid crack propagation may already pose a serious threat to the stability of the dam structure. Single-parameter early warning systems cannot identify this multi-source risk coupling effect, thus creating early warning blind spots and leading to missed warnings.

[0004] Third, safety control measures are passive and lack precision. When the monitoring system issues an early warning, traditional response measures, such as drainage and flood discharge, typically employ a strategy of uniformly discharging water across the entire dam area. This approach cannot precisely reduce pressure at the specific locations of identified cracks, resulting not only in a huge waste of precious water resources but also potentially exacerbating damage to local dam structures due to improper flow distribution, leading to secondary disasters. While existing systems can monitor water storage and decide whether to send early warning information when overall stability anomalies occur, they are clearly insufficient in translating this information into efficient and targeted response actions.

[0005] In summary, there is an urgent need in this field for a dam safety early warning system that can accurately identify cracks in their early stages, integrate multi-source information for comprehensive risk assessment, and drive precise and proactive prevention and control measures, so as to overcome the above-mentioned deficiencies of existing technologies and comprehensively improve the safety operation and maintenance level of dams. Summary of the Invention

[0006] [Technical Issues] This invention aims to solve three core problems in existing dam safety monitoring technologies: First, how to achieve automated and high-precision identification of dam cracks (especially early micro-cracks) to eliminate the lag and subjectivity of manual inspections; second, how to integrate multi-source monitoring data (cracks, displacement, water level) for comprehensive risk assessment to avoid the limitations and blind spots of single-parameter early warning; and third, how to formulate targeted prevention and control measures based on risk assessment results to replace the traditional passive and extensive flood discharge strategy and achieve efficient resource utilization and avoidance of secondary risks.

[0007] [Technical Solution] In a first aspect, the present invention provides a safety early warning system based on dam crack identification, comprising: The crack intelligent recognition layer is used to acquire and process images of the dam surface to identify crack parameters on the dam surface and calculate the crack density index (CDI). A multi-source monitoring layer is used to monitor the displacement and water level of the dam body, and to calculate the displacement offset Sd and the water storage anomaly index Wa based on the monitoring data; The fusion analysis layer is communicatively connected to the crack intelligent identification layer and the multi-source monitoring layer. It is used to receive the crack density index CDI, displacement offset Sd, and water storage anomaly index Wa, and fuse the crack density index CDI, displacement offset Sd, and water storage anomaly index Wa. It calculates a dynamic stability coefficient DSC that comprehensively characterizes the stability of the dam through the dynamic stability coefficient DSC model, and determines the warning level based on the value of the DSC. The intelligent control layer, which is communicatively connected to the fusion analysis layer, is used to execute corresponding response measures based on the warning level.

[0008] Optionally, the dynamic stability coefficient DSC model is:

[0009] Where k is a constant; , , These are the weight coefficients obtained through training based on historical incident cases.

[0010] Optionally, the constant k is 8.5.

[0011] Optionally, the crack intelligent recognition layer includes an image acquisition module and an image processing module; The image acquisition module includes a wind and rain resistant high-definition camera and an infrared thermal imager. The wind and rain resistant high-definition camera has a resolution of no less than 4K and a sampling frequency of 10 frames / second, and is used to capture visible light and temperature anomaly images of the dam surface. The image processing module is configured to: convert the acquired RGB image to Lab color space for adaptive grayscale conversion; determine the optimal threshold based on the maximum information entropy principle to perform entropy optimization segmentation on the image and output a binarized image; and use a convolutional neural network to extract crack parameters from the binarized image and calculate the crack density index CDI.

[0012] Optionally, the adaptive grayscale dynamic calculation of pixel color distance is performed as follows:

[0013] in, The adaptive weighting coefficients are adjusted in real time using an ambient light sensor. The color distance of a pixel; and This refers to the pixel brightness value. and This represents the red-green deviation value of a pixel. and This represents the yellow-blue deviation value of a pixel. The calculation method for the maximum information entropy principle is as follows:

[0014] in, The optimal threshold; Solve for the operator to find the optimal threshold; The candidate grayscale threshold; Let i be the probability density of grayscale value i; Information entropy for the foreground region; Information entropy of the background region; The convolutional neural network is a U-Net network, and the crack parameters include length Lc, average width Wc, and orientation angle θ; the crack density index CDI is calculated as follows:

[0015] in, The length of the dam body; This is the critical width.

[0016] Optionally, the multi-source monitoring layer includes a displacement monitoring module and a water level monitoring module; the displacement monitoring module uses a total station network to track multiple coordinate points on the dam surface and calculate the displacement. The calculation method is as follows:

[0017] in, The Euclidean distance between the coordinate points; To monitor the total number of coordinate points; Index of coordinate points; Here are the real-time coordinates of point i; Let i be the initial coordinates of point i; The water level monitoring module calculates the water storage anomaly index in real time. The calculation formula is as follows:

[0018] in, This represents the rate of change in water level.

