Water conservancy project safety detection early warning method based on artificial intelligence

Abstract

The invention relates to the technical field of water conservancy project detection, and discloses a water conservancy project safety detection early warning method based on artificial intelligence. The method comprises the following steps: acquiring multi-modal monitoring data of a key part through a distributed sensor network, and extracting a dynamic feature sequence in a preset time period through space-time alignment and noise filtering; inputting the image into a deep neural network fused with an attention mechanism, constructing a multi-scale space-time correlation map through hierarchical feature learning, and generating a high-dimensional representation of an engineering structure state; historical accident case data is used as a supervision signal, a hybrid expert model is used for performing multi-task training on high-dimensional representation, and the contribution weight of each monitoring index to the safety risk is obtained; combining real-time environment parameters and structural response characteristics to construct a dynamic threshold adjustment model, adaptively updating an early warning threshold according to a risk probability, and screening out key risk factors of which the contribution weights are greater than the updated threshold; and on the basis of spatial and temporal distribution characteristics, through graph neural network node association reasoning, multi-source early warning information is fused to generate a graded early warning result.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering detection technology, specifically to an artificial intelligence-based method for safety detection and early warning of water conservancy projects. Background Technology

[0002] As a crucial infrastructure, the safety and stability of water conservancy projects directly impact the production, daily life, and ecological environment of the surrounding areas. With the passage of time, changes in the natural environment, and the influence of human factors, various structural defects and safety hazards can gradually emerge, such as dam leakage, slope instability, and pipeline corrosion. Failure to detect and address these issues in a timely manner could lead to serious accidents.

[0003] Traditional safety monitoring of water conservancy projects relies heavily on manual inspections and single-point sensor monitoring, which has significant limitations. Manual inspections are limited by personnel experience and inspection frequency, making it difficult to achieve all-weather, full-coverage monitoring, and their ability to identify hidden defects is weak. Single-point sensor monitoring data has a single dimension, lacking a comprehensive consideration of the overall condition of the project structure, and is prone to misjudgment or omission due to local data deviations. At the same time, water conservancy projects have complex structures, and monitoring data from different parts have significant spatiotemporal correlations. Traditional methods are unable to effectively explore these correlations, resulting in insufficient timeliness and accuracy of early warnings.

[0004] In recent years, with the development of IoT and AI technologies, multi-sensor collaborative monitoring and intelligent data analysis have been increasingly applied to the field of water conservancy project safety monitoring. However, existing technologies often neglect the spatiotemporal synchronization of different sensors during data processing, leading to biases in feature extraction; model construction often employs a single neural network structure, making it difficult to fully capture the deep correlations of multimodal data; and warning thresholds are mostly fixed values, unable to be dynamically adjusted according to real-time environmental changes, making it difficult to adapt to complex and ever-changing engineering operating conditions. These problems result in existing methods still exhibiting issues such as delayed warnings and high false alarm rates in practical applications, failing to meet the high-precision and high-reliability safety assurance requirements of water conservancy projects. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based method for safety detection and early warning of water conservancy projects, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for safety detection and early warning of water conservancy projects based on artificial intelligence, the method comprising:

[0007] S1. Collect multimodal monitoring data of key parts of water conservancy projects through a distributed sensor network, perform spatiotemporal alignment and noise filtering on the raw data, and extract dynamic feature sequences within a preset time period;

[0008] S2. Input the dynamic feature sequence into a deep neural network with an attention mechanism, and construct a multi-scale spatiotemporal correlation map through hierarchical feature learning to generate a high-dimensional representation of the state of the engineering structure.

[0009] S3. Using historical accident case data as a supervision signal, a hybrid expert model is used to train the high-dimensional representation in multiple tasks to obtain the contribution weight of each monitoring indicator to safety risk;

[0010] S4. Combining real-time environmental parameters and structural response characteristics, construct a dynamic threshold adjustment model, adaptively update the early warning threshold based on the risk probability output by the model, and screen out key risk factors whose contribution weight is greater than the updated threshold.

[0011] S5. Based on the spatiotemporal distribution characteristics of key risk factors, a graph neural network is used to perform node association reasoning and integrate multi-source early warning information to generate graded early warning results.

[0012] Preferably, step S1 includes the following:

[0013] A distributed network of vibration sensors, stress sensors, and hydrological sensors deployed on the dam body, slopes, and flood discharge facilities is used to synchronously collect data on vibration acceleration, structural strain, and water level and flow velocity. The collected raw data is timestamped and outliers are removed. A sliding time window is used to divide the continuous data into segments, and time-domain statistics and frequency-domain characteristics within each window are calculated. The time-domain statistics and frequency-domain characteristics are then classified and integrated according to sensor type to form a dynamic feature sequence containing spatial location labels.

