Water conservancy project safety hazard assessment and prediction system and method based on image recognition

Abstract

The present invention relates to the technical field of water conservancy project safety monitoring, and specifically discloses a water conservancy project safety hazard assessment and prediction system and method based on image recognition. The system synchronously collects surface images, internal structure detection data and environmental parameters of water conservancy projects through a multimodal sensing network, adopts a multi-scale convolutional neural network combined with a three-dimensional point cloud registration technology to extract surface visual feature parameters, and utilizes an adaptive time-frequency analysis and a wavelet packet reconstruction algorithm to extract physical feature parameters of internal hidden defects. A dual machine learning framework is constructed, and environmental interference is eliminated through a deep residual network. The causal relationship between features is analyzed based on a gated recurrent unit to screen a set of key feature parameters. A Gaussian process regression model with an adaptive kernel function is used for dynamic risk prediction, and risk mutation points are identified in combination with a multi-scale wavelet transform. Finally, a safety status score and a graded warning signal are generated through a fuzzy comprehensive evaluation algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of water conservancy project safety monitoring, and in particular to a water conservancy project safety hazard assessment and prediction system and method based on image recognition. Background Art

[0002] As critical national infrastructure, the safe operation of water conservancy projects is directly linked to the safety of people's lives and property, as well as the stable development of the economy and society. Traditional water conservancy project safety monitoring relies primarily on regular manual inspections and single-sensor technology, such as point-based monitoring devices like displacement gauges and piezometers. With the advancement of computer vision and artificial intelligence technologies, structural health monitoring methods based on image recognition are increasingly being applied to surface defect detection in water conservancy facilities such as dams and levees. However, existing technologies are often limited to analyzing a single data source, lack the deep integration of multi-source, heterogeneous monitoring data, and are insufficiently capable of identifying the hidden structural defects unique to water conservancy projects. In recent years, deep learning technology has made significant progress in the field of structural damage identification, but its application in water conservancy project safety assessment still faces challenges such as significant environmental interference, inaccurate feature extraction, and poor adaptability of risk prediction models. Furthermore, due to the complex structures and harsh operating environments of water conservancy projects, conventional monitoring methods are unable to meet the needs of real-time and accurate safety warnings. Therefore, there is an urgent need to develop intelligent, multi-dimensional safety hazard assessment and prediction methods.

[0003] The existing technology has the following deficiencies:

[0004] The existing water conservancy project safety monitoring system has significant technical deficiencies in the detection and risk assessment of hidden erosion damage (internal scour / piping) in earth-rockfill dams. Specifically,

[0005] Insufficient multimodal sensor data fusion: Conventional visible light / infrared images cannot penetrate the dam structure, while sonar or geological radar (GPR) are interfered with by water turbidity and metal components, resulting in an internal piping channel miss rate exceeding the industry's acceptable threshold (e.g., >15%).

[0006] Lack of dynamic coupled modeling: Existing survival analysis models (such as the Cox PH model) rely on apparent characteristic parameters (such as crack length and displacement) and fail to integrate multi-physics coupled simulation data of seepage field, stress field, and particle migration (DEM-CFD). This results in systematic deviations in the risk function h(t) when calculating the fine particle loss rate of the dam foundation (measured errors reach ±30%).

[0007] Risk of overfitting due to small samples: Due to the scarcity of internal erosion damage cases (the global database sample size is <50 cases), machine learning models (such as random forest feature selectors) are prone to misclassifying non-critical covariates of water level fluctuations as significant factors (p-value <0.05 but actual contribution <5%), resulting in an inflated survival probability S(t) of at least 20 percentage points. Summary of the Invention

[0008] The purpose of the present invention is to provide a water conservancy project safety hazard assessment and prediction system and method based on image recognition to solve the problems in the above background.

[0009] The purpose of the present invention can be achieved through the following technical solutions:

[0010] A water conservancy project safety hazard assessment and prediction method based on image recognition collects multi-source monitoring data of the target water conservancy project, including surface image data, internal structure detection data, and environmental parameter data;

[0011] Extract features from the surface image data of the target water conservancy project to obtain visual feature parameters reflecting the structural status of the project;

[0012] Conduct time-frequency feature analysis on the internal structure detection data of the target water conservancy project to extract physical characteristic parameters that characterize hidden defects;

[0013] The visual feature parameters and physical feature parameters are input into the pre-trained feature screening model. The dual machine learning framework is used to separate the correlation interference between the feature parameters and output a set of key feature parameters that independently affect the safety status.

[0014] The key characteristic parameter set is input into the dynamic risk prediction model of water conservancy projects, and the risk probability curve that changes with time is obtained through non-parametric modeling. The dynamic risk prediction model of water conservancy projects can automatically identify risk mutation points.

[0015] Based on the risk probability curve and risk mutation points, the safety status assessment results and graded early warning signals of the water conservancy project are generated to achieve real-time prediction of safety hazards.

[0016] As a further solution of the present invention, the feature extraction of the surface image data of the target water conservancy project to obtain visual feature parameters reflecting the structural status of the project specifically includes the following steps:

[0017] A multi-scale fusion convolutional network is used to extract hierarchical features from surface images. The first layer has a convolution kernel size of 7×7 pixels to capture large-scale structural deformation features, the second layer has a convolution kernel size of 3×3 pixels to extract subtle crack texture features, and the third layer uses dilated convolution to expand the receptive field and detect potential seepage areas.

[0018] The extracted multi-scale features are input into the spatial attention mechanism module to generate a feature weight distribution map, which enhances the features of crack-dense areas and structural connections while suppressing the interference features of water reflective areas.

[0019] Use 3D point cloud registration technology to map 2D features to the engineering structure solid model, calculate the spatial continuity index of feature parameters, and eliminate false defect features caused by shooting angles;

[0020] Based on the feature clustering algorithm, similar defect patterns are automatically classified, and a set of visual feature parameters including crack length, distribution density, and strike angle is output. The crack strike angle is quantified in polar coordinate space using an improved Hough transform algorithm.

[0021] As a further solution of the present invention, the time-frequency feature analysis of the internal structure detection data of the target water conservancy project and the extraction of physical characteristic parameters characterizing hidden defects specifically include the following steps:

[0022] An improved short-time Fourier transform is used to perform time-frequency decomposition of sonar detection signals. The window function uses an adaptive Gaussian window, and the window width is dynamically adjusted according to the instantaneous frequency of the signal. A wide window is used in the low-frequency band to ensure frequency resolution, while a narrow window is used in the high-frequency band to improve time resolution.

[0023] A three-dimensional time-frequency feature matrix is ​​constructed, with the first dimension being the time series, the second dimension being the frequency component, and the third dimension being the signal amplitude. The time-frequency pattern basis vectors representing the piping characteristics are extracted using a non-negative matrix decomposition algorithm.

[0024] Match the decomposed time-frequency pattern with the pre-established defect feature library, where the pattern with a matching degree exceeding a threshold is marked as a potential defect signal;

[0025] The marked defect signal is reconstructed by wavelet packet, and its energy entropy and singular value are calculated as physical characteristic parameters, where the energy entropy is used to quantify the randomness of the defect, and the singular value is used to characterize the geometric morphological characteristics of the defect.

