Method for identifying concrete crack of weir body

Abstract

The invention provides a weir body concrete crack identification method, which comprises the following steps of: constructing three scenes of leakage-temperature coupling, crack form acquisition and service safety evaluation, and synchronously acquiring corresponding signals; performing preliminary screening on the leakage-temperature signals by using dynamic texture-edge collaborative filtering to obtain a crack primary judgment graph; analyzing the three-dimensional shape of the fracture to obtain an atlas; inputting the service safety signal into a pre-training network to obtain a safety map; the three maps are sent into a weir body crack cognitive network, a crack recognition map is output, and high-precision comprehensive recognition is achieved. The method can achieve the high-precision and multi-dimensional recognition of the concrete crack of the weir body, improves the accuracy and reliability of crack detection, and provides powerful support for the safety evaluation and maintenance of the weir body structure.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering safety monitoring technology, and more specifically, to a method for identifying cracks in the concrete of a weir. Background Technology

[0002] In water conservancy engineering construction, the stability of the concrete structure of weirs is crucial. The appearance of cracks in the concrete of weirs not only affects their structural integrity but can also lead to problems such as leakage and unstable seepage pressure, thereby affecting the safe operation of the entire water conservancy facility. Traditional crack detection methods mainly rely on manual visual inspection or simple instrument detection, which suffer from drawbacks such as low efficiency, insufficient accuracy, and difficulty in real-time monitoring. In recent years, with the development of sensor technology and signal processing technology, some crack detection methods based on single signals have been proposed, such as using temperature signals or leakage signals for crack identification. However, these methods often provide only limited information and are insufficient to comprehensively assess the impact of cracks on the weir structure.

[0003] In implementing the embodiments of the present invention, the prior art has at least the following problems or defects: Most existing crack detection methods are based on single signal analysis, failing to comprehensively consider multi-dimensional information such as crack leakage characteristics, temperature characteristics, and structural morphology, resulting in insufficient accuracy and reliability of crack identification. Furthermore, traditional crack assessment methods lack real-time monitoring capabilities for dynamic crack changes, making it difficult to promptly detect crack propagation trends and their impact on dam safety. Therefore, a crack identification method capable of integrating multiple signals and morphological information is needed to improve the accuracy and efficiency of crack detection and ensure the safe operation of the dam's concrete structure. Summary of the Invention

[0004] This invention provides a method for identifying cracks in the concrete of a weir, comprising: A leakage-temperature coupled detection scenario, a crack morphology acquisition scenario, and a service safety assessment scenario were constructed. The leakage-temperature coupled detection, crack morphology acquisition, and service safety assessment were performed on the concrete cracks of the dam body, and the raw data of leakage-temperature coupled signals, crack three-dimensional morphology signals, and service safety signals were collected simultaneously. A dynamic texture-edge collaborative filtering algorithm is used to perform primary screening of the leakage-temperature coupling signal to generate a primary crack determination map. Multi-dimensional morphological analysis is performed based on the three-dimensional morphological signal of the fracture to generate a fracture morphological analysis map; wherein, the three-dimensional morphological signal of the fracture includes the fracture opening distance, the fracture trace extension distance, and the fracture depth distance. The service safety signal is input into a pre-trained service safety prediction network to predict the service safety map; The initial crack determination map, crack morphology analysis map, and service safety map are input into the crack recognition network of the weir body to obtain the crack identification map.

[0005] Furthermore, multi-dimensional morphological analysis is performed based on the three-dimensional morphological signal of the fracture to generate a fracture morphological atlas, including: Based on the fracture opening distance, perform primary morphological analysis to generate a primary fracture morphology sub-graph; Secondary morphological corrections are performed on the primary fracture morphology sub-map based on the fracture extension distance and fracture depth distance to generate a fracture morphology analytical map. The primary morphological analysis includes dynamically comparing the fissure opening distance with the allowable opening distance threshold of the weir body design to determine whether the fissure opening is in an over-limit state. The secondary morphological correction includes calculating the ratio of the crack extension distance to the length of the dam structural unit, and calculating the ratio of the crack depth distance to the dam thickness, in order to comprehensively determine the degree to which the crack weakens the integrity of the dam. The multi-dimensional morphological analysis also includes quantifying the surface roughness, surface curvature, and flow channel connectivity of the fracture surface to generate a more comprehensive description of the fracture morphology.

[0006] Furthermore, based on the fracture opening distance, primary morphological analysis is performed to generate a primary fracture morphology sub-map, including: The dam section to be tested is divided into several continuous evaluation units along the axis of the dam body. Within each evaluation unit, the fracture opening distance sequence is statistically analyzed, and the average opening distance value and opening distance fluctuation amplitude of the sequence are calculated. The first opening distance safety index is generated by comparing the average opening distance value with the upper limit of the allowable opening distance, and the second opening distance safety index is generated by comparing the opening distance fluctuation amplitude with the fluctuation tolerance range. A primary fracture morphology sub-map is generated based on the first and second opening distance safety indices.

[0007] Furthermore, secondary morphological corrections are performed on the primary fracture morphology sub-map based on the fracture extension distance and fracture depth distance to generate an analytical fracture morphology map, including: Calculate the deviation ratio between the crack extension distance and the length of the dam structural unit, and calculate the penetration ratio between the crack depth distance and the dam thickness. If the deviation ratio is lower than the preset extension warning ratio and the penetration ratio is lower than the preset penetration warning ratio, then the primary crack morphology sub-graph is convolved with the first correction weight matrix to obtain the crack morphology analytical map. If the deviation ratio is higher than the preset extension warning ratio and the penetration ratio is higher than the preset penetration warning ratio, then the primary crack morphology sub-graph is convolved with the second correction weight matrix to obtain the crack morphology analytical map. The first and second corrected weight matrices are obtained through offline training using historical dam failure samples, and the difference in the numerical distribution of the elements of the two matrices is greater than 50% to ensure that the correction directions are opposite.

[0008] Furthermore, the generation of the primary fracture morphology sub-map based on the first and second opening distance safety indices includes: The first and second opening distance safety indices are respectively subjected to nonlinear normalization to obtain the normalized vector of the two indices. Input the two index normalized vectors into the dual-channel convolution kernel, perform spatial-frequency joint convolution, and output a crack opening risk distribution array; The risk distribution array of fracture openings is pseudo-color mapped to generate a primary fracture morphology sub-map. The kernel size of the dual-channel convolution kernel is dynamically determined by the maximum particle size of the concrete aggregate in the dam body, and the convolution step size adapts to the length of the evaluation unit.

[0009] Furthermore, a dynamic texture-edge collaborative filtering algorithm is used to perform initial screening on the leakage-temperature coupling signal to generate a preliminary crack determination map, including: A joint feature set of texture and edge features is constructed for the leakage-temperature coupled signal; the joint feature set includes thermal image texture entropy, leakage edge gradient, and temperature change ridge; each feature is associated with several background interference suppression operators; The significance response of a feature is calculated based on the weighted response of all background interference suppression operators corresponding to that feature. An adaptive weighted fusion is performed on the saliency responses of all features to obtain a primary crack determination map; The weight vector of the adaptive weighted fusion is updated through online Bayesian inference, and the update rate is adjusted in real time by the rate of change of the water level upstream of the weir.

[0010] Furthermore, the dam crack recognition network includes: The input layer receives the initial crack judgment map, crack morphology analysis map, and service safety map. After unifying the execution scale and grayscale of the three maps, they are sent to the nonlinear fusion layer. The nonlinear fusion layer is used to perform element-level multi-sensory fusion of the unified primary decision tensor, morphological analysis tensor, and security graph tensor to output the crack cognition tensor. The output layer is used to perform channel attention weighting on the crack recognition tensor, and outputs a crack recognition map after being mapped by an activation function. The element-level multisensor fusion is accomplished by three parallel branches, each containing dilated convolutions with different dilation rates to capture the short-, medium-, and long-range spatial correlations of the gaps.