[0014] Preferably, step S2 includes the following:

[0015] A deep neural network architecture integrating spatial attention and channel attention is constructed, comprising a feature encoding layer and an association inference layer. The feature encoding layer uses a convolutional neural network to extract local spatiotemporal features, strengthens the feature weights of key parts through a spatial attention module, and suppresses redundant signal interference through a channel attention module. The association inference layer models the topological relationship between different sensor nodes through a graph convolutional network, dynamically adjusts the node connection strength using an attention mechanism, and fuses the encoded local features with global association features to generate a high-dimensional representation of the engineering structure state.

[0016] Preferably, step S3 includes the following:

[0017] Monitoring data and failure mode labels from historical engineering accident cases are collected and classified into task types such as structural cracks, abnormal seepage, and slope instability. A hybrid expert model is constructed, which includes multiple sub-expert networks and a gating network. The sub-expert networks are trained for different task types, and the gating network dynamically allocates task weights based on the feature distribution of the input high-dimensional representation. The model parameters are optimized using the accident occurrence probability as the loss function. The gradient contribution of each monitoring indicator in different tasks is calculated through backpropagation, and the safety risk contribution weight of each indicator is obtained after normalization.

[0018] Preferably, step S4 includes the following:

[0019] Real-time environmental parameters include rainfall, temperature change rate, and upstream and downstream water level difference; structural response characteristics include vibration dominant frequency offset and strain accumulation rate.

[0020] Preferably, step S4 further includes the following:

[0021] To quantify the dynamic impact of environmental factors on engineering structures, the time-varying curve of rainfall and the gradient characteristics of temperature change rate are first extracted from real-time environmental parameters. A long short-term memory network is used to predict the trend of the time-varying curve and calculate the predicted rainfall value for the next 24 hours. The predicted value is compared with historical data for the same period and converted into an environmental risk coefficient through a membership function. Combined with the gradient characteristics of temperature change rate, a comprehensive environmental impact index is generated by weighted summation.

[0022] Preferably, step S4 further includes the following:

[0023] To characterize the degree of anomaly in the structural response, the power spectral density of the vibration signal and the cumulative curve of the strain data are first extracted from the structural response characteristics. Wavelet transform is used to perform time-frequency decomposition of the power spectral density to calculate the offset of the dominant frequency component. The cumulative curve is linearly fitted, and the strain accumulation rate is obtained through residual analysis. After normalizing the dominant frequency offset and the strain accumulation rate, the comprehensive anomaly value of the structural response is calculated using Euclidean distance.

[0024] Preferably, step S4 further includes the following:

[0025] The comprehensive environmental impact index, the comprehensive structural response anomaly value, and the contribution weights of each monitoring indicator are combined to form a risk assessment vector, which is then input into a pre-trained gradient boosting tree model. The model output is labeled as the real-time risk probability. The product of the basic early warning threshold and the real-time risk probability is adjusted by a correction coefficient, which is dynamically updated according to the project's operating years. Based on the adjusted early warning threshold, indicators with contribution weights greater than the early warning threshold are selected from the monitoring indicators, and their corresponding feature sequences are labeled as key risk factors.

[0026] Preferably, step S5 includes the following:

[0027] Topological relationships are extracted from the spatial location labels of key risk factors to construct a directed graph model of the engineering structure. A graph attention network is used to rank the importance of nodes in the directed graph model and calculate the risk propagation probability of each node. Based on the risk propagation probability, key risk factors are divided into source nodes, intermediate nodes, and terminal nodes. The risk signals of nodes at different levels are aligned in time series, and multi-source information is fused through attention weight allocation to generate a graded early warning result that includes risk level, impact range, and early warning timeliness.

[0028] Preferably, the step of using a graph attention network to rank the importance of nodes in the directed graph model and calculating the risk propagation probability of each node includes:

[0029] An engineering structure graph model is constructed, which includes node features and edge weights. The node features are normalized values ​​of key risk factors, and the edge weights are dynamically calculated based on sensor spatial distance and signal correlation. A bidirectional gated recurrent unit is used as the node update function of the graph neural network, and neighbor node information is fused through a neighborhood aggregation mechanism. An attention mechanism is introduced to weight the contributions of different neighboring nodes, and high-order correlation features of nodes are learned through multi-layer graph convolution operations. The learned node features are input into a fully connected layer, and the risk propagation probability of each node is output.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] Multimodal monitoring data is collected through a distributed sensor network, covering key parts of the water conservancy project, enabling comprehensive perception of the project's status. Spatiotemporal alignment and noise filtering of the raw data eliminate deviations in time and space between different sensor data, reduce the impact of noise from environmental interference, and make the extracted dynamic feature sequences more closely match the actual state of the project structure.