[0026] As a further solution of the present invention, the separation of the correlation interference between the characteristic parameters by the dual machine learning framework specifically includes the following steps:

[0027] Construct a first-level interference cancellation model using a deep residual network structure, taking environmental parameters as input and predicting their interference components on various visual feature parameters and physical feature parameters respectively;

[0028] The original characteristic parameters are subtracted from the predicted interference components to obtain the debiased characteristic parameters. The characteristic parameters whose difference exceeds three times the standard deviation are marked as outliers and the manual review process is initiated.

[0029] A second-level causal analysis model is established, using a gated recurrent unit network structure to perform temporal correlation analysis on debiased feature parameters and calculate the contribution weight of each parameter to historical security events through an attention mechanism.

[0030] The characteristic parameters are sorted according to their contribution weights, and the parameters with weight values ​​in the top 20% are selected to form a key characteristic parameter set, and a causal impact intensity report for each parameter is generated.

[0031] As a further solution of the present invention: the specific implementation of the first-level interference cancellation model includes:

[0032] Adopting a multi-task learning architecture, sharing the underlying network parameters, and simultaneously outputting the interference prediction results for each feature parameter;

[0033] Introducing an adversarial sample generation mechanism during the training phase to enhance the robustness of the model by adding noise and feature perturbations;

[0034] Set a dynamic weighted loss function and adjust the corresponding loss weight of each feature parameter according to its importance;

[0035] Regularly use the latest monitoring data to incrementally train the model to maintain the timeliness of interference prediction.

[0036] As a further solution of the present invention: the dynamic risk prediction model for water conservancy projects is constructed and operated in the following manner:

[0037] A Gaussian process regression framework using an adaptive kernel function is used, where the kernel function type is automatically selected based on the data distribution characteristics of the feature parameters. For periodic features, a periodic kernel is used, while for mutation features, a radial basis kernel is used.

[0038] Construct a hierarchical time window mechanism to divide the input time series data into long-term trend segments, medium-term fluctuation segments, and short-term mutation segments, and use kernel functions of different scales to process them respectively;

[0039] Approximate the posterior distribution through variational inference methods;

[0040] Set up a dynamic warning threshold adjustment mechanism. When the risk probability gradient exceeds the set value at three consecutive time points, the risk mutation point identification process is automatically triggered.

[0041] As a further solution of the present invention: the risk mutation point identification process specifically includes:

[0042] Perform multi-scale wavelet transform on the risk probability curve to extract the local extreme points and inflection point features of the curve;

[0043] Construct a mutation feature verification network to cross-validate the wavelet transform results with the changing trends of key feature parameters;

[0044] A double confirmation mechanism is set up. Only when the wavelet analysis results and the characteristic parameter change trend meet the mutation conditions at the same time, it is finally determined to be a valid risk mutation point;

[0045] The confirmed mutation points are graded in severity and divided into three warning levels according to the magnitude and duration of risk probability changes.

[0046] As a further solution of the present invention: the dynamic risk prediction model for water conservancy projects also includes the following optimization measures:

[0047] Introducing an expert knowledge guidance mechanism, embedding the temporal characteristics of historical major risk events as prior information into the model;

[0048] Establish a dynamic model performance evaluation system to monitor the deviation between the prediction results and the actual safety status in real time;

[0049] Set up an automatic calibration module to automatically start the model parameter adjustment process when the continuous prediction deviation exceeds the allowable range;

[0050] Build a visual explanation interface to intuitively display the formation basis and key influencing factors of the risk probability curve.

[0051] As a further solution of the present invention, the generation of the safety status assessment results and graded warning signals of the water conservancy project specifically includes the following steps:

[0052] Construct a multi-dimensional risk assessment matrix, using the integral value of the risk probability curve, the number and severity of risk mutation points as the main assessment dimensions, and introducing the changing trend of characteristic parameters as an auxiliary assessment dimension;

[0053] A fuzzy comprehensive evaluation algorithm is used to calculate the safety status score, where the weight of each evaluation dimension is dynamically adjusted based on the project type and years of operation. For old projects that have been in operation for more than 20 years, the weight of the severity of the mutation point is increased by 30%;

[0054] Design a three-level warning signal generation mechanism: when the score is in the warning range, a yellow warning is triggered and a diagnostic report is generated; when a high-risk mutation point is detected, an orange warning is triggered and emergency monitoring is initiated; when the score exceeds the danger threshold, a red warning is triggered and automatically sent to the regulatory authorities;

[0055] Establish an early warning feedback optimization system, collect matching data between actual hazard handling results and early warning signals, and dynamically optimize the parameter settings of the assessment model.

[0056] The water conservancy project safety hazard assessment and prediction system based on image recognition includes:

[0057] A multi-source data acquisition module, which is used to collect multi-source monitoring data of the target water conservancy project, including surface image data, internal structure detection data and environmental parameter data;

[0058] A visual feature extraction module is used to extract features from surface image data of a target water conservancy project to obtain visual feature parameters reflecting the structural status of the project;

[0059] A physical feature analysis module, which is used to perform time-frequency feature analysis on the internal structure detection data of the target water conservancy project and extract physical feature parameters that characterize hidden defects;

[0060] A feature optimization and screening module, which is used to input visual feature parameters and physical feature parameters into a pre-trained feature screening model, separate the correlation interference between each feature parameter through a dual machine learning framework, and output a set of key feature parameters that independently affect the safety status;

[0061] A dynamic risk assessment module, which inputs a set of key characteristic parameters into a dynamic risk prediction model for a water conservancy project and obtains a risk probability curve that changes over time through non-parametric modeling, wherein the dynamic risk prediction model for a water conservancy project can automatically identify risk mutation points;

[0062] The early warning decision output module generates the safety status assessment results and graded early warning signals of the water conservancy project according to the risk probability curve and the risk mutation point, thereby realizing the real-time prediction of safety hazards.

[0063] Beneficial effects of the present invention:

[0064] (1) The present invention achieves a comprehensive perception of the structural status of water conservancy projects by constructing a multi-source heterogeneous monitoring data acquisition system and integrating multi-dimensional information such as surface images, internal structure detection, and environmental parameters. In terms of surface image feature extraction, a deep learning architecture based on a multi-scale fusion convolutional network and a spatial attention mechanism is adopted to effectively improve the recognition accuracy and robustness of key defects such as cracks and seepage, and the spatial positioning and parameter quantification of defects are achieved through three-dimensional point cloud registration and feature clustering technology. In terms of internal structure detection, a time-frequency analysis method combining adaptive short-time Fourier transform with non-negative matrix decomposition is introduced, combined with wavelet packet reconstruction and singular value decomposition technology to accurately extract the physical characteristic parameters of hidden defects. Furthermore, a feature screening model is constructed through a dual machine learning framework to effectively separate the correlation interference between various features and extract a set of key characteristic parameters that have an independent impact on the safety status. These high-dimensional, dynamic, and physically meaningful characteristic parameters are input into a dynamic risk prediction model for water conservancy projects based on non-parametric modeling. The model adopts an adaptive kernel function selection mechanism and a hierarchical time window strategy to automatically identify risk mutation points and generate a risk probability curve that evolves over time. On this basis, the system combines multi-scale wavelet transforms with a mutation signature verification network to establish a risk mutation point identification mechanism. Through a dynamic warning threshold adjustment system and a three-level tiered warning mechanism, it achieves intelligent, graded warnings and real-time responses to project safety hazards. This comprehensive technical system not only significantly improves the accuracy and real-time nature of water conservancy project safety hazard identification, but also demonstrates excellent robustness and adaptability in complex environments, greatly enhancing project managers' in-depth understanding of potential risks and the efficiency of emergency response.