[0011] Furthermore, the element-level multisensory fusion outputs a crack cognition tensor according to the following logic: The primary decision tensor, morphological analysis tensor, and security graph tensor are respectively input into three dilated convolution branches with coprime dilation rates to obtain three branch outputs. Perform element-wise multiplication on the outputs of the three branches and then input the result into a sigmoid gate. Then perform element-wise summation on the outputs of the three branches and then input the result into a tanh gate. The sigmoid-gated result and the tanh-gated result are element-wise multiplied to generate the crack cognition tensor. The activation function is a leaky ReLU, and the leakage slope is dynamically adjusted according to the service life of the weir.

[0012] Furthermore, the service safety signals include seepage pressure stability signals, structural strain hazard signals, and crack propagation trend signals; the service safety prediction network outputs three types of signal maps based on the following heterogeneous logic: The seepage pressure stability map is generated by performing a convolution operation between the fracture surface water pressure attenuation convolution kernel and the real-time seepage pressure field, and then inputting the convolution result into the hyperbolic tangent activation function; wherein, the size of the convolution kernel is adjusted synchronously with the thickness of the weir anti-seepage curtain. The structural strain hazard map is generated by cross-correlation calculation between crack trajectory and full-field strain, and then inputting the cross-correlation result into the ReLU activation function; the cross-correlation kernel direction is aligned with the principal stress direction of the dam body in real time. The crack propagation trend map is generated by performing element-wise multiplication between the output of the Laplacian operator of the upstream velocity field and the crack front curvature-driven convolution kernel, and then inputting the product result into the Sigmoid activation function. Finally, the pressure stability map, structural strain hazard map, and crack propagation trend map are spliced ​​along the channel to form a service safety map.

[0013] Furthermore, the assessment of the seepage stability map also includes performing periodic calibration on the convolution kernel of the water pressure decay at the fracture surface, with the calibration period determined by the rate of change of the siltation elevation upstream of the weir. The assessment of the structural strain hazard map also includes directional retraining of the cross-correlation kernel between crack trajectory and full-field strain. The retraining is triggered when the weir encounters a water level exceeding the design check flood. The evaluation of the crack propagation trend map also includes performing scalar annealing on the curvature-driven convolution kernel at the crack front, with the annealing step size dynamically determined by the ratio of crack propagation rate to the elastic modulus of the dam concrete.

[0014] The embodiments of the present invention have at least the following beneficial effects: 1. This invention, by constructing a leakage-temperature coupled detection scenario, a crack morphology acquisition scenario, and a service safety assessment scenario, can comprehensively collect relevant information on concrete cracks in dams. This includes not only the leakage and temperature characteristics of the cracks but also their three-dimensional morphology and service safety status. This multi-dimensional information acquisition method effectively overcomes the problem of traditional crack detection methods relying on only a single signal, resulting in incomplete information. It provides richer and more comprehensive data support for the accurate identification and assessment of cracks, thereby significantly improving the accuracy and reliability of crack detection. It can more timely and accurately detect the existence of cracks and their potential impact on the dam structure, providing a stronger guarantee for the safe operation of the dam.

[0015] 2. This invention employs a dynamic texture-edge collaborative filtering algorithm to perform initial screening of the leakage-temperature coupling signal, generating a preliminary crack determination map. It also utilizes multi-dimensional morphological analysis based on the three-dimensional crack morphology signal to generate a crack morphology analysis map. These techniques enable refined processing and analysis of the crack signal. The dynamic texture-edge collaborative filtering algorithm effectively suppresses background interference, highlights crack features, and improves the saliency of the crack signal. Multi-dimensional morphological analysis accurately describes the geometric features of the crack and the degree of weakening of the dam's integrity. The combination of these technical features allows this invention to accurately identify the location, size, and morphology of cracks in complex engineering environments. This solves the problems of insufficient crack identification accuracy and susceptibility to interference in existing technologies, providing a more accurate basis for subsequent crack assessment and treatment. This helps to take effective maintenance measures in advance and prevent further crack expansion from causing more serious damage to the dam.

[0016] 3. The service safety signal is input into a pre-trained service safety prediction network to predict the service safety map. The initial crack judgment map, crack morphology analysis map, and service safety map are then input into the weir crack recognition network to obtain the crack identification map. This technical solution fully considers the dynamic changes and safety status of concrete cracks in the weir during actual service. Through intelligent analysis and prediction of service safety signals using a pre-trained network model, it can assess the impact of cracks on the service safety of the weir in real time and accurately. Simultaneously, the introduction of the weir crack recognition network enables deep fusion and comprehensive understanding of multi-source information, further improving the accuracy and efficiency of crack identification. Compared with existing technologies, this invention can more comprehensively and accurately assess the degree of danger of cracks, providing a more effective technical means for real-time monitoring and safety early warning of weirs. This helps to promptly identify potential safety hazards, reduce the risk of safety accidents caused by cracks in the weir, and ensure the safe and stable operation of water conservancy facilities. Attached Figure Description

[0017] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein: Figure 1 This is a flowchart illustrating a method for identifying concrete cracks in a weir according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following drawings indicate similar items; therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0019] Traditional crack detection methods rely on a single signal source, making it impossible to simultaneously capture the crack's leakage characteristics, temperature distribution, and three-dimensional morphological features. For example, in monitoring the concrete structure of a large-scale hydraulic engineering dam, simply acquiring thermal image texture entropy through temperature sensors or collecting leakage edge gradients through piezometers makes it difficult to construct a data chain linking crack opening distance, crack extension distance, and depth distance. Specifically, temperature signals cannot quantify the impact of crack surface roughness on the connectivity of seepage channels, while leakage signals cannot resolve the penetration ratio between crack depth distance and dam thickness. The resulting primary judgment map lacks a multi-dimensional morphological correction mechanism, leading to discrepancies between crack propagation trend prediction and structural integrity assessment.

[0020] For example, in a gravity dam monitoring system employing traditional methods, the temperature sensor array and the piezometer network are independently deployed along the dam's axial direction, and the acquired thermal image texture entropy and seepage edge gradient signals are processed using independent filtering algorithms. Due to the lack of a dynamic texture-edge collaborative filtering mechanism, the spatial coupling characteristics of temperature abrupt change ridges and seepage signals are excessively weakened by background noise interference suppression operators. Simultaneously, service safety assessments rely solely on single structural strain signals, failing to integrate seepage pressure stability, crack propagation trends, and cross-correlation calculations of the entire field strain. This results in the crack front curvature-driven convolution kernel failing to dynamically correlate with the upstream velocity field Laplace operator.

[0021] If the above problems are not addressed, the disconnect between the initial crack assessment map and the service safety map will render the dynamic comparison of the deviation ratio between the length of the dam structural unit and the crack extension ineffective. The penetration ratio of the crack depth cannot be mapped in real time to the synchronous adjustment mechanism of the seepage barrier thickness, and the lack of periodic calibration of the crack surface water pressure attenuation convolution kernel will lead to the accumulation of assessment errors in the seepage pressure stability map. Ultimately, the alignment deviation between the crack propagation trend and the principal stress direction will exacerbate the degradation of the cross-correlation kernel of the structural strain hazard map, causing a dynamic imbalance risk in the ratio of the elastic modulus of the dam concrete.

[0022] To address the aforementioned challenges, this application first considers how to simultaneously integrate the leakage, temperature, and three-dimensional morphological features of cracks to overcome the limitations of single-signal-source analysis. Traditional methods process temperature and leakage signals independently, weakening the coupling characteristics. This application recognizes the need to establish a multi-scenario collaborative detection mechanism, acquiring leakage-temperature coupling signals and crack three-dimensional morphological signals in parallel, and introducing service safety signals to form a multi-dimensional data chain. Simultaneously, addressing the lack of morphological correction in the initial judgment map, this application proposes a dynamic texture-edge collaborative filtering algorithm to achieve joint screening of multiple features, avoiding excessive interference from background noise. Furthermore, traditional service safety assessments rely on a single strain signal. This application designs a system that combines pressure stability, crack propagation trends, and full-field strain cross-correlation calculations, generating a dynamic safety map through a prediction network, and ultimately constructing a cognitive network that integrates multi-source data to achieve comprehensive identification.