[0019] Optionally, the warning level includes a three-level warning mechanism, specifically: when At that time, it was at the normal level; when When the time is right, it is a Level 1 warning, triggering the intelligent control layer to push the location of the crack to the terminal device; when When the warning level is 2, the intelligent control layer is triggered to initiate targeted drainage. when At this time, a Level 3 warning is issued, triggering the intelligent control layer to execute full dam drainage and activate emergency broadcasting.

[0020] Optionally, the targeted drainage control algorithm of the intelligent control layer includes: First, the gate opening priority is calculated based on the location of the crack. The calculation method is as follows:

[0021] in, The distance from the crack to the gate. This refers to the local crack density index. The average crack density index; Then, calculate the dynamic drainage volume. The calculation formula is as follows:

[0022] in, Let the projected area of ​​the crack be... This refers to the reservoir's storage capacity.

[0023] Optionally, the system also includes a parameter self-learning module, used to adaptively update the weight coefficients of the dynamic stability coefficient model monthly based on the actual risk assessment results. The calculation method is as follows:

[0024] in, These are the weighting coefficients before the update. This is the updated weighting coefficient. The actual risk index is assessed by experts and ranges from 0 to 10.

[0025] [Beneficial Effects] The safety early warning system based on dam crack identification provided by this invention effectively overcomes the three major bottlenecks of existing technologies by constructing a closed-loop prevention and control system from accurate perception to intelligent diagnosis and then to targeted control, resulting in significant technological progress. Its beneficial effects are specifically reflected in the following aspects: Firstly, in terms of crack identification, early, accurate, and automated crack recognition has been achieved. The image acquisition module, consisting of a weather-resistant high-definition camera and an infrared thermal imager, provides a high-quality, multimodal image data foundation for subsequent processing. The image processing module employs adaptive grayscale technology, adjusting weight coefficients in real time via an ambient light sensor to effectively eliminate interference from complex outdoor lighting. Subsequent entropy-optimized segmentation accurately separates cracks from backgrounds such as vegetation and water stains. Finally, the efficient U-Net convolutional neural network, a high-performance deep learning model, automatically and accurately extracts the key geometric parameters of the cracks and calculates the crack density index (CDI). This series of technological features works together to completely change the traditional reliance on manual inspection, solving the problem of crack detection lag, enabling real-time detection of early-stage cracks <0.2mm, and significantly reducing the subjective misjudgment rate from 15%-30%.

[0026] Secondly, in terms of risk diagnosis, the Dynamic Stability Coefficient (DSC) model adopted in this invention achieves a comprehensive and forward-looking risk assessment based on multi-source information fusion. This model does not simply superimpose multiple parameters, but rather uses a S-shaped function (Logistic function) to nonlinearly fuse the CDI (Crack Indicator) representing crack state, the displacement offset Sd (Structural Deformation), and the water storage anomaly index Wa (External Load) to calculate a comprehensive dynamic stability coefficient (DSC) representing dam stability. The three-level early warning mechanism established based on this DSC value completely breaks through the limitations of single-parameter early warning in existing technologies. Experiments in the embodiments demonstrate that even if the water level does not exceed the threshold, the coupling effect of crack propagation and displacement anomaly can still cause the DSC value to rise to 0.72, thereby triggering a level-two early warning. This multi-parameter coupling analysis makes the early warning more scientific and comprehensive, enabling warnings to be issued 12-48 hours in advance when cracks have not reached the critical value, achieving a leap from single-point threshold alarms to comprehensive system diagnosis. Through retrospective analysis of historical incident cases, it was verified that in several typical cases where cracks did not reach the critical value but were coupled with abnormal displacement, the system model (DSC model) of this invention could issue effective early warnings 12 to 48 hours before the actual incident occurred, proving its forward-looking risk assessment capability.

[0027] Third, in terms of prevention and control execution, the intelligent control layer of this invention achieves a fundamental shift from passive response to precise proactive control. When the system triggers a level-two or higher warning, the intelligent control layer no longer adopts the traditional uniform flood discharge strategy across the entire dam, but instead activates a targeted drainage control algorithm. This algorithm first calculates the gate opening priority based on the crack location, ensuring that gates closer to the crack and with higher risk are opened first; then, it dynamically calculates the drainage volume to precisely match the drainage intensity with the real-time risk level. This strategy significantly reduces the flood discharge volume and noticeably lowers the water pressure in the crack area, greatly conserving water resources and avoiding secondary damage that may be caused by indiscriminate drainage, successfully solving the problem of passive prevention and control measures.

[0028] Fourth, regarding intelligent system evolution, the parameter self-learning module of this invention enables the system to possess adaptive capabilities for continuous optimization. This mechanism allows the system to accumulate experience and learn autonomously during continuous operation, resulting in a significant reduction in the model's false alarm rate and a substantial improvement in the system's reliability and long-term applicability.

[0029] Fifth, regarding multi-scenario adaptability, the core algorithm layer of this invention does not rely on the absolute physical parameters of a specific dam type, but rather makes decisions based on relative relationships such as CDI and DSC, thus enabling compatibility with different dam types such as earth-rock dams and concrete dams. Simultaneously, the hardware used has the capability to operate stably in harsh environments ranging from -30℃ to 70℃, ensuring the system's broad application potential in most regions and climates.