[0065] (2) The proposed method for the assessment and prediction of safety hazards in water conservancy projects based on image recognition and multi-source data fusion not only achieves a refined characterization of the structural status of the project, but also forms a comprehensive and systematic safety status quantitative assessment system by constructing a multi-dimensional risk assessment matrix including risk probability integral, number and intensity of mutation points, characteristic parameter trends and coordinated changes. The system adopts a dynamic weight fuzzy comprehensive evaluation algorithm, combines the project type, operating years and key parameter change characteristics, flexibly adjusts the weight distribution of each assessment dimension, and performs fuzzy processing through trapezoidal membership function, thereby improving the scientificity and robustness of the assessment results. On this basis, the system introduces expert knowledge graphs and typical risk case libraries to guide model predictions in a soft constraint manner, enhancing the interpretability of the assessment process and the credibility of decision-making; at the same time, a dynamic performance evaluation system and automatic calibration mechanism are constructed, which can monitor core indicators such as model prediction accuracy, warning timeliness and feature contribution in real time, and automatically trigger parameter optimization and version update processes when model performance deteriorates, ensuring the stability and adaptability of the system in long-term operation. The integration of this series of intelligent mechanisms not only significantly improves the system's self-optimization and continuous learning capabilities under complex working conditions, but also provides engineering management personnel with a high-precision, explainable, and easy-to-operate auxiliary decision support platform, which helps to achieve the transformation of safety management from passive response to active prevention and control, effectively reduce the probability of major safety accidents, and effectively protect people's lives and property safety and the long-term stable operation of water conservancy projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The present invention will be further described below with reference to the accompanying drawings.

[0067] Figure 1 This is a flowchart of a method for assessing and predicting safety hazards in water conservancy projects based on image recognition according to the present invention;

[0068] Figure 2 It is a flowchart of the water conservancy project safety hazard assessment and prediction system based on image recognition in the present invention. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0070] Example 1, please refer to Figure 1 As shown, the present invention is a method for assessing and predicting safety hazards of water conservancy projects based on image recognition, comprising the following steps:

[0071] Collect multi-source monitoring data of the target water conservancy project, including surface image data, internal structure detection data and environmental parameter data;

[0072] Extract features from the surface image data of the target water conservancy project to obtain visual feature parameters reflecting the structural status of the project;

[0073] Conduct time-frequency feature analysis on the internal structure detection data of the target water conservancy project to extract physical characteristic parameters that characterize hidden defects;

[0074] The visual feature parameters and physical feature parameters are input into the pre-trained feature screening model. The dual machine learning framework is used to separate the correlation interference between the feature parameters and output a set of key feature parameters that independently affect the safety status.

[0075] The key characteristic parameter set is input into the dynamic risk prediction model of water conservancy projects, and the risk probability curve that changes with time is obtained through non-parametric modeling. The dynamic risk prediction model of water conservancy projects can automatically identify risk mutation points.

[0076] Based on the risk probability curve and risk mutation points, the safety status assessment results and graded early warning signals of the water conservancy project are generated to achieve real-time prediction of safety hazards.

[0077] Example 2, the specific implementation process of multi-source monitoring data collection: First, deploy a multimodal sensor network system and install high-resolution industrial camera arrays at key locations of the water conservancy project, including the water-facing and water-retaining surfaces of the dam, gate connections, and flood discharge channels. The camera resolution must be no less than 20 megapixels and equipped with polarizing filters to suppress interference from water surface reflections. Simultaneously, infrared thermal imagers are deployed in areas of structural stress concentration, with a temperature detection accuracy of 0.1°C and a sampling interval of 10 minutes per time to capture thermal anomalies in concrete structures. All optical equipment is equipped with automatic cleaning devices to ensure clear images can be obtained even in rainy and foggy weather.

[0078] For internal structural detection, a distributed sonar sensing system is used in conjunction with a geological radar. A broadband sonar probe is deployed every 50 meters along the dam's axis, with an operating frequency range of 50kHz-500kHz. The probe is calibrated with the assistance of an underwater robot. The geological radar uses an antenna array with a center frequency of 1GHz to perform grid scanning along the project foundation surface, with a scanning density of 20 measurement points per square meter. In particular, three-component microseismic monitoring nodes are densely deployed in historical hidden danger areas, with a sensitivity of 0.1μm / s, to capture microfracture signals within the rock mass.

[0079] The environmental parameter monitoring system consists of a multifunctional meteorological station and a hydrological monitoring unit. The meteorological station is equipped with an anemometer, rain gauge, and atmospheric pressure sensor, with data updated once per second. The hydrological monitoring unit includes a reservoir level gauge, seepage flow meter, and turbidity meter. Reservoir level measurement uses a radar level gauge with an accuracy of ±1mm; seepage flow is monitored using a tracer dilution method with a detection limit of 0.01L / s. All environmental sensors are lightning-proof to ensure normal operation in inclement weather.

[0080] A data synchronization and preprocessing mechanism was established, using GPS disciplined clocks to provide a unified time base for all monitoring devices, with time synchronization errors controlled within 10ms. Raw data was initially verified by edge computing nodes to remove outliers caused by equipment failures, and missing data was remedied using spatiotemporal kriging interpolation. In particular, sonar data was processed to suppress water noise, using adaptive filters to eliminate signal interference caused by turbulence.

[0081] A multi-source data fusion database is constructed, aligning and storing surface image data, internal detection data, and environmental parameters in time series. Data tags include metadata such as spatial coordinates, acquisition time, and device number. A data quality assessment system is established to score each batch of incoming data based on completeness, accuracy, and consistency. Scores below a threshold automatically trigger data re-acquisition. The database supports a multi-level caching mechanism to ensure millisecond-level response times even with high concurrent access.

[0082] We implement a secure data transmission and backup strategy, using a layered encryption transmission protocol. Monitoring data is transmitted concurrently via both wired fiber and wireless 5G channels, automatically switching to local storage when network latency exceeds 200ms. We also establish an off-site disaster recovery backup center, perform incremental backups daily, and retain the last 30 days of rolling backup data. All data transmission processes are digitally signed to prevent data tampering.