[0032] When processing dynamic feature sequences, deep neural networks that integrate attention mechanisms can focus on features that significantly impact engineering safety through hierarchical feature learning. At the same time, they can construct multi-scale spatiotemporal correlation maps to effectively explore the intrinsic connections between data from different monitoring points and different time periods, thereby generating a high-dimensional representation that better reflects the overall state of the engineering structure and showing the state change patterns of the engineering structure from local to global and from short-term to long-term.

[0033] By using historical accident case data as monitoring signals and employing a hybrid expert model for multi-task training, the model can learn the correlation patterns between various monitoring indicators and safety risks from historical experience. This allows the model to determine the contribution weight of each indicator, making risk assessment more closely aligned with the risk evolution patterns in actual engineering operations. This weight allocation based on real-world case training can accurately identify the monitoring indicators that play a dominant role in safety risk assessment.

[0034] A dynamic threshold adjustment model, constructed by combining real-time environmental parameters and structural response characteristics, can flexibly update the early warning threshold based on real-time risk probabilities, avoiding the limitations of fixed thresholds when the environment changes. By screening key risk factors with contribution weights greater than the updated threshold, the model can focus on the factors that have the greatest impact on the current safety status, reducing interference from irrelevant information.

[0035] Based on the spatiotemporal distribution characteristics of key risk factors, using graph neural networks for node association reasoning can further analyze the spatial propagation path and temporal evolution trend of key risk factors. By integrating multi-source early warning information to generate graded early warning results, the early warning information becomes more targeted and hierarchical, making it easier for relevant personnel to take corresponding countermeasures according to the early warning level, and better meeting the actual needs of water conservancy project safety management. Attached Figure Description

[0036] Figure 1 This is a schematic diagram illustrating the working principle of the artificial intelligence-based water conservancy project safety detection and early warning method described in this invention.

[0037] Figure 2 This is a flowchart of multimodal data acquisition and feature extraction.

[0038] Figure 3 A flowchart for multi-task training and weight calculation;

[0039] Figure 4 Flowchart for calculating environmental risk coefficient;

[0040] Figure 5 This is a flowchart for dynamic threshold adjustment and key factor screening. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Please see Figure 1This invention provides an artificial intelligence-based method for safety detection and early warning of water conservancy projects, the method comprising:

[0043] S1. Multimodal monitoring data of key parts of the water conservancy project are collected through a distributed sensor network. The raw data is then spatiotemporally aligned and noise filtered to extract dynamic feature sequences within a preset time period. Specifically, multi-dimensional data is acquired using a sensor network deployed in various key areas of the project. Time synchronization technology ensures the consistency of data from different sensors in the time dimension. Filtering algorithms are used to remove noise interference from the data, and dynamic feature sequences reflecting changes in the project's status are extracted according to a set time interval.

[0044] S2. The dynamic feature sequence is input into a deep neural network with an attention mechanism. Through hierarchical feature learning, a multi-scale spatiotemporal correlation map is constructed, generating a high-dimensional representation of the engineering structure's state. The deep neural network, through a multi-level learning process, gradually mines deep features in the data. The introduction of the attention mechanism allows the network to focus on features that have a significant impact on engineering safety. Combined with multi-scale analysis methods, correlation maps at different temporal and spatial scales are constructed, ultimately forming a high-dimensional data representation that comprehensively reflects the engineering structure's state.

[0045] S3. Using historical accident case data as supervisory signals, a hybrid expert model is employed for multi-task training of the high-dimensional representation to obtain the contribution weights of each monitoring indicator to safety risks. Relevant data from past engineering accidents are collected, organized, and labeled, and used as training targets to guide the hybrid expert model in learning. The model analyzes different types of safety risks through multi-task training, thereby calculating the contribution weights of each monitoring indicator in the safety risk assessment.

[0046] S4. Combining real-time environmental parameters and structural response characteristics, a dynamic threshold adjustment model is constructed. The warning threshold is adaptively updated based on the risk probability output by the model, and key risk factors with contribution weights greater than the updated threshold are selected. Real-time environmental parameter and structural response data are collected and input into the dynamic threshold adjustment model. The model automatically adjusts the warning threshold based on the calculated risk probability, and then compares the contribution weights of each monitoring indicator with the updated threshold to select key risk factors exceeding the threshold.

[0047] S5. Based on the spatiotemporal distribution characteristics of key risk factors, a graph neural network is used for node association reasoning to integrate multi-source early warning information and generate tiered early warning results. The graph neural network is used to model and analyze the temporal and spatial distribution relationships of key risk factors. Through inter-node association reasoning, the intrinsic connections between risk factors are uncovered. By integrating early warning information from multiple sources, the final tiered early warning results are generated according to risk level classification standards.