[0030] In summary, this invention, through a complete system comprised of a four-layer architecture—an intelligent crack identification layer, a multi-source monitoring layer, a fusion analysis layer, and an intelligent control layer—organically integrates image recognition technology, multi-source sensing technology, a nonlinear fusion diagnostic model, intelligent control algorithms, and a self-learning mechanism. This forms a comprehensive dam safety early warning system that integrates early warning, precise diagnosis, accurate handling, self-optimization, and wide adaptability. Its technical effects are synergistic and significant: it not only enables early detection of hidden dangers and comprehensive risk assessment but also drives precise and efficient prevention and control measures, greatly ensuring the structural safety of the dam and the safety of life and property downstream. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 The system flowchart provided for this invention. Detailed Implementation

[0033] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Example 1 This embodiment describes in detail the composition and workflow of a safety early warning system based on dam crack identification according to the present invention.

[0035] I. System Composition The safety early warning system consists of four layers: a crack intelligent identification layer, a multi-source monitoring layer, a fusion analysis layer, and an intelligent control layer, forming a closed-loop prevention and control system.

[0036] 1. Intelligent crack recognition layer, responsible for the automatic identification and quantification of cracks on the dam surface, including: Image acquisition module: Composed of weather-resistant high-definition cameras and infrared thermal imagers deployed on the dam surface. The weather-resistant high-definition cameras have a resolution of at least 4K, a sampling frequency of 10 frames per second, a focal length of 12-120mm, and an IP68 protection rating. They are deployed in groups of 50 meters with staggered coverage to capture high-definition visible light images of the dam surface. The infrared thermal imagers are used to simultaneously capture areas with abnormal temperatures.

[0037] Image processing module: Configured to execute an image processing workflow. First, the acquired RGB image is converted to Lab color space for adaptive grayscale conversion, and the weighting coefficients are adjusted in real time using an ambient light sensor. (sunny ,cloudy day To eliminate lighting interference, adaptive grayscale dynamically calculates the color distance of pixels. The calculation method is as follows:

[0038] The parameters are defined as follows: Pixel color distance, used to quantify the overall color difference between any two pixels (i and j) in the dam surface image. The larger the value, the more significant the difference in color features between the two points. This index helps to distinguish the split marks from background interference (such as vegetation and water stains) in subsequent processing.

[0039] , : Pixel brightness value, representing the L channel values ​​of the i-th and j-th pixels in the CIE Lab color space, with a value range of 0 to 100 (0 represents pure black, 100 represents pure white), reflecting the brightness of the pixel.

[0040] , and , Pixel chromaticity values ​​represent the values ​​of the a and b channels of a pixel in the Lab color space. The a channel represents red-green hue, with a value range of -128 to 127 (positive values ​​lean towards red, negative values ​​lean towards green); the b channel represents yellow-blue hue, with a value range of -128 to 127 (positive values ​​lean towards yellow, negative values ​​lean towards blue).

[0041] : Adaptive weighting coefficient, which is used to dynamically balance the contribution ratio of brightness difference and chromaticity difference in color distance calculation, and its value ranges from 0 to 1. The value is dynamically adjusted based on the ambient light intensity sensed in real time by an ambient light sensor. For example, when the light is strong, the weight of brightness difference is increased; when the light is uniform, the weight of color difference is increased.

[0042] : A sign function whose value is determined by the preset direction and angle of the crack. The decision is made to ensure that the directionality of the color distance calculation is consistent with the enhancement of crack features.

[0043] Furthermore, the system uses a preset crack direction angle. A preliminary assessment of the safety risk level is then made for the identified cracks. Specifically, the direction and angle of the cracks are considered. Corresponding to the principal stress direction of the dam structure, different value ranges characterize the potential threat of cracks to the stability of the dam: (1) When the direction angle Cracks approaching 0° or 180°, roughly parallel to the principal stress direction of the dam body, are classified as low- to medium-risk cracks. These cracks are often caused by temperature contraction or horizontal pressure and typically extend along weak stress surfaces. The system marks them as regular monitoring targets, with relatively lenient warning thresholds.

[0044] (2) When the direction angle Cracks approaching 90°, roughly perpendicular to the direction of the principal stress in the dam body, are classified as high-risk cracks. These cracks directly cut into the main load-bearing or seepage-proof structures of the dam body, are prone to propagation under water seepage, and may lead to structural instability. The system will activate an enhanced monitoring mode for these cracks, lower the warning threshold, and prioritize alarms.

[0045] (3) When the direction angle Cracks located between 30° and 60°, roughly parallel to the stress direction of the dam's slope or arch, are identified as key areas of concern. These cracks are common in structures such as arch dams and may exacerbate stress concentration. The system marks them as key monitoring targets and performs trend analysis by correlating them with historical stress data.

[0046] Sign function The value of is set based on the above-mentioned angle interval division and risk logic to ensure that the color distance calculation in subsequent image processing can be targeted to enhance the crack features of different risk types, so that the monitoring results output by the system are consistent with the actual structural safety requirements of the dam.