[0083] Example 3, the specific implementation process of surface image data feature extraction: First, a multi-scale fusion convolutional network architecture is constructed. The network adopts a three-level hierarchical feature extraction design. The first level uses a large-size convolution kernel of 7×7 pixels to perform preliminary feature extraction on the input image. This level sets 32 feature channels with a step size of 2 pixels, which can effectively capture large-scale deformation features such as the overall deformation of the dam body and structural dislocation. The second level uses a fine convolution kernel of 3×3 pixels, and the number of feature channels is increased to 64. The feature map size is kept unchanged by zero padding, and it is specifically used to extract fine crack texture features with a width of more than 0.1mm. The third level introduces a void convolution layer with an expansion rate of 3. While maintaining the computational complexity, the receptive field is expanded to 5 times the area of ​​the original image, focusing on detecting possible seepage areas and potential weak points. A batch normalization layer and a ReLU activation function are configured after each convolution layer to avoid the gradient disappearance problem.

[0084] A spatial attention mechanism is implemented for feature enhancement. The feature maps extracted by the multi-scale convolutional network are fed into the attention module. This module first generates a channel attention vector through global average pooling and then applies a 3×3 convolution kernel to generate a spatial attention weight map. Specifically, to address the characteristics of water conservancy projects, prior knowledge constraints are introduced into the attention calculation. A base weight coefficient of no less than 0.7 is set for known crack-prone areas (such as structural joints and stress concentration zones), while the weight for reflective areas on the water surface is strictly limited to no more than 0.2. A skip connection structure is used at the output of the attention module to preserve important information from the original features. 3D point cloud registration and feature mapping are performed. A pre-acquired laser point cloud model of the engineering structure is used as a reference. This model contains no fewer than 10 million 3D points with a spacing of less than 2 mm. Feature points extracted from the 2D image are registered using an improved ICP (Iterative Closest Point) algorithm. A dynamic threshold mechanism is implemented during the registration process to automatically reduce the weight of a feature point if the matching error between the feature point and the point cloud model exceeds 5 mm. After registration is completed, the continuity index of each feature point in three-dimensional space is calculated, including the neighborhood point density change rate and normal vector consistency. Pseudo features that do not conform to spatial continuity (such as false cracks caused by shooting angles) are filtered out, and the filtering ratio is controlled within 15% of the total number of feature points.

[0085] Feature clustering and parameter quantification were performed, using an improved DBSCAN density clustering algorithm to categorize verified feature points. A dynamic neighborhood radius parameter was set, using a 3mm radius for crack feature areas and a 10mm radius for seepage areas. Parameters were calculated for the clustered feature set: crack length was obtained by summing broken lines in three-dimensional space; distribution density was calculated by counting the number of feature points per unit area; and crack strike angles were calculated using an improved Hough transform algorithm in polar coordinates, calculating the angle between the principal direction and the dam axis with an accuracy of 1 degree. The final output visual feature parameter set includes the 3D coordinates, geometric parameters, and confidence score for each defect cluster, forming a complete representation of the structural state.

[0086] Example 4, the specific implementation process of the time-frequency feature analysis of internal structure detection data: First, adaptive short-time Fourier transform processing is implemented. Aiming at the non-stationary characteristics of the water conservancy project sonar detection signal, a Gaussian window function with dynamic window width adjustment function is designed. The initial width of the window function is set to 1 / 10 of the center frequency of the signal. During the processing, the instantaneous frequency change of the signal is monitored in real time. When the frequency is detected to be lower than 200kHz, the window width is automatically expanded to 1.5 times the original value to improve the frequency resolution. When the frequency is higher than 400kHz, the window width is reduced to 0.7 times the original value to enhance the time resolution. The overlap rate of each analysis window is controlled at 75% to ensure the continuity of the time-frequency analysis. For multi-channel sonar data, beamforming technology is used for pre-processing to eliminate the interference caused by the multipath effect. According to actual measurements, this method can improve the frequency resolution of time-frequency analysis by 40% and the time resolution by 35%, which is significantly better than the traditional method with fixed window length.

[0087] A three-dimensional time-frequency feature characterization system was constructed, organizing the time-frequency analysis results output by the short-time Fourier transform into a three-dimensional matrix structure. The time axis uses 10ms as the basic unit, the frequency axis is divided into 1 / 3 octave bands, and the amplitude axis adopts a logarithmic scale. To reduce computational complexity, the amplitude data was normalized, and the dynamic range was controlled within 60dB. Focusing on the internal defect characteristics of water conservancy projects, special attention was paid to the time-frequency energy distribution patterns in the 50-150kHz and 300-500kHz frequency bands, which correspond to the typical characteristic frequencies of piping and cracks, respectively. During the three-dimensional matrix construction process, the sensor spatial coordinates corresponding to each data point were synchronously recorded to provide a basis for subsequent defect location.

[0088] Non-negative matrix factorization and feature extraction were performed, using a multi-layer iterative non-negative matrix factorization algorithm to process the three-dimensional time-frequency matrix. The decomposition layer was set to five, and three basis vectors were extracted in each layer. Sparsity constraints were introduced during the decomposition process to ensure that each basis vector only represented the characteristics of a single type of defect. In particular, a dedicated optimization objective function was set for piping characteristics to enhance its ability to extract low-frequency continuous signals. After decomposition, the obtained basis vectors were clustered, and redundant vectors with a similarity greater than 90% were removed. Ultimately, the 10 to 15 most representative time-frequency patterns were retained as the foundational elements of the feature library. These basis vectors were classified and stored according to the defect type they represented, and the corresponding project location and severity level were labeled.

[0089] Implement defect feature matching and verification, establish a dynamic threshold matching mechanism, and calculate similarity between real-time time-frequency patterns and the feature library. This similarity calculation uses an improved cosine similarity metric, assigning higher weight to areas sensitive to amplitude changes. When the match exceeds a preset threshold (the default setting is 0.85), the signal for that period is marked as a potential defect. To reduce false alarm rates, a duration verification condition is set, requiring a match to be considered valid only if it persists for more than 100ms. For confirmed defect signals, key parameters such as start time, duration, center frequency, and energy intensity are recorded to form a preliminary defect feature description.

[0090] Wavelet packet reconstruction and feature quantification are performed, and the db8 wavelet basis function is selected to perform a 5-layer wavelet packet decomposition on the marked defect signal. The optimal decomposition path is automatically selected based on the signal characteristics. During the reconstruction process, the detail coefficients of layers 2-4 are retained to effectively separate the defect characteristics from the background noise. The energy entropy calculation uses an improved algorithm based on Shannon entropy, with a window length of 50ms and a step size of 10ms, which dynamically reflects the random changes in the defect signal. Singular value decomposition is performed in the joint time-frequency domain, and the first three main singular values ​​and their ratios are extracted as defect morphological features. The final output physical feature parameter set includes the energy entropy time series curve, the singular value distribution map, and the defect spatial location information, providing a quantitative basis for subsequent risk assessment.

[0091] In Example 5, a first-level interference elimination model is first constructed. The model uses a deep residual network as the basic architecture and contains 12 residual blocks. Each residual block consists of two 3×3 convolutional layers and jump connections. The input layer is designed as a multi-channel structure, which receives time series data of environmental parameters such as temperature, humidity, and water level, and the sampling frequency is unified as 1 time / hour. The model output layer adopts a bifurcated structure and simultaneously generates interference prediction values ​​for visual feature parameters and physical feature parameters. During the model training phase, a sliding window method is used to construct training samples, with the window length set to 30 days and the step length being 1 day. The loss function uses smooth L1 loss, which has better robustness to abnormal interference values. When the model is deployed, the number of parameters is compressed to 40% of the original model through knowledge distillation technology to ensure real-time operation capabilities on edge devices.