[0023] like Figure 1 As shown, this application proposes a method for identifying concrete cracks in a dam, comprising the following steps: S1. Construct a leakage-temperature coupled detection scenario, a crack morphology acquisition scenario, and a service safety assessment scenario. Perform leakage-temperature coupled detection, crack morphology acquisition, and service safety assessment on the concrete cracks of the dam body, and simultaneously acquire the raw data of leakage-temperature coupled signals, crack three-dimensional morphology signals, and service safety signals. S2. The dynamic texture-edge collaborative filtering algorithm is used to perform primary screening of the leakage-temperature coupling signal to generate a primary crack determination map; S3. Perform multi-dimensional morphological analysis based on the three-dimensional morphological signal of the fracture to generate a fracture morphological analysis map; wherein, the three-dimensional morphological signal of the fracture includes the fracture opening distance, the fracture trace extension distance, and the fracture depth distance. S4. Input the service safety signal into the pre-trained service safety prediction network to predict the service safety map; S5. Input the initial crack judgment map, crack morphology analysis map, and service safety map into the weir crack recognition network to obtain the crack identification map.

[0024] The leakage-temperature coupled detection scenario refers to an environment setting that simultaneously collects seepage and temperature distribution data in the crack area of ​​the dam. This can be achieved using a distributed fiber optic sensor and thermal imager deployment to capture seepage anomalies and temperature field distortions caused by cracks. The dynamic texture-edge collaborative filtering algorithm is a signal processing technique that combines thermal imaging texture features with seepage edge gradient features. This can be implemented using texture entropy calculation and edge gradient operator joint filtering to suppress environmental noise and enhance the significance of crack-related signals. The three-dimensional morphological signal of the crack includes the crack opening distance, crack extension distance, and crack depth distance. This can be acquired using three-dimensional laser scanning or structured light imaging technology to quantify the geometric characteristics of the crack in three spatial dimensions. Multi-dimensional morphological analysis refers to multi-angle quantitative analysis of the crack's geometric parameters. This can be achieved using a combination of statistical analysis and proportional calculation to assess the impact of cracks on structural integrity. The service safety prediction network is a deep learning model trained on historical data. This can be implemented using a fusion architecture of convolutional neural networks and time series models to predict the potential threat of cracks to the long-term safety of the dam. Among them, the dam crack recognition network refers to a feature extraction and decision-making model that integrates multi-source data. Specifically, it can be implemented by combining multi-channel dilated convolution with attention mechanism to generate crack recognition results by integrating data from different scenarios.

[0025] The core innovation of this application lies in constructing three collaborative scenarios: leakage-temperature coupling detection, crack morphology acquisition, and service safety assessment. By simultaneously acquiring multi-dimensional signals and employing dynamic texture-edge collaborative filtering, multi-dimensional morphology analysis, and deep learning network fusion technology, it achieves joint analysis of multi-source data on seepage characteristics, morphological features, and structural safety impacts during crack identification, overcoming the limitations of single signal detection.

[0026] The working process and principle of this application are as follows: First, three detection scenarios are constructed: a leakage-temperature coupling detection scenario, a crack morphology acquisition scenario, and a service safety assessment scenario. Corresponding detection and assessment are performed in these scenarios, simultaneously acquiring raw data of leakage-temperature coupling signals, crack morphology signals, and service safety signals.

[0027] For leakage-temperature coupled signals, a dynamic texture-edge collaborative filtering algorithm is used for initial screening to generate a preliminary crack determination map. This algorithm can simultaneously consider the texture and edge features of the signal, improving the accuracy of screening.

[0028] The three-dimensional morphological signal of the fracture includes information in three dimensions: fracture opening distance, fracture trajectory extension distance, and fracture depth distance. Based on this information, multi-dimensional morphological analysis is performed to generate a fracture morphological atlas. Multi-dimensional analysis can comprehensively reflect the spatial characteristics of the fracture.

[0029] Service safety signals are input into a pre-trained service safety prediction network, which then generates a service safety map. The pre-trained network can learn key features of service safety based on historical data.

[0030] Finally, the initial crack identification map, crack morphology analysis map, and service safety map are input into the weir crack recognition network to obtain the final crack identification map. The recognition network achieves comprehensive crack identification by fusing multi-source information.

[0031] The steps are closely interconnected, forming a complete information processing chain. Leakage-temperature coupled detection provides a preliminary judgment, crack morphology acquisition provides spatial characteristics, and service safety assessment provides potential risk prediction. The three are combined and ultimately integrated and analyzed by a cognitive network, thereby achieving a comprehensive and accurate identification of concrete cracks in the dam body.

[0032] As a preferred embodiment, the solution of this application is specifically implemented as follows: Three detection systems were deployed on the concrete weir of a large reservoir dam: The leakage-temperature coupled detection system consists of an infrared thermal imager and a piezometer, which are uniformly arranged along the axis of the weir. The thermal imager collects the surface temperature distribution, and the piezometer measures the changes in internal seepage pressure.

[0033] The crack morphology acquisition system uses a 3D laser scanner to perform high-precision scanning of the dam surface. The scanning resolution is set to 0.1 mm to capture minute cracks.

[0034] The service safety assessment system includes sensors such as strain gauges and displacement gauges, which are deployed at key parts of the dam.

[0035] The three systems collect data synchronously, with a sampling frequency of once per hour. The collected raw data is preprocessed before entering the subsequent analysis process.

[0036] For the leakage-temperature coupled signal, the thermal image texture entropy and leakage edge gradient features are first extracted. Then, a dynamic texture-edge collaborative filtering algorithm is applied. This algorithm employs an adaptive weight fusion mechanism, dynamically adjusting the filtering parameters according to changes in the upstream water level. After filtering, a preliminary crack determination map is generated.

[0037] The processing of the three-dimensional morphology signal of the fracture includes: calculating the statistical characteristics of the fracture opening distance sequence; analyzing the relationship between the fracture trace extension distance and the length of the dam structural unit; and evaluating the ratio of the fracture depth distance to the dam thickness. Based on these analytical results, a multi-dimensional analytical map of the fracture morphology is generated.

[0038] A neural network model pre-trained based on service safety signal input. This model consists of multiple convolutional networks and fully connected layers, capable of learning the spatiotemporal characteristics of signals such as strain and displacement. The network output is a probabilistic graph representing the service safety status.

[0039] Finally, the three types of maps mentioned above are input into the dam crack recognition network. This network employs a multi-scale convolutional structure, enabling it to capture crack features at different scales. The last layer of the network uses the softmax activation function to output the final crack recognition result.

[0040] The entire identification process runs in real time on the edge computing server deployed on-site, and the identification results are presented to the monitoring personnel through a visual interface.

[0041] This application further proposes a method for generating a fracture morphology analysis map based on three-dimensional fracture morphology signals through multi-dimensional morphological analysis. This method includes: generating a primary fracture morphology sub-map based on the fracture opening distance; and generating a fracture morphology analysis map by performing secondary morphological correction on the primary fracture morphology sub-map based on the fracture trace extension distance and fracture depth distance. The primary morphological analysis includes dynamically comparing the fracture opening distance with the allowable opening distance threshold of the weir design to determine whether the fracture opening is in an excessive state. The secondary morphological correction includes proportional calculation of the fracture trace extension distance and the length of the weir structural unit, as well as proportional calculation of the fracture depth distance and the weir thickness, to comprehensively determine the degree to which the fracture weakens the integrity of the weir. The multi-dimensional morphological analysis also includes quantifying the fracture surface roughness, fracture surface curvature, and fracture surface seepage channel connectivity to generate a more comprehensive fracture morphology description.

[0042] The primary morphological analysis generates a first safety index by dynamically comparing the fracture opening distance with the allowable opening distance threshold, and a second safety index by comparing the opening distance fluctuation amplitude with the fluctuation tolerance range. Secondary morphological correction calculates the deviation ratio between the fracture extension distance and the structural unit length, and the penetration ratio between the fracture depth distance and the dam thickness, and performs convolution correction using a first or second correction weight matrix trained on historical dam failure samples. Fracture surface roughness is quantified using the root mean square value of the surface profile, curvature is calculated using the reciprocal of the radius of curvature, and seepage channel connectivity is based on seepage path topology analysis.