[0047] Secondly, based on the principle of maximum information entropy, an optimal threshold T is determined, and entropy-optimized segmentation is performed on the image to output a binarized image that eliminates interference from vegetation, water stains, etc. The calculation method for the maximum information entropy principle is as follows:

[0048] The definitions and calculation logic of each parameter are as follows: The optimal segmentation threshold, calculated using this formula, is used to binarize a grayscale image into a foreground (cracked area) and a background (normal surface of the dam). The optimal threshold operator is defined as finding the value of t that maximizes the "total information entropy" in the formula within the range of gray value t (0-255), and defining this value of t as the optimal threshold T. Candidate grayscale thresholds: all possible grayscale thresholds are traversed, with a value range of 0-255 (covering all pixel grayscale levels of the grayscale image). Each t corresponds to a segmentation scheme that "divides the image into foreground (≤t) and background (>t)". The probability density of gray value i: represents the proportion of pixels with gray value i (i ranges from 0 to 255) in the dam image being processed in the current image to the total number of pixels in the image; Foreground region information entropy: When the candidate threshold is t, pixels with gray values ​​≤ t constitute the "foreground region" (preliminarily determined to be crack-related regions). This summation term represents the information entropy of the foreground region. The larger the entropy value, the more uniform the gray distribution in the foreground region. Background region information entropy: When the candidate threshold is t, pixels with gray values ​​> t constitute the "background region" (preliminarily determined to be the normal surface of the dam or the interference region). This summation term represents the information entropy of the background region. The larger the entropy value, the more uniform the gray distribution in the background region.

[0049] Finally, a U-Net convolutional neural network is used to extract crack features from the binarized image (input image scale is 512×512 pixels), outputting the crack length Lc, average width Wc, and orientation angle θ, and the crack density index CDI is calculated according to the following formula:

[0050] The meanings of each parameter are as follows: This parameter represents the total length of the dam structure obtained from the current dam design drawings. It is used to calculate the absolute length of cracks. Normalization to a relative proportion is used to eliminate differences between dams of different sizes.

[0051] This is the critical width, and its value is not a fixed empirical value, but a dynamic threshold determined comprehensively based on the specific dam design drawings, dam type, design strength of building materials, and industry safety standards. For concrete gravity dams, The allowable crack width can be determined based on the design specifications; for earth-rock dams, it can be set based on the impermeability of the core wall or seepage barrier material. This can be achieved by measuring the average crack width. With the dam body specifically designed By making comparisons, we can more scientifically assess the danger of cracks in the width dimension.

[0052] The CDI index combines the relative length and relative width of cracks; a higher value indicates a greater potential threat to the dam structure in terms of both spatial distribution and opening scale. It also incorporates factors based on strike angle. The initial risk assessment and the CDI index provide core quantitative basis for quantitative and refined evaluation of the severity of cracks.

[0053] 2. Multi-source monitoring layer, responsible for acquiring other key parameters of the dam structure and environment, including: Displacement monitoring module: Using a total station network with points arranged in a triangular grid (point spacing ≤ 20 meters), it tracks more than 100 coordinate points on the dam surface and calculates the displacement according to the following formula. :

[0054] The meanings of each parameter are as follows: n: Total number of monitoring coordinate points, which refers to the total number of coordinate points deployed by the total station network on the dam surface for tracking displacement, covering key areas of the dam body (such as the dam crest, dam waist, and dam foundation). : Coordinate point index, used to distinguish different monitoring coordinate points, with values ​​from 1 to n (e.g., i=1 corresponds to the first coordinate point, i=2 corresponds to the second coordinate point), traversing all coordinate points to calculate the displacement contribution of each point and sum them; The real-time coordinates of coordinate point i refer to the real-time three-dimensional coordinates (usually including X, Y, and Z axis coordinates) of the i-th coordinate point obtained by total station network measurement within the monitoring period (e.g., every hour), reflecting the current spatial position of the point. The initial coordinates of coordinate point i refer to the three-dimensional coordinates of the i-th coordinate point measured and recorded by the total station network when the dam is in a normal and stable state. These coordinates serve as the benchmark value for judging displacement in subsequent monitoring cycles (usually calibrated after the dam is completed or maintained).

[0055] : The Euclidean distance between the current coordinates and the initial coordinates of the i-th point, i.e., the absolute displacement of that point.

[0056] The constant "10" in the denominator of this formula has the dimension of length, measured in millimeters (mm). Its function is to form a dimensionless exponential term together with the numerator (displacement Euclidean distance), serving as a characteristic scale to measure the significant contribution of a single-point displacement. This value is determined based on a comprehensive analysis of extensive historical monitoring data and typical threshold ranges for displacement anomaly early warning in dam structural safety specifications.

[0057] The formula obtains the overall displacement offset by summing the displacement contributions of all monitoring points. Exponential function The design makes the contribution of small displacements gradual, while the contribution of abnormally large displacements is significantly amplified, thus more sensitively capturing structural deformation anomalies of the dam body.