[0092] Interference component calculation and anomaly detection are performed, and the original characteristic parameters are subtracted element-by-element from the interference prediction values ​​to obtain the debiased characteristic parameters. An adaptive scaling factor is introduced into the difference calculation process to normalize the characteristic parameters according to their dimensions. Characteristic parameters with a difference exceeding three standard deviations are automatically marked as outliers by the system, triggering a three-level processing flow: first, comparison with historical data for the same period, second, checking the sensor's operating status, and finally, submission to the manual review interface. The review interface displays the original value, interference prediction value, and debiasing results of the parameter over the past 72 hours for engineers to make their final judgment.

[0093] A second-level causal analysis model was established, employing a two-layer gated recurrent unit network with 128 hidden units per layer and an input sequence length of 30 days. A multi-head attention mechanism with four heads was implemented after the gated recurrent unit to capture feature correlations across different timescales. The model was trained using a curriculum learning strategy, first using annual data to learn long-term trends and then using quarterly data to optimize short-term fluctuations. Causal contribution weights were calculated using perturbation analysis, which sequentially masks each input feature and observes the change in the model output. Greater changes indicate higher contribution weights. Monte Carlo sampling was used for weight calculation, with 100 iterations repeated and the average taken to ensure stability.

[0094] The key feature parameter set is generated and verified. All feature parameters are sorted in descending order based on their causal contribution weights, and the top 20% of parameters are selected as the key feature set. To ensure the rationality of the selection, three verification conditions are set: the weight value must be more than twice the average value of all parameters; it must remain stable for three consecutive evaluation cycles; and it must be consistent with engineering mechanics theory. A detailed analysis report is generated for each verified key feature parameter set, including information such as the weight value of each parameter, the main influencing period, and the strength of the correlation with other parameters. The system automatically conducts a comprehensive evaluation of the key feature set every month, triggering an immediate update mechanism when the status of the engineering structure changes significantly.

[0095] Specific optimization measures for the first-level interference elimination model:

[0096] The multi-task learning architecture was optimized, with a shared feature extraction layer at the bottom of the model consisting of four convolutional layers and two fully connected layers, outputting a 256-dimensional shared feature vector. The upper task-specific layers employed a parallel structure. The visual feature interference prediction branch consisted of three fully connected layers, while the physical feature interference prediction branch consisted of two fully connected layers and a gated recurrent unit layer. The loss functions for both branches employed a dynamic weighting scheme, with an initial weight of 1:1. The weights were automatically adjusted after every 100 training rounds based on validation set performance, with adjustments not exceeding 20%.

[0097] An adversarial training enhancement mechanism is introduced. During the training data preprocessing phase, a specific perturbation strategy is designed for each type of feature parameter: ±5% Gaussian noise is added to visual features such as crack length; a random time shift (up to ±3 sampling points) is applied to physical features such as energy entropy; and a combination of numerical and temporal perturbations is applied to environmental parameters. The proportion of adversarial examples generated is controlled between 30% and 50%, and is dynamically adjusted with each training round. During adversarial training, the model uses the FGSM method to calculate the perturbation direction, and the perturbation amplitude is adaptively determined based on the current model performance.

[0098] A dynamic weight adjustment system was established. The importance of feature parameters was assessed based on three metrics: engineering sensitivity (determined by expert knowledge), measurement reliability (calculated based on sensor accuracy), and time series stability (measured by the coefficient of variation). Loss function weights were updated monthly, taking into account the weighted average of these three metrics. To maintain training stability, a weight change smoothing mechanism was implemented, ensuring that weight adjustments between successive cycles did not exceed 15%. The system retained a record of weight adjustments for the past 12 months for model performance analysis.

[0099] We implement an incremental learning update strategy and establish a data sliding window mechanism to keep the most recent three years of data online. Earlier data is compressed and stored as feature vectors. The model performs incremental training monthly, with training data consisting of newly added data and randomly sampled historical data (in a 1:1 ratio). This incremental training utilizes an elastic weight fixation algorithm to impose stronger constraints on important parameters, preventing new data from overwriting old knowledge.

[0100] Example 6, specific implementation process of constructing and operating a dynamic risk prediction model for water conservancy projects:

[0101] First, an adaptive kernel function selection mechanism was established. Taking into account the multi-scale characteristics of water conservancy project monitoring data, an intelligent kernel function selector was designed. This selector automatically determines the optimal kernel function combination by analyzing the statistical properties of characteristic parameters. For parameters exhibiting obvious periodicity (such as diurnal variations in reservoir water levels), a periodic kernel function is automatically selected, with the period length determined through spectral analysis. For parameters with abrupt changes (such as sudden crack expansion), a radial basis kernel function is used, with its length scale dynamically adjusted through maximum likelihood estimation. During the model initialization phase, a 30-day observation analysis is performed on each key characteristic parameter to establish an initial kernel function configuration. The system automatically reassesses the characteristic changes of each parameter weekly and adjusts the kernel function type and related hyperparameters as necessary. This mechanism ensures that the model can adapt to the risk prediction needs of different water conservancy project types and different operation stages.

[0102] A tiered time window processing strategy is implemented, dividing the input time series data into three levels for processing: For long-term trends (data over one year), an annual sliding window with a kernel function length scale of 90 days is used to capture slow changes such as project aging; for medium-term fluctuations (data from one month to one year), a quarterly sliding window with a kernel function length scale of 30 days is used to monitor seasonal variations; for short-term sudden changes (data within one month), a daily sliding window with a kernel function length scale of 7 days is used to capture sudden risks. The results of each level of analysis are weighted and integrated to generate the final forecast. The weights are dynamically adjusted based on the project type, for example, earth-rockfill dams focus on medium- and long-term trends, while concrete dams focus on short-term changes. The windows are divided using an overlapping approach, with the overlap ratio controlled at 50% to ensure forecast continuity.

[0103] A variational inference method is used for efficient computation, and a variational posterior approximation framework based on the Gaussian distribution is designed to reduce the computational complexity of traditional Gaussian process regression from cubic to linear. The variational distribution family uses the mean field approximation and contains 100 induction points, which are selected from the training data using the k-means clustering algorithm. The evidence lower bound is optimized using natural gradient descent, and the learning rate is set to adaptive adjustment mode with an initial value of 0.01 and a decay of 10% every 100 iterations. To maintain numerical stability, a diagonal perturbation term of 1e-6 is added to the covariance matrix calculation. The model performs a full training once a week and incremental updates are performed daily to ensure a balance between prediction accuracy and timeliness.