[0048] Example 1: See Figure 2 In step S1, the construction of the distributed sensor network covers key components such as the dam body, slopes, and spillway facilities. Vibration and stress sensors are deployed at different elevations on the dam body. Sensors are arranged differently on different slope sections according to slope gradient and geological conditions. Hydrological and vibration sensors are installed at key nodes of the spillway facilities, such as gates and the inner walls of the spillway. The vibration sensors are triaxial accelerometers, capable of simultaneously acquiring vibration acceleration data in three directions. Their sampling frequency is set to 100Hz, which can capture structural vibrations caused by various factors such as water flow impact, equipment operation, and external vibrations. The stress sensors are fiber optic strain gauges. This type of sensor has the characteristics of anti-electromagnetic interference and high measurement accuracy. They are embedded in key stress areas inside the dam concrete and at the connection points of the slope anchorage structure to record the strain changes of the structure under load in real time. Hydrological sensors include water level gauges and flow velocity meters. Water level gauges are installed at the front and rear of the dam and at the inlet and outlet of the spillway, while flow velocity meters are deployed at representative sections of the spillway where the water flow is rapid, to simultaneously acquire water level fluctuation data and water flow velocity data.

[0049] The collected raw data is first timestamped. Using GPS synchronization technology, the time recording accuracy of all sensors is controlled to the millisecond level, ensuring consistency in time across data collected from different locations and types of sensors, avoiding data analysis errors caused by time deviations. Outliers in the raw data are identified and processed using the 3σ criterion: the mean and standard deviation of each sensor's data sequence are calculated, and data exceeding ±3 times the standard deviation are identified as outliers. These outliers are then imputed using linear interpolation, constructing a linear function from the adjacent normal data points to calculate and replace the outlier, maintaining the continuity of the data sequence. A sliding time window is used to divide the continuous data segment, with a window size of 60 seconds and a sliding step of 10 seconds. Each window contains 60 seconds of data, with a 50-second overlap between adjacent windows. For the data within each window, time-domain statistics are calculated, including mean, variance, peak value, and kurtosis. The mean reflects the central tendency of the data, the variance reflects the dispersion, the peak value captures extreme values, and the kurtosis describes the steepness of the data distribution. Simultaneously, the time-domain data is converted to the frequency domain using Fourier transform to extract frequency domain features, including dominant frequency, spectral energy, and bandwidth. Dominant frequency refers to the frequency component with the highest energy, spectral energy reflects the energy magnitude of that frequency component, and bandwidth refers to the frequency range where energy is concentrated. The extracted time-domain statistics and frequency domain features are categorized and integrated according to the type of vibration sensor, stress sensor, and hydrological sensor. A corresponding spatial location label is added to each feature, containing the specific location of the sensor (such as dam elevation, slope section, flood discharge facility node, etc.) and coordinate information. This ultimately forms a dynamic feature sequence containing spatial information, which can simultaneously reflect the temporal variation pattern and spatial distribution characteristics of the monitoring data.

[0050] In step S2, the deep neural network architecture fusing spatial attention and channel attention consists of a feature encoding layer and an association inference layer. The feature encoding layer employs a three-layer convolutional neural network. The first convolutional operation uses 3×3 kernels, with a total of 64 kernels and a stride of 1. By sliding the convolutional kernels across the input data, it extracts local temporal features of the data and captures short-term data change patterns. The second convolutional layer uses 5×5 kernels, increasing the number to 128, with a stride of 2. While further extracting deeper local spatiotemporal features, it also reduces the dimensionality of the feature map, focusing on more critical feature information. The third convolutional layer uses 3×3 kernels, with a total of 256 kernels and a stride of 1. By increasing the number of convolutional kernels, it enhances the expressive power of the features and strengthens the characterization of complex features. The spatial attention module is connected after the third convolutional layer. Its operation is as follows: First, global average pooling and max pooling are performed on the feature maps output by the convolution, resulting in two distinct feature maps. These two feature maps are then fed into a convolutional network with a single hidden layer. This network learns to output a spatial attention weight map, where each element corresponds to the importance of its position on the feature map. The weight map is multiplied by the input feature map, strengthening the feature weights of critical areas (such as stress concentration zones in dams and unstable slope sections) while relatively weakening the feature weights of less important areas. The channel attention module adopts a squeeze-excitation structure. First, global average pooling is performed on the feature maps, compressing the features of each channel into a single value to obtain a channel descriptor. This descriptor reflects the overall strength of the corresponding channel's features. Then, the channel descriptor is fed into a two-layer fully connected network. The first fully connected layer reduces the dimensionality of the descriptor, while the second fully connected layer increases it. A sigmoid activation function is then used to generate channel attention weights. These weights are multiplied by the feature map to suppress redundant signal channels that contribute little to engineering safety detection and early warning, highlighting the features of important signal channels.