[0058] Water level monitoring module: Utilizing water level sensors (accuracy ±0.1% FS) deployed at the upper, middle, and lower reaches of the reservoir, the module monitors the water level in real time. When the absolute value of the rate of change of water level |dV / dt| exceeds a threshold (preferred in this embodiment), the water level is monitored. When the water storage anomaly index is reached, it is calculated using the following formula. :

[0059] The meanings of each parameter are as follows: The rate of change of water level over time is obtained by differential calculation of water level sensor data.

[0060] : Take the absolute value of the rate of change, used to accumulate abnormal fluctuations in all directions.

[0061] , The start and end times of the integration are usually defined as a monitoring and analysis cycle.

[0062] The formula is calculated in The integral of the absolute value of the rate of change of water level over a period of time. It does not concern itself with whether the water level rises or falls, but only with whether the fluctuations are drastic. Normal, slow water level changes contribute very little, while rapid rises or falls in water level over a short period of time are significantly captured by this integral, thus quantifying anomalies in external hydraulic loads.

[0063] 3. The fusion analysis layer, communicating with the crack intelligent identification layer and the multi-source monitoring layer, is used for comprehensive risk assessment. It receives CDI, Sd, and Wa, and performs nonlinear fusion calculations using the dynamic stability coefficient DSC model. The model is as follows:

[0064] Wherein, k is a constant, and in this embodiment, the constant k is preferably taken as 8.5; , , These are weight coefficients obtained through training based on historical incident cases. Preferably, in this embodiment... , , The values ​​are 0.5, 0.3, and 0.2 respectively.

[0065] The weighting coefficients β1, β2, and β3 were obtained through training based on historical datasets. Specifically, the historical dataset was constructed around the correspondence between "multi-source monitoring parameters and actual hazard results," and its core fields included: the crack density index CDI, displacement offset Sd, and water storage anomaly index Wa as input features, and the expert hazard score as the output label.

[0066] The training process for weight coefficients β1, β2, and β3 includes data preprocessing, model training, and validation. Data preprocessing includes: cleaning outlier data; standardizing input features such as CDI, Sd, and Wa (e.g., Z-Score normalization) to eliminate the influence of dimensions; and dividing the dataset into training and test sets. During model training, the expert risk rating is used as the target, with CDI, Sd, and Wa as features. The cross-entropy loss function is employed, and gradient descent or its variant (Adam algorithm) is used to optimize the loss function, iteratively updating the weight coefficients β1, β2, and β3 until convergence. The constant k in the DSC formula remains fixed during training. Model validation can assess the deviation between predicted and actual scores based on metrics such as mean squared error (MSE), ensuring that the trained weight coefficients conform to business logic.

[0067] The weight coefficients obtained through the above training method, such as β1=0.5, β2=0.3, and β3=0.2 in this embodiment, directly reflect the relative weights of the three factors of cracks, displacement, and water level on the stability of the dam body, so that the DSC value can scientifically characterize the comprehensive risk.

[0068] The dynamic stability coefficient, a constant k in the DSC model, is used to adjust the steepness of the Logistic function curve, and its value directly affects the system's sensitivity to changes in risk. The larger the value of k, the steeper the curve near the threshold, and the more sensitive the system is to changes in parameters; the smaller the value of k, the flatter the curve.

[0069] This embodiment determines the optimal value of k using a hierarchical adaptation verification method. Specifically, after determining a set of weight coefficients β1, β2, and β3 through training with historical data, these weight coefficients are fixed, and a verification set containing multiple historical incident cases is used to test the impact of different k values, such as 5.0, 6.5, 8.5, and 10.0, on the accuracy of the early warning. The evaluation index is the matching rate between the warning level to which the calculated DSC value falls and the actual warning level that the case should be in.

[0070] Experimental verification shows that when k is set to 8.5, the system's early warning level matching rate reaches over 90%, achieving the best overall early warning effect. If k is less than 8.5, the curve is too flat, which may lead to insensitivity to risk accumulation and missed warnings; if k is greater than 8.5, the curve is too steep, which may lead to overreaction to normal parameter fluctuations and false alarms. Furthermore, the value of k=8.5 has good compatibility with the subsequent parameter self-learning module, maintaining the stability and reliability of the early warning level even when the weight coefficient β is fine-tuned.

[0071] In addition, this layer is equipped with a three-level early warning mechanism: specifically: when At that time, it was at the normal level; when At this time, it is a Level 1 warning, triggering the intelligent control layer to push the location of the crack to the terminal device; when At this time, it is a level two warning, triggering the intelligent control layer to initiate targeted drainage; when At this time, it is a Level 3 warning, which triggers the intelligent control layer to execute full dam drainage and activate emergency broadcasting.

[0072] The thresholds (0.3, 0.6, 0.8) of the three-level early warning mechanism are optimized values ​​determined based on statistical analysis of historical incident cases, numerical simulation, and engineering experience. Through retrospective analysis of a large amount of historical monitoring data and corresponding incident records, it was found that when the DSC value is below 0.3, the dam's operation is stable with no abnormal risks; when the DSC value is between 0.3 and 0.6, the system begins to show slight abnormal characteristics, requiring early warning and monitoring of crack locations; when the DSC value is between 0.6 and 0.8, the risk coupling effect becomes apparent, requiring targeted intervention (targeted drainage); when the DSC value exceeds 0.8, the dam enters a high-risk state, requiring immediate and comprehensive emergency measures. Verification through case studies shows that this set of thresholds can effectively balance early warning sensitivity and false alarm rate while ensuring safety, providing clear and operable emergency response guidelines.