[0104] A dynamic early warning threshold adjustment system is constructed. The early warning threshold is not a fixed value, but an intelligent parameter that is dynamically adjusted according to the project's operating status. The basic threshold is set at twice the standard deviation of the average daily change rate of the risk probability. A preliminary early warning is triggered when the gradient value exceeds this threshold at three consecutive time points. The system simultaneously monitors multiple auxiliary indicators: the persistence of gradient changes, the degree of coordinated changes in related characteristic parameters, and abnormal environmental conditions. These factors are combined to calculate the final early warning level through a fuzzy logic system. The threshold adjustment module recalculates the probability fluctuation range under the normal operation of the project every month as a new baseline reference. During special periods such as flood season or periods of active earthquakes, the system automatically increases the monitoring frequency and appropriately lowers the threshold to enhance risk sensitivity.

[0105] Specific implementation process of the risk mutation point identification process:

[0106] Multiscale wavelet transform analysis was performed, using the db4 wavelet basis function to perform a five-layer decomposition of the risk probability curve, obtaining detail coefficients and approximate coefficients at different scales. A shift-invariant wavelet transform algorithm was used during the decomposition process to avoid positioning errors caused by phase distortion. Local extreme points and zero-crossing points were calculated for the wavelet coefficients at each scale and used as potential candidate mutation signatures. To reduce noise interference, an amplitude threshold filtering mechanism was implemented to retain only significant extreme points exceeding three times the background noise level. The wavelet analysis window was set to a seven-day window length, with daily sliding updates to ensure real-time mutation detection.

[0107] A mutation signature verification network was constructed, employing a dual-channel architecture. One channel processes wavelet analysis results and consists of three convolutional layers and one long short-term memory (LSTM) layer. The other channel processes the changing trends of key feature parameters and consists of two fully connected layers and one attention layer. The outputs of the two channels are fused at the decision layer, and the mutation confidence is calculated using a sigmoid function. The network is trained using a contrastive learning approach, with positive samples drawn from data 72 hours prior to historical risk events and negative samples from data during normal operation. The verification network is retrained quarterly to ensure adaptability to changes in project status. When wavelet analysis detects a potential mutation point, the system automatically extracts data from the 24 hours before and after the event and feeds it into the verification network for secondary verification results.

[0108] A dual-confirmation decision-making mechanism is implemented, and risk mutation points are only finally confirmed when both wavelet analysis and the verification network give a positive conclusion. The confirmation conditions for wavelet analysis include: synchronous extreme points are detected on at least three adjacent scales, and the amplitude of the extreme points shows a monotonically increasing trend. The confirmation threshold for the verification network is a confidence score exceeding 0.85. For confirmed mutation points, the system records their complete characteristics, including start time, peak intensity, and duration, and automatically correlates them with major events in the engineering log to continuously improve the decision-making rule base. To reduce the risk of false alarms, a silent period mechanism is implemented to suspend the detection of new mutation points within 6 hours after confirming a mutation point to avoid repeated alarms.

[0109] A three-level warning system has been established. Level 1 (yellow) corresponds to a single-day increase in risk probability of 10% to 20% that persists for no more than 24 hours, triggering enhanced routine inspections. Level 2 (orange) corresponds to a single-day increase of 20% to 50%, or an increase exceeding 10% for two consecutive days, initiating special inspections and emergency preparedness. Level 3 (red) corresponds to a single-day increase exceeding 50% or a sustained deterioration, directly notifying management and initiating emergency response plans. Once the warning level is determined, the system automatically generates a report containing the following elements: location of the sudden change point, analysis of possible causes, affected project areas, and recommended countermeasures. All warning events are entered into the knowledge base, serving as the foundation for continuous model optimization.

[0110] Specific implementation process of model optimization measures:

[0111] By embedding an expert knowledge guidance mechanism, a knowledge graph of water conservancy project risks is constructed, containing time-series characteristic patterns of over 200 typical risk cases. Key nodes in these patterns are annotated by domain experts, such as the changing patterns of characteristic parameters in the 30 days before accelerated crack expansion. During the model prediction process, the similarity between the current data and various patterns in the knowledge graph is calculated in real time. When the similarity exceeds a threshold, the confidence weight of the prediction result is automatically adjusted. Expert knowledge is incorporated into the model as soft constraints, influencing the loss function through regularization terms rather than directly overriding data-driven results. The knowledge base is updated every six months to incorporate new engineering cases and research results.

[0112] A dynamic performance evaluation system was established, with multi-dimensional evaluation metrics designed: prediction accuracy (risk event detection rate), timeliness (early warning lead time), stability (volatility in prediction results), and interpretability (clarity of feature contributions). The system monitors these metrics in real time, triggering a performance alert if any metric falls below the preset threshold for five consecutive days. The evaluation process utilizes a sliding window approach, with long-term metrics evaluated over a one-year window, a mid-term window of one quarter, and a short-term window of one month, providing a comprehensive understanding of model performance. Evaluation results are intuitively displayed in a radar chart, helping operations personnel quickly understand model status.

[0113] An automated calibration process was implemented, triggered by conditions such as a forecast deviation exceeding 15% for three consecutive days or a significant change in the contribution ranking of key feature parameters. The calibration process utilizes Bayesian optimization to adjust kernel function hyperparameters and variational distribution parameters while maintaining the core model structure. A model snapshot was automatically created before calibration. After calibration, the model was run in a shadow environment for 48 hours to verify the results. Once improvements were confirmed, the model was deployed to the production environment. To maintain forecast continuity, the calibration process employed a gradual update strategy, running the old and new versions in parallel for a period of time, gradually switching traffic.

[0114] A visual interpretation interface and an interactive risk probability curve display system were developed to support multi-level drill-down analysis: At the macro level, overall risk trends are displayed; at the meso level, mutation points and warning periods are marked; and at the micro level, the specific contributions of each characteristic parameter are revealed. Heat maps are used to display the changes in characteristic parameter importance over time, while flow diagrams are used to illustrate the interactions between parameters. The interface also integrates a what-if analysis function, allowing users to adjust the value of a characteristic parameter and observe the impact on risk prediction in real time. All visualization elements support multi-dimensional filtering by project location, parameter type, and time range to meet the decision-making needs of different users.

[0115] Example 7, specific implementation process of safety status assessment and early warning signal generation:

[0116] The specific implementation process for constructing a multidimensional risk assessment matrix is ​​as follows: First, an assessment system consisting of five core dimensions is established: the integral value of the risk probability curve over the past seven days reflects the degree of short-term risk accumulation; the number of risk mutation points over the past 30 days indicates system instability; the average severity of mutation points reflects the intensity of the risk; the slope changes of key characteristic parameters indicate development trends; and the coordinated changes between parameters detect abnormal patterns. Each dimension is assigned a standardized scoring range of 0-100 and normalized using the range method. The baseline values ​​of the dimensions vary for different types of water conservancy projects: concrete dams focus on the coordinated changes in crack parameters, while earth-rockfill dams focus on the mutation characteristics of seepage parameters. The assessment matrix is ​​updated hourly, and a complete record of the past 90 days is retained for trend analysis.