[0043] Specifically, in the initial morphological analysis stage, when the fracture opening distance exceeds the allowable opening distance threshold, an over-limit state judgment is triggered, generating the first opening distance safety index. When the opening distance fluctuation amplitude exceeds the fluctuation tolerance range, a second opening distance safety index is generated. Together, they constitute the initial fracture morphology sub-map. In the secondary morphological correction stage, the deviation ratio and penetration ratio reflect the degree of fracture expansion in the planar and longitudinal directions, respectively. When both are below the warning ratio, the first correction weight matrix is ​​used to enhance the characteristics of the safe area; when they are above the warning ratio, the second correction weight matrix is ​​used to strengthen the characteristics of the risk area. The fracture surface roughness is calculated by obtaining surface contour data through laser scanning, the curvature is obtained by fitting the curvature of a three-dimensional point cloud, and the connectivity of the seepage channel is evaluated based on the number of seepage path branches and the intersection density. By superimposing the above quantitative parameters, a multi-dimensional morphological analysis map containing spatial distribution, geometric features, and seepage characteristics is formed.

[0044] As a preferred embodiment, the solution of this application is specifically implemented as follows: The process of generating a fracture morphology atlas by performing multi-dimensional morphological analysis based on the three-dimensional morphological signal of the fracture includes the following steps: First, a primary morphological analysis is performed based on the fracture opening distance to generate a primary fracture morphology sub-graph. Specifically, the fracture opening distance is dynamically compared with the allowable opening distance threshold of the dam design to determine whether the fracture is in an excessive state. For example, the allowable opening distance threshold can be set to 0.2 mm; when the detected fracture opening distance exceeds this threshold, it is determined to be in an excessive state.

[0045] Furthermore, secondary morphological corrections are performed on the primary fracture morphology sub-map based on the fracture extension distance and fracture depth distance to generate an analytical fracture morphology map. Specifically, the ratio of the fracture extension distance to the length of the dam structural unit and the ratio of the fracture depth distance to the dam thickness are calculated to comprehensively determine the degree to which the fracture weakens the integrity of the dam. For example, when the fracture extension distance exceeds 50% of the length of the dam structural unit or the fracture depth distance exceeds 30% of the dam thickness, it can be determined that the fracture has significantly weakened the integrity of the dam.

[0046] Furthermore, multi-dimensional morphological analysis also includes quantifying the surface roughness, curvature, and connectivity of seepage channels on the fracture surface to generate a more comprehensive description of the fracture morphology. Specifically, laser scanning technology can be used to measure the surface roughness, curvature analysis algorithms can be used to calculate the surface curvature, and water injection tests can be used to evaluate the connectivity of seepage channels on the fracture surface.

[0047] This application further proposes dividing the dam section under inspection into several continuous evaluation units along the axis of the weir body; within each evaluation unit, statistically analyzing the crack opening distance sequence and calculating the average opening distance value and the opening distance fluctuation amplitude of the sequence; comparing the average opening distance value with the upper limit of the allowable opening distance to generate a first opening distance safety index, and comparing the opening distance fluctuation amplitude with the fluctuation tolerance range to generate a second opening distance safety index; and generating a primary crack morphology sub-map based on the first and second opening distance safety indices.

[0048] The evaluation units are divided by dynamically adjusting the unit length, which is determined based on the aggregate size distribution of the dam's concrete. The average opening distance is calculated using the sliding window method, with the window size proportional to the evaluation unit length. The opening distance fluctuation amplitude is obtained by calculating the sequence standard deviation, with the standard deviation threshold set based on historical crack propagation data. The first and second opening distance safety indices are processed using nonlinear normalization, and the normalization parameters are dynamically updated with the dam's service life.

[0049] In some embodiments, after the dam section is divided into multiple continuous evaluation units, the crack opening distance data within each unit is continuously collected and formed into a time series. The average opening distance value is calculated using a moving average algorithm, with the window size set to one-fifth to one-third of the evaluation unit length. The opening distance fluctuation amplitude is calculated using a standard deviation algorithm, with the fluctuation tolerance range set to 10% to 20% of the average opening distance value based on the fatigue characteristics of concrete materials. The first opening distance safety index is generated by comparing the average opening distance value with the design allowable upper limit; an early warning is triggered when the ratio exceeds 0.8. The second opening distance safety index is generated by the difference between the standard deviation and the fluctuation tolerance range; an anomaly is marked when the difference exceeds 0.1. After normalization, the two safety indices are input into a dual-channel convolution kernel. The convolution kernel size is set to three to five times the maximum aggregate size, and the convolution step size is adaptively adjusted to half the evaluation unit length. The final generated primary crack morphology sub-map visualizes the safety indices using pseudo-color mapping technology, with a color gradient set to three levels: red, yellow, and green. Red represents high-risk areas, and green represents safe areas.

[0050] As a preferred embodiment, the solution of this application is specifically implemented as follows: The dam section to be tested is divided into several continuous evaluation units along the axis of the dam body. For example, for a 500-meter-long concrete gravity dam, an evaluation unit can be divided into 10 continuous evaluation units, with each unit being 50 meters long.

[0051] Within each evaluation unit, a sequence of fracture opening distances is statistically analyzed, and the average opening distance and the amplitude of the opening distance fluctuation are calculated. Specifically, a high-precision laser rangefinder can be used to measure the fracture opening distance every 1 meter along the fracture direction, obtaining a set of opening distance data. For a 50-meter-long evaluation unit, approximately 50 opening distance data points can be obtained. The arithmetic mean of these 50 data points is then calculated as the average opening distance, and the difference between the maximum and minimum opening distance values ​​is calculated as the opening distance fluctuation amplitude.

[0052] The first opening distance safety index is generated by comparing the average opening distance value with the upper limit of the allowable opening distance, and the second opening distance safety index is generated by comparing the opening distance fluctuation amplitude with the fluctuation tolerance range. For example, assuming the upper limit of the allowable opening distance is 2 mm and the fluctuation tolerance range is ±0.5 mm, if the average opening distance value of a certain evaluation unit is 1.8 mm, then the first opening distance safety index can be set to 0.9 (1.8 / 2); if the opening distance fluctuation amplitude is 0.4 mm, then the second opening distance safety index can be set to 0.8 (0.4 / 0.5).

[0053] A primary fracture morphology sub-map is generated based on the first and second opening distance safety indices. These two safety indices can be used as two-dimensional coordinates to create a scatter plot, with each evaluation unit corresponding to one scatter point. Then, an interpolation algorithm is used to generate a continuous color distribution map, forming the primary fracture morphology sub-map.

[0054] This application further proposes a secondary morphological correction method based on the crack extension distance and crack depth distance to generate a crack morphology analysis map. This includes: calculating the deviation ratio between the crack extension distance and the length of the dam structural unit, and calculating the penetration ratio between the crack depth distance and the dam thickness; if the deviation ratio is lower than a preset extension warning ratio and the penetration ratio is lower than a preset penetration warning ratio, then the primary crack morphology sub-map is convolved with a first correction weight matrix to obtain a crack morphology analysis map; if the deviation ratio is higher than a preset extension warning ratio and the penetration ratio is higher than a preset penetration warning ratio, then the primary crack morphology sub-map is convolved with a second correction weight matrix to obtain a crack morphology analysis map; the first correction weight matrix and the second correction weight matrix are obtained through offline training using historical dam failure samples, and the difference in the numerical distribution of the elements of the two matrices is greater than 50% to ensure that the correction directions are opposite.

[0055] The deviation ratio is calculated as the ratio of crack extension distance to structural unit length, and the penetration ratio is calculated as the ratio of crack depth distance to dam thickness. The extension warning ratio is set at 15% of the structural unit length, and the penetration warning ratio is set at 30% of the dam thickness. The elements of the first correction weight matrix are concentrated in the positive range, and the elements of the second correction weight matrix are concentrated in the negative range. Both matrices are optimized from historical samples using a backpropagation algorithm, with the training objective being to minimize the mean square error between the corrected map and the actual damaged area.