[0073] When a level-two warning is triggered and targeted drainage needs to be initiated, the system achieves real-time linkage with the control mechanism through a pre-set hierarchical communication link. Specifically: (1) Decision transmission: The fusion analysis layer transmits control decisions containing information such as target gate number, opening priority and dynamic drainage volume Qd to the central controller (such as programmable logic controller PLC) of the intelligent control layer through a highly reliable industrial Ethernet, such as using Profinet or Modbus-TCP protocol.

[0074] (2) Command issuance and execution: The central controller of the intelligent control layer generates specific control commands such as "gate opening degree" and "opening duration" based on the received decision, and issues them to the local control cabinet of the target flood discharge gate through fieldbus (such as RS485) or wireless communication module suitable for remote dam areas (such as 4G / 5G industrial module).

[0075] (3) Driving action: The gate local control cabinet drives its electric actuator (such as a servo motor) or hydraulic actuator (such as a hydraulic pump) to precisely control the opening and closing of the gate, thereby achieving targeted drainage of the crack area with a flow rate of Qd.

[0076] When the fusion analysis layer determines that a level three warning has been issued, the intelligent control layer simultaneously performs the following three operations, the specific implementation of which is as follows: (1) Dam Drainage Control: The central controller (such as an industrial PLC) of the intelligent control layer sends an "emergency full open" command to the local control units (LCUs) of all floodgates in the reservoir through a pre-set industrial Ethernet or fieldbus network (such as Profinet, Modbus TCP / IP). After receiving the command, each LCU drives the electric actuator or hydraulic system of the gate to open the gate to the maximum safe opening degree to achieve the maximum design flow rate for flood discharge. The system's control logic ensures that all gates are started in a preset sequence or in parallel to avoid flow impact caused by simultaneous opening.

[0077] (2) Emergency Broadcast Activation: The intelligent control layer generates a hard-wired trigger signal through the digital output (DO) module, or sends a broadcast command to an independent IP network broadcast system via network protocols (such as TCP / IP, SNMP). The content management server of the broadcast system has multiple sets of evacuation notification voice files (such as notifications for different hazard levels and different times). The command sent by the intelligent control layer contains a hazard level code, and the broadcast system automatically selects and plays the corresponding pre-stored voice file in a loop to ensure that the information accurately and promptly covers the dam area and the designated downstream area.

[0078] (3) Hazard Data Reporting: The communication module built into the intelligent control layer uploads the encapsulated hazard data packet (containing the DSC value at the trigger time, crack location and parameters, displacement monitoring data, warning level, timestamp, etc.) to the provincial or national unified water conservancy safety monitoring cloud platform in real time through a standard API interface (such as a RESTful API based on HTTP / HTTPS). The data packet is in a standardized JSON or XML format, which facilitates the platform to receive, parse, and trigger higher-level emergency response.

[0079] 4. The intelligent control layer, communicating with the fusion analysis layer, is used to execute precise prevention and control measures. Depending on the warning level, its response measures include: for a Level 1 warning, pushing the location of the crack to the terminal equipment; for a Level 2 warning, initiating targeted drainage; and for a Level 3 warning, executing full dam drainage and activating emergency broadcasting.

[0080] When performing targeted drainage, it employs a specific control algorithm, namely: First, the gate opening priority is calculated based on the location of the crack. The calculation method is as follows:

[0081] The calculation and definition of the key parameters are as follows: The shortest spatial distance from the geometric center of the identified crack to the target floodgate; (Local Crack Density Index): Used to accurately quantify the risk level of areas with concentrated cracks, it is a key basis for determining "prioritizing the opening of adjacent gates." Its calculation is based on all crack parameters within a local area centered on the target crack. As a preferred implementation, this local area can be defined as a circular region centered on the crack center with a radius twice the longest dimension of the crack, or a corresponding rectangular region. Within this region, the crack density index is calculated using parameters such as the total length and average width of the cracks within the region, according to the CDI calculation formula. (Mean Crack Density Index): This index characterizes the average crack risk level across the entire dam surface at the current moment, serving as a benchmark for assessing whether local risks are abnormal. Its calculation scope covers all dam surfaces monitored by the image acquisition module (such as the upstream face, crest, and abutments). By statistically analyzing all crack parameters identified across the entire dam area, the average crack density index for the entire monitoring area is calculated according to the CDI calculation formula.

[0082] This priority formula takes into account "spatial proximity" ( ) and "risk anomaly" ( Two factors are considered. The closer the gate is to the crack, the higher its priority base. At the same time, if the risk index of the local area where the crack is located is significantly higher than the average level of the dam body as a whole (i.e., the ratio > 1), the opening priority of the gate will be further amplified, thereby ensuring rapid and precise intervention in high-risk areas.