[0051] The inference layer takes the features output from the feature encoding layer as input. It first constructs a topological graph of sensor nodes, treating each sensor as a node, with node features being encoded local features. The graph convolutional network contains two layers: the first layer has 128 hidden units, and the second has 64. An adjacency matrix models the connections between different nodes, with elements determined based on the physical distance and signal correlation between sensors. An attention mechanism generates attention coefficients by calculating feature similarity between nodes, dynamically adjusting node connection strength to focus more on strongly correlated node pairs (e.g., sensors in adjacent dam sections or hydrological sensors along the same flow path). The encoded local features are then fused with the globally correlated features obtained through the graph convolutional network, combining them to form a comprehensive feature that includes both local details and global correlations. This comprehensive feature is then mapped to a high-dimensional space through a fully connected layer, generating a high-dimensional representation of the engineering structure's state. This high-dimensional representation comprehensively and deeply reflects the overall state of the engineering structure.

[0052] Example 2: See Figure 3 In step S3, when collecting historical engineering accident case data, records of various accidents similar in type and scale to current water conservancy projects from the past 30 years, both domestically and internationally, are included. Monitoring data before and after the accident is extracted from these records, including multi-dimensional data such as vibration acceleration, structural strain, and water level and flow velocity. Corresponding failure mode labels are also compiled. Failure modes are categorized into task types such as structural cracks, abnormal seepage, and slope instability. Monitoring data for structural cracks mainly includes strain abrupt changes captured by stress sensors, frequency shifts recorded by vibration sensors, and displacement changes in related locations. Monitoring data for abnormal seepage involves abnormal fluctuations in water level differences, sudden increases or decreases in seepage flow, and changes in soil moisture content. Monitoring data for slope instability includes displacement data at different slope depths, stress distribution in the anchoring structure, and pore water pressure in the surrounding soil.

[0053] The constructed hybrid expert model comprises five sub-expert networks and one gated network. Each sub-expert network adopts the same three-layer fully connected structure: the first layer has 256 hidden units, the second layer has 128 hidden units, and the third layer has 64 hidden units. Each layer uses the ReLU activation function to enhance the model's ability to fit nonlinear features. These sub-expert networks are trained independently for different task types, such as structural cracks, seepage anomalies, and slope instability. By learning the correlation between monitoring data and failure modes for the corresponding task, they develop specialized risk identification capabilities. The gated network consists of two fully connected layers and a softmax activation function. The first layer has 128 hidden units, and the number of units in the second layer is the same as the number of sub-expert networks. By receiving a high-dimensional representation of the input, the fitness of each sub-expert network in the current task is calculated, and task weights are dynamically allocated, enabling the model to automatically select the more suitable sub-expert network for processing based on the characteristics of the input data.

[0054] During model training, the probability of accident occurrence is used as the loss function, and the cross-entropy loss function is employed to optimize the model parameters. During backpropagation, the gradient contribution of each monitoring indicator in different tasks is calculated. Specifically, the gradient of each monitoring indicator's features is solved at the model output layer to obtain the indicator's influence on the loss function, thus reflecting the indicator's importance in risk assessment. Subsequently, the gradient contributions of the same monitoring indicator across all tasks are normalized. By mapping the gradient contribution values ​​to the range of 0-1, the contribution weights of each monitoring indicator in different safety risk types are obtained. These weights quantify the influence of different indicators on various types of risks.

[0055] In step S4, the real-time environmental parameters collected cover rainfall, temperature change rate, and upstream and downstream water level difference. Rainfall data is acquired through multiple rain gauges distributed at different locations within the project area. These rain gauges are evenly distributed around the dam, on the top of the slope, and near the spillway area, recording cumulative rainfall hourly to form a continuous rainfall sequence. The temperature change rate is collected by temperature sensors installed on the structural surface and in the surrounding environment. The ratio of the temperature difference between two adjacent hours to time is calculated to reflect the rate of temperature change. The upstream and downstream water level difference is calculated from the water level data from the upstream and downstream water level gauges. The upstream water level gauge is installed in the reservoir area at a certain distance from the dam, while the downstream water level gauge is installed in the river channel downstream of the dam. Data is transmitted between the two gauges in real time, and the upstream and downstream water level difference is calculated by the difference. The value is dynamically updated according to water level changes.

[0056] Structural response characteristics mainly include vibration dominance frequency offset and strain accumulation rate. Vibration dominance frequency offset is obtained through spectral analysis of signals collected by vibration sensors. First, the vibration signal is converted from the time domain to the frequency domain to determine the current dominant frequency. Then, it is compared with the reference dominant frequency recorded under normal engineering conditions, and the difference between the two is calculated to obtain the vibration dominance frequency offset. Strain accumulation rate is obtained by processing strain data over a period of time. Continuous strain monitoring data is collected, and a linear fitting method is used to obtain a trend line of strain change over time. The slope of the trend line is used to calculate the strain increment per unit time, i.e., the strain accumulation rate, thus reflecting the cumulative deformation of the structure under load. These real-time environmental parameters and structural response characteristics together constitute an important basis for assessing the safety status of the project, providing fundamental data for subsequent risk analysis and early warning threshold adjustment.