[0056] In some embodiments, when both the deviation ratio and the penetration ratio are below the warning threshold, a first correction weight matrix is ​​used to positively enhance the primary crack morphology sub-map, reducing the probability of misjudgment in low-risk areas. When both exceed the warning threshold, a second correction weight matrix is ​​used for negative suppression, strengthening the abnormal signal in high-risk areas. The difference in matrix element values ​​is enforced through offline training to ensure that the two matrices produce opposite correction directions during convolution. For example, in a dam with a structural unit length of ten meters, the second matrix correction is triggered when the crack extension distance reaches 1.5 meters; if the depth distance exceeds three meters, the penetration ratio threshold is triggered simultaneously. Through the dual-matrix dynamic switching mechanism, the accuracy of the crack morphology analysis map is improved, effectively distinguishing the different impacts of local surface cracking and deep penetrating cracks on structural integrity.

[0057] As a preferred embodiment, the solution of this application is specifically implemented as follows: Calculate the deviation ratio between the crack extension distance and the length of the dam structural unit, and calculate the penetration ratio between the crack depth distance and the dam thickness. For example, if the length of a dam structural unit is 10 meters and the measured crack extension distance is 3 meters, the deviation ratio is 30%. If the dam thickness is 5 meters and the measured crack depth distance is 2 meters, the penetration ratio is 40%.

[0058] Furthermore, the preset extension warning ratio is 50%, and the preset penetration warning ratio is 60%. Therefore, if the deviation ratio of 30% is lower than the preset extension warning ratio of 50% and the penetration ratio of 40% is lower than the preset penetration warning ratio of 60%, the primary fracture morphology sub-graph is convolved with the first correction weight matrix to obtain the fracture morphology analytical map.

[0059] In some embodiments, the first corrected weight matrix can be set as a 3x3 matrix with element values ​​distributed between 0.8 and 1.2. The primary fracture morphology sub-map is convolved with this matrix to obtain the corrected fracture morphology analytical map.

[0060] In another scenario, if the deviation ratio is higher than the preset extension warning ratio and the penetration ratio is higher than the preset penetration warning ratio, the primary fracture morphology sub-graph is convolved with the second correction weight matrix to obtain the fracture morphology analytical map. The second correction weight matrix can also be set as a 3x3 matrix, but its element values ​​are distributed between 1.5 and 2.5.

[0061] The first and second corrected weight matrices are obtained through offline training using historical dam failure samples. For example, 100 sets of historical dam failure sample data can be collected, and machine learning algorithms can be used to train on these samples to obtain two different corrected weight matrices. The numerical distribution of the elements in the two matrices differs by more than 50% to ensure that the correction directions are opposite. This design ensures differentiated treatment for different degrees of crack conditions.

[0062] This application further proposes to perform nonlinear normalization on the first and second opening distance safety indices to obtain a normalized vector of the two indices; input the normalized vector of the two indices into a dual-channel convolution kernel, perform spatial-frequency joint convolution, and output a crack opening risk distribution array; perform pseudo-color mapping on the crack opening risk distribution array to generate a primary crack morphology sub-map; the kernel size of the dual-channel convolution kernel is dynamically determined by the maximum particle size of the concrete aggregate in the dam body, and the convolution step size adapts to the length of the evaluation unit.

[0063] Nonlinear normalization employs a piecewise exponential function to compress safety indicators, eliminating dimensional differences while preserving nonlinear relationships. The dual-channel convolution kernel includes spatial and frequency domain convolution branches. The spatial branch captures the spatial correlation of local risk distributions through deformable convolution kernels, while the frequency branch extracts the amplitude-frequency characteristics of risks through Fast Fourier Transform. Pseudo-color mapping uses the HSV color space to map risk values ​​to hue and saturation parameters, enhancing the visual recognizability of risk gradients. When dynamically adjusting the kernel size, the kernel side length is set to 1.5 times the aggregate particle size based on the measured maximum aggregate particle size, ensuring that the convolution operation covers the risk distribution area around a single aggregate. When adaptively changing the convolution stride, the stride is shortened by 15% for every 10% increase in evaluation cell length, ensuring consistent coverage of convolution operations within cells of different scales.

[0064] In some embodiments, the first and second opening-distance safety indices are nonlinearly normalized to form two normalized vectors. These two vectors are input to the spatial and frequency domain branches of a dual-channel convolutional kernel, respectively. The spatial branch extracts risk distribution features in the spatial dimension using a deformable convolutional kernel, while the frequency domain branch converts the risk signal to the frequency domain using a fast Fourier transform, extracting high-frequency abrupt changes and low-frequency trend components. The two outputs are superimposed after inverse Fourier transform to generate a crack opening risk distribution array. This array is converted into a gradient image from blue to red using pseudo-color mapping, where red areas represent high-risk areas and blue areas represent low-risk areas. Dynamic adjustment of the kernel size allows the convolution operation to adapt to changes in the scale of local risk features caused by different aggregate distributions, avoiding feature omissions due to a fixed kernel size. Adaptive stride variation ensures that the sampling density of the convolution operation is inversely proportional to the cell length when the evaluation cell length changes, maintaining the integrity of risk feature extraction. The resulting primary crack morphology sub-map accurately reflects the risk gradient at different spatial locations, providing high-precision input for subsequent secondary morphology correction.

[0065] As a preferred embodiment, the solution of this application is specifically implemented as follows: The first and second opening-distance safety indices are respectively subjected to nonlinear normalization to obtain normalized vectors for the two indices. The nonlinear normalization uses the sigmoid function for mapping, compressing the original index values ​​to the 0-1 interval.

[0066] Two index-normalized vectors are input into a dual-channel convolution kernel, which performs joint spatial-frequency convolution to output a risk distribution array of fracture openings. The spatial branch of the dual-channel convolution kernel uses a 3x3 convolution kernel, and the frequency branch uses a Fourier-transformed spectral convolution. The risk distribution array is obtained by weighted summation of the joint convolution results.

[0067] The risk distribution array at the fracture opening is subjected to pseudo-color mapping to generate a primary fracture morphology sub-map. The JET color mapping scheme is used to map the risk values ​​from low to high as a gradient color spectrum of blue-cyan-yellow-red.

[0068] The kernel size of the dual-channel convolution kernel is dynamically determined by the maximum particle size of the concrete aggregate in the weir, and the convolution step size adapts to the length of the evaluation unit. For example, when the maximum aggregate particle size is 40mm, the kernel size is set to 5x5; when the evaluation unit length is 10m, the convolution step size is set to 2.

[0069] This application further proposes a dynamic texture-edge collaborative filtering algorithm to perform primary screening of the seepage-temperature coupled signal and generate a primary crack determination map. This includes: constructing a joint texture-edge feature set of the seepage-temperature coupled signal; the joint feature set includes thermal image texture entropy, seepage edge gradient, and temperature change ridge; associating each feature with several background interference suppression operators; calculating the saliency response of each feature based on the weighted response of all background interference suppression operators corresponding to each feature; performing adaptive weighted fusion on the saliency responses of all features to obtain the primary crack determination map; the weight vector of the adaptive weighted fusion is updated through online Bayesian inference, and the update rate is adjusted in real time by the rate of change of the upstream water level of the weir.

[0070] The texture-edge joint feature set captures the non-uniformity of temperature distribution through thermal image texture entropy, extracts abrupt changes in the leakage path through leakage edge gradient, and identifies abnormal change regions in the temperature field through temperature abrupt change ridges. Background interference suppression operators are designed for different features; for example, thermal image texture entropy uses local entropy filtering to suppress environmental thermal noise, leakage edge gradient uses directional gradient threshold filtering to eliminate seepage fluctuation interference, and temperature abrupt change ridges use a ridge tracking algorithm to eliminate equipment heating interference. Weighted response calculation is performed through parallel processing of multiple operators, with the weight of each operator dynamically adjusted based on historical interference patterns. During adaptive weighted fusion, Bayesian inference adjusts the weight update step size according to the real-time water level change rate; for example, when the water level rise rate exceeds 0.5 meters per second, the weight update interval is shortened to the millisecond level.