[0083] Then, calculate the dynamic drainage volume. The calculation formula is as follows:

[0084] in, The projected area of ​​the crack on the dam surface can be estimated using the crack length Lc and average width Wc identified from the image. This refers to the reservoir's storage capacity.

[0085] Meanwhile, the safety early warning system also has a parameter self-learning module, which is used to adaptively update the weight coefficients of the dynamic stability coefficient DSC model monthly based on the actual risk assessment results. The calculation method is as follows:

[0086] in, These are the weighting coefficients before the update. This is the updated weighting coefficient. The actual risk index is assessed by experts and ranges from 0 to 10.

[0087] The learning rate in this formula is set to a fixed value of 0.05. This value is an empirical value determined after extensive simulation verification using historical cases, and its core purpose is to ensure the stability and convergence of the weight coefficient updates. If the learning rate is too large (e.g., >0.1), a single update may cause drastic jumps in the weight coefficients, leading to instability in the DSC calculation results and confusion in the warning levels; if the learning rate is too small (e.g., <0.01), the weight coefficient updates will be slow, and the system's adaptive efficiency will be too low. Testing has shown that a learning rate of 0.05 achieves the best balance between update speed and system stability.

[0088] The actual risk index in this formula is derived by senior water conservancy engineering experts or expert groups through a comprehensive assessment of risk events that occur during the monthly system maintenance cycle and the effectiveness of their handling. To ensure the objectivity and operability of the assessment, the following multi-dimensional quantitative assessment framework is used as the basis for expert scoring: Table 1. Evaluation Dimensions and Scoring Reference Table

[0089] Experts will combine specific monitoring data, on-site investigation findings, and engineering experience from the aforementioned dimensions to score each dimension. Then, through weighted averaging or comprehensive analysis, they will ultimately provide an actual risk index (which can be an integer or a decimal) between 0 and 10. This index quantitatively represents the actual risk level encountered by the dam during that period, serving as a reliable monitoring signal for the self-learning module.

[0090] In addition, at the end of each month, the system compares the highest DSC predicted value (or average DSC value) recorded by the system during the month with the actual risk index (normalized by dividing by 10) as assessed by experts. If experts believe that the actual risk is higher than the system's prediction (index / 10 > DSC), the system will make a slight upward adjustment according to the formula. The coefficients are adjusted to make the model more sensitive to similar risks in the future; conversely, they are fine-tuned downwards. Through this continuous "supervised learning," the system can continuously align itself with the actual engineering safety logic and improve the accuracy of early warnings.

[0091] Example 2 Combination Figure 1 This embodiment uses a specific work scenario (after continuous heavy rain, the system detects an anomaly in section 3 of the dam) as an example to explain in detail the workflow of the safety early warning system in Embodiment 1: 1. The crack intelligent recognition layer initiates the image acquisition and image processing process: High-definition cameras and infrared thermal imagers deployed on section 3 of the dam captured visible light and thermal imaging sequences of the dam surface. The image processing module then processed the visible light images: first, the acquired RGB images were converted to Lab color space for adaptive grayscale conversion, and weighting coefficients were adjusted based on ambient light sensor data to eliminate rain and fog interference; next, entropy optimization segmentation was performed to generate a binarized image highlighting crack features; finally, a U-Net convolutional neural network accurately extracted crack parameters from the image: length Lc = 8.2m, average width Wc = 2.3mm. Based on this, the system calculated the crack density index (CDI) to be 0.48.

[0092] 2. Synchronous operation of multi-source monitoring layers: The total station network detected a shift in the coordinate points of the dam body, and the displacement shift Sd = 3.7 was calculated; the water level sensor detected a sharp rise in the water level, and the water storage anomaly index Wa = 15.3 was calculated.

[0093] 3. The fusion analysis layer performs fusion analysis and early warning decision-making: The fusion analysis layer receives CDI (value 0.48), Sd (value 3.7), and Wa (value 15.3). These multi-source parameters are input into the Dynamic Stability Coefficient (DSC) model for nonlinear fusion calculation. Substituting these parameters into the formula and using preset coefficients, the DSC value is calculated at this point:

[0094] According to the preset three-level early warning mechanism, the DSC value of 0.72 falls into the (0.6, 0.8) range, and the system immediately triggers a level two early warning.

[0095] 4. The intelligent control layer begins intelligent control and precise execution: The intelligent control layer responds to the level 2 early warning and activates the targeted drainage control algorithm.

[0096] First, calculate the gate opening priority: the system locates the crack near gate No. 3 (distance D=20m) and obtains the local crack density index. =0.48 and average crack density index =0.12, substitute into the formula:

[0097] Gate No. 3 has been given the highest opening priority.

[0098] Then, calculate the dynamic drainage volume and execute it based on reservoir capacity. DSC=0.72, CDI=0.48 and crack projected area Calculate the drainage volume:

[0099] Finally, the system automatically opened Gate 3 to 60% to implement precise drainage until the monitoring showed that the water level dropped by 2.1 meters, thus effectively reducing pressure on the cracked area.