[0057] Example 3: See Figure 4 In step S4, to quantify the dynamic impact of environmental factors on the engineering structure, the time-varying curve of rainfall and the gradient characteristics of the temperature change rate are extracted from real-time environmental parameters. The time-varying curve of rainfall, with time on the horizontal axis and rainfall amount on the vertical axis, is converted from discrete hourly rainfall data collected by the rain gauge into a continuous curve using linear interpolation, allowing the curve to intuitively reflect the continuous trend of rainfall change over time. The gradient characteristics of the temperature change rate are obtained by calculating the difference between the temperature change rates of two adjacent time periods, for example, calculating the difference between the temperature change rate of the current hour and the previous hour, thus reflecting the change in the rate of temperature change.

[0058] A Long Short-Term Memory (LSTM) network was used to predict the trend of time-varying rainfall curves. This network consists of one input layer, two hidden layers, and one output layer. The input layer has 12 nodes, corresponding to rainfall data from the past 12 hours. The first hidden layer contains 64 LSTM units, and the second hidden layer contains 32 LSTM units to capture long-term dependencies in the rainfall data. The output layer has 24 nodes, corresponding to the predicted rainfall for the next 24 hours. During network training, historical rainfall data from the past 10 years was used as training samples. The network parameters were adjusted to gradually reduce the deviation between the predicted results and the actual data. The predicted rainfall for the next 24 hours was compared with historical rainfall data for the same period (the same month and time period in the past 5 years), and the deviation rate was calculated as (predicted value - historical average for the same period) / historical average for the same period. The deviation rate was converted into an environmental risk coefficient using a triangular membership function. This function categorized the deviation rate into three levels: low, medium, and high, each corresponding to a different risk coefficient value. The smaller the deviation rate, the lower the environmental risk coefficient, and vice versa.

[0059] By combining the gradient characteristics of the temperature change rate, it is standardized by mapping the gradient characteristic value to the range of 0-1 using the formula (gradient characteristic value - minimum gradient characteristic value) / (maximum gradient characteristic value - minimum gradient characteristic value). Then, the standardized temperature change rate gradient characteristic is weighted and summed with the environmental risk coefficient to generate a comprehensive environmental impact index, calculated as follows:

[0060] I E =0.6×C R +0.4×G T

[0061] Among them, I E C represents the comprehensive environmental impact index. R G represents the environmental risk coefficient. T This represents the standardized temperature change rate gradient characteristic.

[0062] To characterize the degree of anomalousness in the structural response, the power spectral density of the vibration signal and the cumulative curve of strain data are extracted from the structural response characteristics. The power spectral density of the vibration signal is calculated using the Welch method, which divides the vibration signal into multiple overlapping segments, performs a Fourier transform on each segment and calculates the power spectrum, and then averages these power spectra to obtain the final power spectral density, thus reflecting the energy distribution of the signal at different frequency components. The cumulative curve of strain data is plotted with time on the horizontal axis and cumulative strain on the vertical axis. The cumulative strain value is obtained by subtracting the strain value at each moment from the strain value at the initial moment, and then connecting the cumulative strain values ​​at different moments to form the cumulative curve.

[0063] Wavelet transform was used to perform time-frequency decomposition of the power spectral density, with the db4 wavelet basis function selected and a decomposition level of 5. Wavelet coefficients for different frequency bands were obtained through decomposition, and these coefficients were then reconstructed to obtain the signal components for each frequency band. The dominant frequency component, i.e., the frequency component with the highest energy, was identified from the reconstructed signal components. The difference between the current dominant frequency component and the reference dominant frequency component recorded under normal engineering conditions was calculated to obtain the dominant frequency offset. A linear fit was performed on the cumulative curve, and the slope of the fitted line was solved using the least squares method. This slope reflects the average rate of change of strain over time. By calculating the residual between the cumulative curve and the fitted line, the degree of deviation of the strain data from the linear trend was analyzed, obtaining the strain accumulation rate, i.e., the strain increment per unit time. The dominant frequency offset and the strain accumulation rate were normalized separately, and their values ​​were mapped to the range of 0-1 using the formula (eigenvalue - minimum eigenvalue) / (maximum eigenvalue - minimum eigenvalue). Then, the Euclidean distance between these two normalized features was calculated and used as the comprehensive anomaly value of the structural response. The larger this value, the higher the degree of anomaly in the structural response. Through the above process, the dynamic impact of environmental factors on engineering structures and the degree of abnormality in structural response can be comprehensively quantified, providing key parameters for subsequent risk assessment.