[0071] Specifically, in the leakage-temperature coupled signal processing, the thermal image texture entropy is first calculated using a sliding window to determine the local entropy value. The window size is set to three times the concrete aggregate particle size to cover typical temperature anomaly areas. The leakage edge gradient is calculated using the Sobel operator to determine the anisotropic gradient magnitude, with the gradient direction aligned with the main seepage path of the weir. Temperature abrupt change ridges are identified by extracting the extreme points of ridge curvature using continuous wavelet transform, with the ridge length threshold set to twice the minimum crack propagation length. Background interference suppression operators perform noise reduction on each feature, such as applying median filtering to the thermal image texture entropy, with the filter kernel size proportional to the concrete pouring layer thickness. In the weighted response calculation stage, the outputs of each operator are weighted using normalized Euclidean distance, with the weight coefficients dynamically adjusted according to the current environmental noise level. During adaptive fusion, the weight vector is updated online using a Bayesian probability model. The model's prior distribution is initialized based on historical water level data, and the posterior distribution's update frequency is adjusted according to the real-time water level change rate. When the upstream water level fluctuates drastically, the fusion weights are recalculated at a higher frequency to ensure that the crack determination map can quickly respond to changes in seepage conditions, thereby improving the robustness of the primary screening results.

[0072] As a preferred embodiment, the solution of this application is specifically implemented as follows: A joint feature set of texture and edge features for the leakage-temperature coupled signal is constructed. The joint feature set includes thermal image texture entropy, leakage edge gradient, and temperature abrupt ridge. Each feature is associated with several background interference suppression operators. For example, the thermal image texture entropy feature is associated with Gaussian filtering and median filtering operators, the leakage edge gradient feature is associated with Sobel and Canny operators, and the temperature abrupt ridge feature is associated with morphological opening and closing operators.

[0073] Furthermore, the saliency response of each feature is calculated based on the weighted responses of all background interference suppression operators corresponding to that feature. Specifically, for the thermal image texture entropy feature, Gaussian filtering and median filtering operators are applied to process the original thermal image, and the difference in texture entropy before and after filtering is calculated as the saliency response. For the leakage edge gradient feature, Sobel and Canny operators are applied to extract the edges, and the weighted superposition of the two edge maps is calculated as the saliency response. For the temperature change ridge feature, morphological opening and closing operators are applied to process the temperature map, and the temperature change ridge is extracted as the saliency response.

[0074] Therefore, an adaptive weighted fusion is performed on the saliency responses of all features to obtain a preliminary crack determination map. The weight vector of the adaptive weighted fusion is updated through online Bayesian inference, and the update rate is adjusted in real time by the rate of change of the upstream water level. For example, when the upstream water level changes rapidly, the update step size of the weight vector is increased to quickly adapt to environmental changes; when the upstream water level changes slowly, the update step size of the weight vector is decreased to maintain the stability of the determination result.

[0075] This application further proposes a dam crack cognition network, including an input layer for receiving a primary crack judgment map, a crack morphology analysis map, and a service safety map. After unifying the scale and grayscale of the three maps, the network is fed into a nonlinear fusion layer. The nonlinear fusion layer performs element-level multi-sensory fusion on the unified primary judgment tensor, morphology analysis tensor, and service safety map tensor to output a crack cognition tensor. The output layer performs channel attention weighting on the crack cognition tensor and outputs a crack identification map after activation function mapping. The element-level multi-sensory fusion is completed by three parallel branches, each containing dilated convolutions with different expansion rates to capture the short-, medium-, and long-range spatial correlations of cracks.

[0076] The input layer uses an interpolation algorithm to adjust the resolution of the three maps to the same size and employs histogram matching technology to unify the grayscale distribution range and eliminate data scale differences. The nonlinear fusion layer uses three parallel branches with dilated convolution kernels of 1, 3, and 5 expansion rates respectively. The branch with expansion rate 1 extracts detailed features of the crack edges, the branch with expansion rate 3 captures the correlation features between cracks and adjacent structural units, and the branch with expansion rate 5 analyzes the distribution pattern of cracks within the overall dam body. The output layer calculates the weight coefficients of each channel through a channel attention mechanism. The weight coefficients are dynamically adjusted based on the variance distribution of the crack perception tensor; for example, the channel weights in the crack leading edge region are increased to 0.7-0.9, while the weights in the background region are decreased to 0.1-0.3. The activation function is a leaky ReLU, with the leakage slope dynamically calculated based on the ratio of the dam body concrete carbonation depth to its service life. When the service life exceeds 20 years, the leakage slope is set to 0.05-0.1.

[0077] In some embodiments, the input layer first adjusts the pixel size of the primary crack determination map, crack morphology analysis map, and service safety map to 512×512 using bilinear interpolation, and normalizes the grayscale range to 0-255. In the nonlinear fusion layer, the size of the dilated convolution kernels of the three branches is fixed at 3×3, and the dilation rates are coprime to avoid overlapping feature responses. During element-level multisensory fusion, the primary determination tensor, morphology analysis tensor, and safety map tensor are input into the three branches respectively, and the branch outputs are multiplied and summed to generate the crack cognition tensor. The output layer filters high-variance feature channels through a channel attention module. For example, the channel weight for crack regions with abnormal seepage channel connectivity is increased to 0.85, while the channel weight for regions with stable structural strain is reduced to 0.15. The leaky ReLU function uses a slope of 0.08 in the negative region to prevent gradient vanishing. This scheme reduces the false detection rate of crack identification maps to 2.3%-3.1% by fusing local morphology and global distribution features of cracks through multi-scale hollow convolution and dynamic weight allocation, which improves the accuracy by about 15.7% compared with traditional single-branch fusion methods.

[0078] As a preferred embodiment, the solution of this application is specifically implemented as follows: The dam crack recognition network comprises an input layer, a nonlinear fusion layer, and an output layer. The input layer receives a primary crack determination map, a crack morphology analysis map, and a service safety map. After scaling and grayscale normalization of the three maps, they are fed into the nonlinear fusion layer. The nonlinear fusion layer performs element-level multi-sensory fusion on the unified primary determination tensor, morphology analysis tensor, and service safety map tensor, outputting a crack recognition tensor. The output layer performs channel attention weighting on the crack recognition tensor, and after activation function mapping, outputs a crack identification map.

[0079] Element-level multisensor fusion is accomplished by three parallel branches, each containing dilated convolutions with different dilation rates to capture short-, medium-, and long-range spatial correlations in the gaps. Specifically, the first branch uses dilated convolutions with a dilation rate of 1, the second branch uses dilated convolutions with a dilation rate of 3, and the third branch uses dilated convolutions with a dilation rate of 5. All three branches have a 3x3 kernel size and a stride of 1.

[0080] Furthermore, the channel attention weighting mechanism is implemented using the Squeeze-and-Excitation (SE) module. First, global average pooling is performed on the crack cognition tensor to obtain the channel descriptor. Then, channel weights are generated through two fully connected layers and a non-linear activation function. Finally, the channel weights are multiplied by the original feature map to obtain the weighted feature map.

[0081] Therefore, the activation function used is Leaky ReLU, whose mathematical expression is f(x) = max(0.01x, x). Compared with standard ReLU, Leaky ReLU has a small slope on the negative half-axis, which can alleviate the problem of neuron death and improve the expressive power of the network.

[0082] This application further proposes to perform element-wise multiplication on the outputs of the three branches and then input them into a sigmoid gate, and then perform element-wise summation on the outputs of the three branches and then input them into a tanh gate; perform element-wise multiplication on the sigmoid gate result and the tanh gate result to generate a crack cognitive tensor; the activation function adopts a leaky ReLU, and the leak slope is dynamically adjusted by the service life of the weir.

[0083] In this design, three hollow convolutional branches with coprime expansion rates are set to 2, 3, and 5, respectively, to cover short-range local details, mid-range morphological correlations, and long-range structural trends of the cracks through different expansion rates. A sigmoid gate is used to nonlinearly compress the element-level product results, filtering out high-confidence feature regions; a tanh gate normalizes the element-level summation results, preserving the overall distribution characteristics of the multi-branch fusion. In the leaky ReLU activation function, the leakage slope is adjusted linearly according to the service life of the weir; for every 10 years of service life, the leakage slope increases by 0.02 to adapt to the nonlinear response changes caused by the aging of concrete materials.