[0117] A dynamic weighted fuzzy comprehensive evaluation algorithm was implemented, and an adjustable weight distribution system was designed. The basic weights were set as follows: risk integral value 30%, number of mutation points 25%, severity 25%, characteristic trend 15%, and coordinated change 5%. Weight adjustment takes into account two key factors: the project type is determined through a preset template. For example, the arch dam template increases the characteristic trend weight to 20%. The operating life is adjusted using a piecewise function. For projects under 10 years old, the weight of mutation points is reduced by 5%, while for projects over 20 years old, the weight of mutation points is increased by 30%. Fuzzy processing uses a trapezoidal membership function to soften the score of each dimension and avoid assessment fluctuations caused by rigid thresholds. An expert rule base is introduced into the evaluation process to automatically modify the scoring results when specific parameter combination patterns appear.

[0118] The generation and distribution of three-level warning signals are implemented. The trigger condition for the yellow warning is that the comprehensive score enters the range of 30-50. The system automatically generates a diagnostic report containing the following contents: ranking of major risk factors, recent parameter change curve, and comparative analysis of similar projects. The orange warning is for two situations: a mutation point is detected where the risk probability increases by more than 25% in a single day, or the score is in the range of 50-70 for three consecutive days. The emergency monitoring plan activated at this time includes: drone inspections of key areas, intensified monitoring of seepage volume, and structural vibration frequency detection. The red warning is triggered when the score exceeds 70 or an extreme mutation occurs. The warning information is pushed to the command center of the regulatory department in real time through a dedicated channel, accompanied by detailed drawings of the project structure, emergency plans, and resource scheduling suggestions. All warning signals are recorded with timestamps accurate to the second and automatically associated with the video surveillance screen.

[0119] The specific implementation process for establishing an early warning feedback optimization system is as follows: a closed-loop learning mechanism is established. After an early warning is issued, the system continuously tracks the subsequent processing process and records the following key data: warning response time, on-site confirmation results, and the effectiveness of the response measures. An algorithm for evaluating early warning-event matching is developed, providing quantitative scores based on three dimensions: temporal consistency (warning lead time), spatial consistency (location accuracy), and intensity consistency (severity matching). Parameter optimization and adjustment are performed monthly, recalibrating the weight distribution and threshold settings of the evaluation matrix based on the matching results. A conservative optimization strategy is adopted during the optimization process, with single adjustments not exceeding 10% of the original value to ensure system stability. A typical case library is established, and highly matched early warning events are abstracted into rule patterns to guide subsequent evaluations.

[0120] Example 8, please refer to Figure 2 As shown in FIG, the water conservancy project safety hazard assessment and prediction system based on image recognition includes:

[0121] A multi-source data acquisition module, which is used to collect multi-source monitoring data of the target water conservancy project, including surface image data, internal structure detection data and environmental parameter data;

[0122] A visual feature extraction module is used to extract features from surface image data of a target water conservancy project to obtain visual feature parameters reflecting the structural status of the project;

[0123] A physical feature analysis module, which is used to perform time-frequency feature analysis on the internal structure detection data of the target water conservancy project and extract physical feature parameters that characterize hidden defects;

[0124] A feature optimization and screening module, which is used to input visual feature parameters and physical feature parameters into a pre-trained feature screening model, separate the correlation interference between each feature parameter through a dual machine learning framework, and output a set of key feature parameters that independently affect the safety status;

[0125] A dynamic risk assessment module, which inputs a set of key characteristic parameters into a dynamic risk prediction model for a water conservancy project and obtains a risk probability curve that changes over time through non-parametric modeling, wherein the dynamic risk prediction model for a water conservancy project can automatically identify risk mutation points;

[0126] The early warning decision output module generates the safety status assessment results and graded early warning signals of the water conservancy project according to the risk probability curve and the risk mutation point, thereby realizing the real-time prediction of safety hazards.

[0127] The present invention works by constructing a multimodal sensing network that simultaneously collects multi-source monitoring data on water conservancy project surface images, internal structures, and environmental parameters using a high-resolution industrial camera array, infrared thermal imagers, sonar detection systems, and environmental sensors. A data fusion mechanism is then established to align time and space. For surface image data, a multiscale fusion convolutional network combined with a spatial attention mechanism is used for feature extraction. Three-dimensional point cloud registration technology is used to accurately map two-dimensional features to physical structures, outputting visual feature parameters including crack distribution and structural deformation. Adaptive time-frequency analysis is then performed on the internal detection data, extracting physical feature parameters that characterize hidden defects through non-negative matrix factorization and wavelet packet reconstruction. A dual machine learning framework is then constructed. The first stage uses a deep residual network to eliminate environmental interference, while the second stage uses a gated recurrent unit to analyze the causal contribution of feature parameters and screen a set of key features. Based on this feature set, a Gaussian process regression model with an adaptive kernel function is established. Dynamic risk prediction is performed using a hierarchical time window and variational inference methods. Multi-scale wavelet transforms are then used to intelligently identify risk mutation points. Finally, a multi-dimensional assessment matrix was constructed, combined with a fuzzy comprehensive evaluation algorithm to generate a safety status score. This system triggered graded warnings based on risk levels, and established a warning feedback mechanism to continuously optimize model performance. This system implements intelligent processing throughout the entire process, from data collection, feature extraction, risk prediction, to warning decision-making, significantly improving the accuracy and timeliness of identifying safety hazards in water conservancy projects.

[0128] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for assessing and predicting safety hazards in water conservancy projects based on image recognition, characterized in that: The following steps are involved: Collect multi-source monitoring data of the target water conservancy project, including surface image data, internal structure detection data and environmental parameter data; Feature extraction is performed on the surface image data of the target water conservancy project to obtain visual feature parameters reflecting the structural status of the project. The specific steps include: A multi-scale fusion convolutional network is used to extract hierarchical features from surface images. The first layer has a convolution kernel size of 7×7 pixels to capture large-scale structural deformation features, the second layer has a convolution kernel size of 3×3 pixels to extract subtle crack texture features, and the third layer uses dilated convolution to expand the receptive field and detect potential seepage areas. The extracted multi-scale features are input into the spatial attention mechanism module to generate a feature weight distribution map, which enhances the features of crack-dense areas and structural connections while suppressing the interference features of water reflective areas. Use 3D point cloud registration technology to map 2D features to the engineering structure solid model, calculate the spatial continuity index of feature parameters, and eliminate false defect features caused by shooting angles; Based on the feature clustering algorithm, similar defect patterns are automatically classified and a set of visual feature parameters including crack length, distribution density, and strike angle is output. The crack strike angle is quantified in polar coordinate space using an improved Hough transform algorithm. Perform time-frequency feature analysis on the internal structure detection data of the target water conservancy project to extract physical characteristic parameters that characterize hidden defects. The specific steps include: An improved short-time Fourier transform is used to perform time-frequency decomposition of sonar detection signals. The window function uses an adaptive Gaussian window, and the window width is dynamically adjusted according to the instantaneous frequency of the signal. A wide window is used in the low-frequency band to ensure frequency resolution, while a narrow window is used in the high-frequency band to improve time resolution. A three-dimensional time-frequency feature matrix is ​​constructed, with the first dimension being the time series, the second dimension being the frequency component, and the third dimension being the signal amplitude. The time-frequency pattern basis vectors representing the piping characteristics are extracted using a non-negative matrix decomposition algorithm. Match the decomposed time-frequency pattern with the pre-established defect feature library, where the pattern with a matching degree exceeding a threshold is marked as a potential defect signal; The marked defect signal is reconstructed by wavelet packet, and its energy entropy and singular value are calculated as physical characteristic parameters, where the energy entropy is used to quantify the randomness of the defect, and the singular value is used to characterize the geometric morphological characteristics of the defect; The visual feature parameters and physical feature parameters are input into the pre-trained feature screening model. The dual machine learning framework is used to separate the correlation interference between the feature parameters and output a set of key feature parameters that independently affect the safety status. The key characteristic parameter set is input into the dynamic risk prediction model of water conservancy projects, and the risk probability curve that changes with time is obtained through non-parametric modeling. The dynamic risk prediction model of water conservancy projects can automatically identify risk mutation points. Based on the risk probability curve and risk mutation points, the safety status assessment results and graded early warning signals of the water conservancy project are generated to achieve real-time prediction of safety hazards.