[0064] Example 4: See Figure 5 In step S4, the comprehensive environmental impact index, the comprehensive structural response anomaly value, and the contribution weights of each monitoring indicator are combined to form a risk assessment vector. The comprehensive environmental impact index is a comprehensive indicator obtained by integrating environmental factors such as rainfall prediction trends and temperature change rate gradient characteristics; its value ranges from 0 to 1. The comprehensive structural response anomaly value is calculated based on the dominant frequency shift of the vibration signal and the strain accumulation rate, also within the 0-1 range, and is used to reflect the current degree of structural anomaly. The contribution weights of each monitoring indicator are derived from the training results of a hybrid expert model, covering multiple dimensions such as vibration acceleration, structural strain, and water level and flow velocity. The weight values ​​of each indicator are normalized and range from 0 to 1. These data are arranged in a fixed order, with the comprehensive environmental impact index as the first element, the comprehensive structural response anomaly value as the second element, and the contribution weights of each monitoring indicator arranged sequentially to form a multi-dimensional risk assessment vector. The dimension of the vector depends on the number of monitoring indicators.

[0065] The gradient boosting tree model is constructed using the XGBoost algorithm. This model comprises 100 decision trees, with a maximum depth of 6 for each tree to avoid overfitting the training data. The learning rate is set to 0.1, iteratively improving the model's prediction accuracy. The minimum sum of sample weights for each leaf node is set to 1 to ensure sufficient sample support for each leaf node. Before deployment, the model is pre-trained using engineering monitoring data and corresponding risk records from the past five years. By adjusting parameters such as the number and depth of the trees, the model accurately captures the correlation between the risk assessment vector and the actual risk level. The constructed risk assessment vector is input into the pre-trained gradient boosting tree model. Through ensemble calculations across multiple decision trees, the model outputs a value between 0 and 1, which is labeled as the real-time risk probability. A higher value indicates a greater safety risk to the current engineering structure.

[0066] The basic early warning threshold is determined based on the project's design standards, material properties, and historical safe operation data, with an initial value set at 0.6. The dynamic threshold adjustment model adjusts the basic early warning threshold based on real-time risk probability and the project's operational lifespan. A correction coefficient is introduced during the adjustment process, and this coefficient is dynamically updated with the project's operational lifespan. When the project's operational lifespan is 0-10 years, the overall structural performance is relatively stable, and the correction coefficient is set to 0.8; when the operational lifespan is 10-20 years, the structure may exhibit some degree of aging, and the correction coefficient is adjusted to 1.0; when the operational lifespan exceeds 20 years, the degree of structural aging deepens, and the correction coefficient increases to 1.2. The adjusted early warning threshold is obtained by multiplying the basic early warning threshold by the real-time risk probability, multiplying by the correction coefficient, and then adding this result back to the basic early warning threshold. For example, when the basic early warning threshold is 0.6, the real-time risk probability is 0.7, and the project's operational lifespan is 15 years (correction coefficient 1.0), the adjusted early warning threshold is 0.6 × 0.7 × 1.0 + 0.6, and the calculated result serves as the current early warning judgment standard.

[0067] Based on the adjusted warning threshold, the contribution weight of each monitoring indicator is compared with this threshold one by one, and indicators with a contribution weight greater than the warning threshold are selected. These indicators have a high influence in the current risk assessment, and their corresponding feature sequences are marked as key risk factors. For example, if the adjusted warning threshold is 0.5, the contribution weight of a vibration sensor is 0.6, and the contribution weight of another stress sensor is 0.4, then the vibration sensor is identified as a key risk factor, and its collected vibration acceleration dynamic feature sequence will be the focus of subsequent risk analysis. The key risk factors selected in this way can reflect the links in the current engineering structure most likely to cause safety problems, providing clear analytical objects for subsequent node correlation reasoning and graded warning.

[0068] Example 5: In step S5, topological relationships are extracted from the spatial location labels of key risk factors to construct a directed graph model of the engineering structure. The spatial location labels of key risk factors contain information about the specific engineering location of the sensors, such as the specific elevation of the dam, the specific slope section of the slope, and the specific nodes of the flood discharge facilities, along with precise coordinate data. Based on this information, the physical connection relationships and signal transmission paths between different key risk factors are determined, thereby constructing a directed graph model. For example, a bidirectional directed edge is set between the sensors at the bottom of the dam and those in the middle of the dam, since they belong to the same dam structure and have a stress transmission relationship; a unidirectional directed edge is set between the hydrological sensors at the inlet and outlet of the flood discharge channel, since the water flow direction is fixed; and a directed edge is set between the sensors at the top and bottom of the slope, considering the stress distribution and displacement transmission under gravity.