[0084] In some embodiments, the spatial features output by the three branches are used to highlight important features with high co-occurrence through element-wise multiplication, and spatial attention weights in the 0-1 interval are generated by sigmoid gating. Element-wise summation preserves the complementary information of each branch, and feature distribution corrections in the -1 to 1 interval are generated by tanh gating. After multiplying the two, interference from insignificant regions is suppressed while preserving the complete gradient information of multi-scale features. The leaky ReLU dynamically adjusts the gradient in the negative interval according to the service life of the weir. For example, a 0.15 leak slope is used for a 20-year-old weir, and a 0.18 slope is used for a 30-year-old weir. This allows the crack identification model to adapt to the degree of concrete carbonation depth and elastic modulus degradation, improving the prediction accuracy of crack propagation trends under different aging states.

[0085] As a preferred embodiment, the specific implementation of this application is as follows: The primary crack determination map, crack morphology analysis map, and service safety map are preprocessed through an input layer to form a primary determination tensor, a morphology analysis tensor, and a safety map tensor. These three tensors are input into three hollow convolutional branches with dilation rates of 2, 3, and 5, respectively. The kernel size of each branch is set to 5×5, with coprime dilation rates to avoid overlap of receptive fields between different branches. The three branches extract features from the local texture features, mid-range morphological correlations, and long-range seepage pressure effects of the cracks. The branch outputs are multiplied element-wise and then input into a sigmoid function to generate spatial attention weights. Simultaneously, the element-wise summation of the branch outputs is used to generate feature modulation coefficients via a tanh function. Finally, the attention weights and modulation coefficients are multiplied element-wise to form the crack cognition tensor. In the leaky ReLU activation function, the leakage slope is dynamically adjusted according to the dam's construction time; specifically, the slope value increases linearly by 0.02 for every ten years of service life, with a maximum of 0.2.

[0086] This application further proposes a method for generating a seepage stability map by performing a convolution operation between a fracture surface water pressure attenuation convolution kernel and a real-time seepage pressure field, and then inputting the convolution result into a hyperbolic tangent activation function; wherein the size of the convolution kernel is adjusted synchronously with the thickness of the weir's seepage prevention curtain; a structural strain hazard map is generated by performing a cross-correlation operation between the fracture trajectory and the overall field strain, and then inputting the cross-correlation result into a ReLU activation function; wherein the cross-correlation kernel direction is aligned with the principal stress direction of the weir in real time; a crack propagation trend map is generated by performing an element-wise multiplication between the Laplace operator output of the upstream velocity field and a crack front curvature-driven convolution kernel, and then inputting the product result into a Sigmoid activation function; finally, the seepage stability map, structural strain hazard map, and crack propagation trend map are spliced ​​along the channel to form a service safety map.

[0087] The logic for adjusting the size of the convolution kernel for attenuating water pressure on the fracture surface is as follows: when the thickness of the seepage barrier increases, the size of the convolution kernel shrinks linearly to match the reduced seepage path; the cross-correlation kernel direction alignment is achieved by capturing the principal stress direction of the weir body in real time through the built-in gyroscope, driving the cross-correlation kernel to rotate in the same direction; the scale annealing step size of the convolution kernel driven by the curvature of the fracture front is dynamically calculated based on the reciprocal of the ratio of the fracture propagation rate to the elastic modulus, and the step size shortens when the ratio increases.

[0088] Specifically, in the pressure stability assessment, the convolution kernel size is adjusted synchronously with the thickness of the seepage barrier to ensure accurate capture of water pressure attenuation characteristics under different thicknesses. The hyperbolic tangent activation function maps the convolution results to the [-1,1] interval, eliminating dimensional differences. In the structural strain hazard assessment, the cross-correlation kernel direction is aligned with the principal stress direction in real time, enabling precise quantification of the spatial correlation between crack trajectories and the strain field. The ReLU activation function filters out negative correlation interference. In crack propagation trend prediction, the product of the curvature-driven convolution kernel and the output of the velocity field Laplace operator reflects the influence of hydrodynamics on crack propagation. The Sigmoid function compresses the results to the probability space. After the three-channel maps are stitched together, the independent features of each assessment dimension are retained along the spatial dimension, forming a multi-dimensional service safety map that includes pressure, strain, and propagation trends.

[0089] As a preferred embodiment, the specific implementation of this application is as follows: When generating the seepage stability map, a convolution operation is performed on the real-time collected seepage field data using a fracture surface water pressure attenuation convolution kernel. The size of the convolution kernel is dynamically adjusted to 3×3 or 5×5 according to the measured thickness of the seepage barrier. The convolution result is processed by a hyperbolic tangent activation function to generate the seepage stability map, wherein the weight of the convolution kernel is calibrated every 24 hours based on the upstream siltation elevation sensor data. The generation of the structural strain hazard map is achieved by cross-correlation operation between the fracture trajectory data and the strain field data. The rotation angle of the cross-correlation kernel is matched in real time with the azimuth angle output by the strain principal direction sensor. When the water level monitoring data exceeds the design flood level by 0.5 meters, the gradient descent retraining process of the cross-correlation kernel is automatically triggered. The crack propagation trend map is calculated by multiplying the gradient tensor of the velocity field after processing with the Laplace operator with a 5×5 convolution kernel generated by the curvature feature of the crack front edge element by element. The product result is input into the Sigmoid function to generate a probability distribution map. The scale annealing step size of the convolution kernel is set to one-thousandth of the output value per minute of the concrete elastic modulus tester.

[0090] This application further proposes that, during the evaluation of crack propagation trend maps, scale annealing be performed on the convolution kernel driven by the curvature of the crack front, with the annealing step size dynamically determined by the ratio of crack propagation rate to the elastic modulus of the dam concrete.

[0091] The scaling annealing of the convolution kernel, driven by the curvature of the crack front, is achieved through a successive reduction in the kernel size. The annealing step size is calculated based on the ratio of the real-time monitored crack propagation rate to the concrete's elastic modulus. This ratio is acquired in real-time by strain sensors embedded on both sides of the crack and a concrete elastic modulus tester. When the crack propagation rate accelerates, the annealing step size decreases logarithmically, rapidly reducing the kernel size to capture abrupt changes in local curvature at the crack front. Conversely, when the concrete's elastic modulus decreases, the annealing step size increases linearly, increasing the kernel size to cover a wider area affected by crack propagation. During annealing, the kernel's weight matrix is ​​Gaussian smoothed according to its current size to avoid abrupt changes in characteristic response caused by scale variations.

[0092] In some embodiments, the crack propagation rate is obtained through displacement difference calculation at crack monitoring points, and the elastic modulus of concrete is obtained through ultrasonic wave propagation velocity inversion. The ratio of the two is input into the annealing step size calculation module, and the output step size value controls the scale annealing rate of the convolution kernel. In each iteration, the convolution kernel size decays exponentially with the current step size value, while its spatial receptive field is adjusted synchronously. The annealed convolution kernel is element-wise multiplied with the Laplace operator output of the upstream velocity field, and a crack propagation trend map is generated through the Sigmoid activation function. This process, by dynamically adjusting the receptive scale of the convolution kernel, enables the model to adapt to different combinations of material properties and crack propagation rates, accurately capturing the curvature change characteristics of the crack front, thereby improving the spatiotemporal resolution and accuracy of crack propagation trend prediction.

[0093] As a preferred embodiment, the specific implementation of this application is as follows: During the evaluation of the crack propagation trend map, the scale annealing operation of the convolution kernel driven by the crack front curvature is dynamically executed. When the ratio of the crack propagation rate to the elastic modulus of the dam concrete is in the low threshold range, the annealing step size is set to a larger value to accelerate parameter updates; when the ratio enters the high threshold range, the annealing step size is automatically switched to a smaller value to improve the accuracy of parameter adjustment. During the annealing process, the scale parameters of the convolution kernel are iteratively optimized based on the real-time calculated crack propagation direction and concrete material properties. The elastic modulus data is collected in real time by an acoustic sensor array embedded inside the dam body, while the crack propagation rate is continuously measured by distributed fiber optic sensors arranged on the crack monitoring section.

[0094] Through the above technical solution, this application achieves dynamic optimization of the crack propagation trend prediction model, effectively solving the prediction deviation problem caused by the asynchronous change of material parameters and crack dynamic propagation in traditional methods. By establishing a dynamic correlation mechanism between crack propagation rate and material mechanical properties, the timeliness and accuracy of crack propagation trend prediction are significantly improved, ensuring reliable crack development status assessment results can be obtained at different stages of material performance degradation.