2. The method for assessing and predicting safety hazards of water conservancy projects based on image recognition according to claim 1 is characterized in that: The dual machine learning framework is used to separate the correlation interference between the feature parameters, specifically including the following steps: Construct a first-level interference cancellation model using a deep residual network structure, taking environmental parameters as input and predicting their interference components on various visual feature parameters and physical feature parameters respectively; The original characteristic parameters are subtracted from the predicted interference components to obtain the debiased characteristic parameters. The characteristic parameters whose difference exceeds three times the standard deviation are marked as outliers and the manual review process is initiated. A second-level causal analysis model is established, using a gated recurrent unit network structure to perform temporal correlation analysis on debiased feature parameters and calculate the contribution weight of each parameter to historical security events through an attention mechanism. The characteristic parameters are sorted according to their contribution weights, and the parameters with weight values ​​in the top 20% are selected to form a key characteristic parameter set, and a causal impact intensity report for each parameter is generated.

3. The method for assessing and predicting safety hazards of water conservancy projects based on image recognition according to claim 2 is characterized in that: The specific implementation of the first-level interference cancellation model includes: Adopting a multi-task learning architecture, sharing the underlying network parameters, and simultaneously outputting the interference prediction results for each feature parameter; Introducing an adversarial sample generation mechanism during the training phase to enhance the robustness of the model by adding noise and feature perturbations; Set a dynamic weighted loss function and adjust the corresponding loss weight of each feature parameter according to its importance; Regularly use the latest monitoring data to incrementally train the model to maintain the timeliness of interference prediction.

4. The method for assessing and predicting safety hazards of water conservancy projects based on image recognition according to claim 1 is characterized in that: The dynamic risk prediction model for water conservancy projects is constructed and operated in the following ways: A Gaussian process regression framework using an adaptive kernel function is used, where the kernel function type is automatically selected based on the data distribution characteristics of the feature parameters. For periodic features, a periodic kernel is used, while for mutation features, a radial basis kernel is used. Construct a hierarchical time window mechanism to divide the input time series data into long-term trend segments, medium-term fluctuation segments, and short-term mutation segments, and use kernel functions of different scales to process them respectively; Approximate the posterior distribution through variational inference methods; Set up a dynamic warning threshold adjustment mechanism. When the risk probability gradient exceeds the set value at three consecutive time points, the risk mutation point identification process is automatically triggered.

5. The method for assessing and predicting safety hazards of water conservancy projects based on image recognition according to claim 4 is characterized in that: The risk mutation point identification process specifically includes: Perform multi-scale wavelet transform on the risk probability curve to extract the local extreme points and inflection point features of the curve; Construct a mutation feature verification network to cross-validate the wavelet transform results with the changing trends of key feature parameters; A double confirmation mechanism is set up. Only when the wavelet analysis results and the characteristic parameter change trend meet the mutation conditions at the same time, it is finally determined to be a valid risk mutation point; The confirmed mutation points are graded in severity and divided into three warning levels according to the magnitude and duration of risk probability changes.

6. The method for assessing and predicting safety hazards of water conservancy projects based on image recognition according to claim 5 is characterized in that: The dynamic risk prediction model for water conservancy projects also includes the following optimization measures: Introducing an expert knowledge guidance mechanism, embedding the temporal characteristics of historical major risk events as prior information into the model; Establish a dynamic model performance evaluation system to monitor the deviation between the prediction results and the actual safety status in real time; Set up an automatic calibration module to automatically start the model parameter adjustment process when the continuous prediction deviation exceeds the allowable range; Build a visual explanation interface to intuitively display the formation basis and key influencing factors of the risk probability curve.

7. The method for assessing and predicting safety hazards of water conservancy projects based on image recognition according to claim 1 is characterized in that: The generation of the water conservancy project safety status assessment results and graded warning signals specifically includes the following steps: Construct a multi-dimensional risk assessment matrix, using the integral value of the risk probability curve, the number and severity of risk mutation points as the main assessment dimensions, and introducing the changing trend of characteristic parameters as an auxiliary assessment dimension; A fuzzy comprehensive evaluation algorithm is used to calculate the safety status score, where the weight of each evaluation dimension is dynamically adjusted based on the project type and years of operation. For old projects that have been in operation for more than 20 years, the weight of the severity of the mutation point is increased by 30%; Design a three-level warning signal generation mechanism: when the score is in the warning range, a yellow warning is triggered and a diagnostic report is generated; when a high-risk mutation point is detected, an orange warning is triggered and emergency monitoring is initiated; when the score exceeds the danger threshold, a red warning is triggered and automatically sent to the regulatory authorities; Establish an early warning feedback optimization system, collect matching data between actual hazard handling results and early warning signals, and dynamically optimize the parameter settings of the assessment model.

8. The water conservancy project safety hazard assessment and prediction system based on image recognition is characterized by: The method for evaluating and predicting safety hazards of water conservancy projects based on image recognition according to any one of claims 1 to 7 comprises: A multi-source data acquisition module, which is used to collect multi-source monitoring data of the target water conservancy project, including surface image data, internal structure detection data and environmental parameter data; A visual feature extraction module is used to extract features from surface image data of a target water conservancy project to obtain visual feature parameters reflecting the structural status of the project; A physical feature analysis module, which is used to perform time-frequency feature analysis on the internal structure detection data of the target water conservancy project and extract physical feature parameters that characterize hidden defects; A feature optimization and screening module, which is used to input visual feature parameters and physical feature parameters into a pre-trained feature screening model, separate the correlation interference between each feature parameter through a dual machine learning framework, and output a set of key feature parameters that independently affect the safety status; A dynamic risk assessment module, which inputs a set of key characteristic parameters into a dynamic risk prediction model for a water conservancy project and obtains a risk probability curve that changes over time through non-parametric modeling, wherein the dynamic risk prediction model for a water conservancy project can automatically identify risk mutation points; The early warning decision output module generates the safety status assessment results and graded early warning signals of the water conservancy project according to the risk probability curve and the risk mutation point, thereby realizing the real-time prediction of safety hazards.

Images