[0100] 5. System self-learning and optimization: At the end of this month's cycle, the parameter self-learning module is activated. Based on the effectiveness of the emergency response, experts assess the actual emergency index as 7 (range 0-10).

[0101] The system substitutes the difference between the actual risk index (7) and the predicted DSC value (0.72) into the learning formula:

[0102]

[0103]

[0104] This allowed for fine-tuning of the model weight coefficients. , and This allows the model to more closely approximate reality in subsequent predictions. This continuous self-optimization process enables the system to accumulate experience, ultimately significantly reducing the model's false alarm rate from 22% to 6%.

[0105] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A safety early warning system based on dam crack identification, characterized in that, include: The crack intelligent recognition layer is used to acquire and process images of the dam surface to identify crack parameters on the dam surface and calculate the crack density index (CDI). A multi-source monitoring layer is used to monitor the displacement and water level of the dam body, and to calculate the displacement offset Sd and the water storage anomaly index Wa based on the monitoring data; The fusion analysis layer is communicatively connected to the crack intelligent identification layer and the multi-source monitoring layer. It is used to receive the crack density index CDI, displacement offset Sd, and water storage anomaly index Wa, and fuse the crack density index CDI, displacement offset Sd, and water storage anomaly index Wa. It calculates a dynamic stability coefficient DSC that comprehensively characterizes the stability of the dam through the dynamic stability coefficient DSC model, and determines the warning level based on the value of the DSC. The intelligent control layer, which is communicatively connected to the fusion analysis layer, is used to execute corresponding response measures based on the warning level.

2. The system as described in claim 1, characterized in that, The dynamic stability coefficient DSC model is as follows: Where k is a constant; , , These are the weight coefficients obtained through training based on historical incident cases.

3. The system as described in claim 2, characterized in that, The constant k is 8.

5.

4. The system as described in claim 1, characterized in that, The crack intelligent recognition layer includes an image acquisition module and an image processing module; The image acquisition module includes a wind and rain resistant high-definition camera and an infrared thermal imager. The wind and rain resistant high-definition camera has a resolution of no less than 4K and a sampling frequency of 10 frames / second, and is used to capture visible light and temperature anomaly images of the dam surface. The image processing module is configured to: convert the acquired RGB image to Lab color space for adaptive grayscale conversion; determine the optimal threshold based on the maximum information entropy principle to perform entropy optimization segmentation on the image and output a binarized image; and use a convolutional neural network to extract crack parameters from the binarized image and calculate the crack density index CDI.

5. The system as described in claim 4, characterized in that, The adaptive grayscale dynamic calculation of pixel color distance is performed as follows: in, The adaptive weighting coefficients are adjusted in real time using an ambient light sensor. The color distance of a pixel; and This refers to the pixel brightness value. and This represents the red-green deviation value of a pixel. and This represents the yellow-blue deviation value of a pixel. The calculation method for the maximum information entropy principle is as follows: in, The optimal threshold; Solve for the operator to find the optimal threshold; The candidate grayscale threshold; Let i be the probability density of grayscale value i; Information entropy for the foreground region; Information entropy of the background region; The convolutional neural network is a U-Net network, and the crack parameters include length Lc, average width Wc, and orientation angle θ; the crack density index CDI is calculated as follows: in, The length of the dam body; This is the critical width.

6. The system as described in claim 1, characterized in that, The multi-source monitoring layer includes a displacement monitoring module and a water level monitoring module; the displacement monitoring module uses a total station network to track multiple coordinate points on the dam surface and calculates the displacement. The calculation method is as follows: in, The Euclidean distance between the coordinate points; To monitor the total number of coordinate points; Index of coordinate points; Here are the real-time coordinates of point i; Let i be the initial coordinates of point i; The water level monitoring module calculates the water storage anomaly index in real time. The calculation formula is as follows: in, This represents the rate of change in water level.

7. The system as described in claim 1, characterized in that, The warning levels include a three-level warning mechanism, specifically: when At that time, it was at the normal level; when When the time is right, it is a Level 1 warning, triggering the intelligent control layer to push the location of the crack to the terminal device; when When the warning level is 2, the intelligent control layer is triggered to initiate targeted drainage. when At this time, a Level 3 warning is issued, triggering the intelligent control layer to execute full dam drainage and activate emergency broadcasting.

8. The system as described in claim 1, characterized in that, The targeted drainage control algorithm of the intelligent control layer includes: First, the gate opening priority is calculated based on the location of the crack. The calculation method is as follows: in, The distance from the crack to the gate. This refers to the local crack density index. The average crack density index; Then, calculate the dynamic drainage volume. The calculation formula is as follows: in, Let the projected area of ​​the crack be... This refers to the reservoir's storage capacity.

9. The system as described in claim 1, characterized in that, The system also includes a parameter self-learning module, which is used to adaptively update the weight coefficients of the dynamic stability coefficient model monthly based on the actual risk assessment results. The calculation method is as follows: in, These are the weighting coefficients before the update. This is the updated weighting coefficient. The actual risk index is assessed by experts and ranges from 0 to 10.

Images