[0095] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for identifying cracks in the concrete of a dam, characterized in that, Includes the following steps: A leakage-temperature coupled detection scenario, a crack morphology acquisition scenario, and a service safety assessment scenario were constructed. The leakage-temperature coupled detection, crack morphology acquisition, and service safety assessment were performed on the concrete cracks of the dam body, and the raw data of leakage-temperature coupled signals, crack three-dimensional morphology signals, and service safety signals were collected simultaneously. A dynamic texture-edge collaborative filtering algorithm is used to perform primary screening of the leakage-temperature coupling signal to generate a primary crack determination map. Multi-dimensional morphological analysis is performed based on the three-dimensional morphological signal of the fracture to generate a fracture morphological analysis map; wherein, the three-dimensional morphological signal of the fracture includes the fracture opening distance, the fracture trace extension distance, and the fracture depth distance. The service safety signal is input into a pre-trained service safety prediction network to predict the service safety map; The initial crack determination map, crack morphology analysis map, and service safety map are input into the crack recognition network of the weir body to obtain the crack identification map.

2. The method according to claim 1, characterized in that, Multi-dimensional morphological analysis is performed based on the three-dimensional morphological signal of the fracture to generate a fracture morphological atlas, including: Based on the fracture opening distance, perform primary morphological analysis to generate a primary fracture morphology sub-graph; Secondary morphological corrections are performed on the primary fracture morphology sub-map based on the fracture extension distance and fracture depth distance to generate a fracture morphology analytical map. The primary morphological analysis includes dynamically comparing the fissure opening distance with the allowable opening distance threshold of the weir body design to determine whether the fissure opening is in an over-limit state. The secondary morphological correction includes calculating the ratio of the crack extension distance to the length of the dam structural unit, and calculating the ratio of the crack depth distance to the dam thickness, in order to comprehensively determine the degree to which the crack weakens the integrity of the dam. The multi-dimensional morphological analysis also includes quantifying the surface roughness, surface curvature, and flow channel connectivity of the fracture surface to generate a more comprehensive description of the fracture morphology.

3. The method according to claim 2, characterized in that, Based on the fracture opening distance, a primary morphological analysis is performed to generate a primary fracture morphology sub-image, including: The dam section to be tested is divided into several continuous evaluation units along the axis of the dam body. Within each evaluation unit, the fracture opening distance sequence is statistically analyzed, and the average opening distance value and opening distance fluctuation amplitude of the sequence are calculated. The first opening distance safety index is generated by comparing the average opening distance value with the upper limit of the allowable opening distance, and the second opening distance safety index is generated by comparing the opening distance fluctuation amplitude with the fluctuation tolerance range. A primary fracture morphology sub-map is generated based on the first and second opening distance safety indices.

4. The method according to claim 3, characterized in that, Secondary morphological corrections are performed on the primary fracture morphology sub-map based on the fracture extension distance and fracture depth distance to generate an analytical fracture morphology map, including: Calculate the deviation ratio between the crack extension distance and the length of the dam structural unit, and calculate the penetration ratio between the crack depth distance and the dam thickness. If the deviation ratio is lower than the preset extension warning ratio and the penetration ratio is lower than the preset penetration warning ratio, then the primary crack morphology sub-graph is convolved with the first correction weight matrix to obtain the crack morphology analytical map. If the deviation ratio is higher than the preset extension warning ratio and the penetration ratio is higher than the preset penetration warning ratio, then the primary crack morphology sub-graph is convolved with the second correction weight matrix to obtain the crack morphology analytical map. The first and second corrected weight matrices are obtained through offline training using historical dam failure samples, and the difference in the numerical distribution of the elements of the two matrices is greater than 50% to ensure that the correction directions are opposite.

5. The method according to claim 4, characterized in that, The primary fracture morphology sub-map generated based on the first and second opening-space safety indices includes: The first and second opening distance safety indices are respectively subjected to nonlinear normalization to obtain the normalized vector of the two indices. Input the two index normalized vectors into the dual-channel convolution kernel, perform spatial-frequency joint convolution, and output a crack opening risk distribution array; The risk distribution array of fracture openings is pseudo-color mapped to generate a primary fracture morphology sub-map. The kernel size of the dual-channel convolution kernel is dynamically determined by the maximum particle size of the concrete aggregate in the dam body, and the convolution step size adapts to the length of the evaluation unit.

6. The method according to claim 1, characterized in that, The leakage-temperature coupling signal is initially screened using a dynamic texture-edge collaborative filtering algorithm to generate a preliminary crack determination map, including: A joint feature set of texture and edge features is constructed for the leakage-temperature coupled signal; the joint feature set includes thermal image texture entropy, leakage edge gradient, and temperature change ridge; each feature is associated with several background interference suppression operators; The significance response of a feature is calculated based on the weighted response of all background interference suppression operators corresponding to that feature. An adaptive weighted fusion is performed on the saliency responses of all features to obtain a primary crack determination map; The weight vector of the adaptive weighted fusion is updated through online Bayesian inference, and the update rate is adjusted in real time by the rate of change of the water level upstream of the weir.

7. The method according to claim 1, characterized in that, The weir crack recognition network includes: The input layer receives the initial crack judgment map, crack morphology analysis map, and service safety map. After unifying the execution scale and grayscale of the three maps, they are sent to the nonlinear fusion layer. The nonlinear fusion layer is used to perform element-level multi-sensory fusion of the unified primary decision tensor, morphological analysis tensor, and security graph tensor to output the crack cognition tensor. The output layer is used to perform channel attention weighting on the crack recognition tensor, and outputs a crack recognition map after being mapped by an activation function. The element-level multisensor fusion is accomplished by three parallel branches, each containing dilated convolutions with different dilation rates to capture the short-, medium-, and long-range spatial correlations of the gaps.

8. The method according to claim 7, characterized in that, The element-level multisensory fusion outputs the crack cognition tensor according to the following logic: The primary decision tensor, morphological analysis tensor, and security graph tensor are respectively input into three dilated convolution branches with coprime dilation rates to obtain three branch outputs. Perform element-wise multiplication on the outputs of the three branches and then input the result into a sigmoid gate. Then perform element-wise summation on the outputs of the three branches and then input the result into a tanh gate. The sigmoid-gated result and the tanh-gated result are element-wise multiplied to generate the crack cognition tensor. The activation function is a leaky ReLU, and the leakage slope is dynamically adjusted according to the service life of the weir.

9. The method according to claim 1, characterized in that, The service safety signals include permeability stability signals, structural strain hazard signals, and crack propagation trend signals; the service safety prediction network outputs three types of signal maps based on the following heterogeneous logic: The seepage pressure stability map is generated by performing a convolution operation between the fracture surface water pressure attenuation convolution kernel and the real-time seepage pressure field, and then inputting the convolution result into the hyperbolic tangent activation function; wherein, the size of the convolution kernel is adjusted synchronously with the thickness of the weir seepage prevention curtain; The structural strain hazard map is generated by cross-correlation calculation between crack trajectory and full-field strain, and then inputting the cross-correlation result into the ReLU activation function; the cross-correlation kernel direction is aligned with the principal stress direction of the dam body in real time. The crack propagation trend map is generated by performing element-wise multiplication between the output of the Laplacian operator of the upstream velocity field and the crack front curvature-driven convolution kernel, and then inputting the product result into the Sigmoid activation function. Finally, the pressure stability map, structural strain hazard map, and crack propagation trend map are spliced ​​along the channel to form a service safety map.

10. The method according to claim 9, characterized in that, The assessment of the seepage stability map also includes periodic calibration of the convolution kernel for the attenuation of water pressure on the fracture surface, with the calibration period determined by the rate of change of the siltation elevation upstream of the weir. The assessment of the structural strain hazard map also includes directional retraining of the cross-correlation kernel between crack trajectory and full-field strain. The retraining is triggered when the weir encounters a water level exceeding the design check flood. The evaluation of the crack propagation trend map also includes performing scalar annealing on the curvature-driven convolution kernel at the crack front, with the annealing step size dynamically determined by the ratio of crack propagation rate to the elastic modulus of the dam concrete.

Images