[0069] A graph attention network is used to rank the importance of nodes in a directed graph model and calculate the risk propagation probability of each node. The constructed engineering structure graph model includes node features and edge weights. Node features are normalized values ​​of key risk factors, obtained by mapping the raw data of key risk factors to a range of 0-1, referencing the maximum and minimum values ​​of this type of data in historical records. Edge weights are dynamically calculated based on sensor spatial distance and signal correlation. Spatial distance is calculated using sensor coordinate data, and signal correlation is obtained by analyzing the similarity of long-term monitoring data from two sensors. For example, two sensors that are close to each other and whose monitoring data show highly consistent trends have higher edge weights, while sensors that are far apart and whose data correlation is weaker have lower edge weights.

[0070] The graph attention network consists of two graph attention layers and one fully connected layer. The first graph attention layer uses a bidirectional gated recurrent unit (GRU) as the node update function. This function can handle long-term dependencies in sequential data and is suitable for fusing neighbor node information from different time points. During processing, each node collects feature information from its surrounding neighbors through a neighborhood aggregation mechanism, while simultaneously introducing an attention mechanism to weight the contributions of different neighbors. The calculation of the attention coefficient comprehensively considers node feature similarity and edge weights; neighbor nodes with high feature similarity and large edge weights are assigned higher attention coefficients, and vice versa. The second graph attention layer has the same structure as the first layer. By aggregating neighborhood information again, it further mines higher-order correlation features between nodes, enabling node features to reflect risk associations over a wider range. The node features output from the two graph attention layers are input into the fully connected layer, which contains one hidden layer and one output layer. The hidden layer performs a non-linear transformation on the node features, and the output layer maps the results to the range of 0-1 through an activation function, obtaining the risk propagation probability of each node. This probability reflects the likelihood of the risk of that node spreading to other nodes.

[0071] Based on the risk propagation probability, key risk factors are categorized into source nodes, intermediate nodes, and terminal nodes. Nodes with a risk propagation probability greater than 0.7 are classified as source nodes; these nodes have a strong risk diffusion capability and are the main sources of risk. Nodes with a risk propagation probability between 0.3 and 0.7 are classified as intermediate nodes; these nodes can both receive risk propagation from other nodes and transmit risk to other nodes, acting as risk relays. Nodes with a risk propagation probability less than 0.3 are classified as terminal nodes; these nodes have a weak risk diffusion capability and primarily receive risk propagation from other nodes.

[0072] Time series alignment of risk signals from different levels of nodes is performed. A dynamic time warping algorithm adjusts the time axis of risk signals from different nodes to eliminate time offsets caused by signal acquisition delays or transmission speed differences, enabling direct comparison of the risk status of each node at the same point in time. Then, multi-source information fusion is achieved through attention weight allocation, assigning different weights to risk signals from different levels of nodes: higher weights for source nodes, lower weights for intermediate nodes, and lower weights for terminal nodes. The fused information generates a tiered early warning result, which includes risk level, impact range, and warning timeliness. The risk level is determined based on the fused risk value, the impact range is determined based on the spatial distribution of key risk factors and risk propagation paths, and the warning timeliness is determined by referring to the risk propagation probability and the development speed of similar risks in historical data, ultimately forming a comprehensive and detailed tiered early warning result.

[0073] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for safety detection and early warning of water conservancy projects based on artificial intelligence, characterized in that, Including the following steps: S1. Collect multimodal monitoring data of key parts of water conservancy projects through a distributed sensor network, perform spatiotemporal alignment and noise filtering on the raw data, and extract dynamic feature sequences within a preset time period; S2. Input the dynamic feature sequence into a deep neural network with an attention mechanism, and construct a multi-scale spatiotemporal correlation map through hierarchical feature learning to generate a high-dimensional representation of the state of the engineering structure. S3. Using historical accident case data as a supervision signal, a hybrid expert model is used to train the high-dimensional representation in multiple tasks to obtain the contribution weight of each monitoring indicator to safety risk; S4. Combining real-time environmental parameters and structural response characteristics, construct a dynamic threshold adjustment model, adaptively update the early warning threshold based on the risk probability output by the model, and screen out key risk factors whose contribution weight is greater than the updated threshold. S5. Based on the spatiotemporal distribution characteristics of key risk factors, a graph neural network is used to perform node association reasoning and integrate multi-source early warning information to generate graded early warning results.

2. The method for safety detection and early warning of water conservancy projects based on artificial intelligence according to claim 1, characterized in that, Step S1 includes the following: A distributed network of vibration sensors, stress sensors, and hydrological sensors deployed on the dam body, slopes, and flood discharge facilities is used to synchronously collect vibration acceleration, structural strain, and water level and flow velocity data. The collected raw data is timestamped and outliers are removed. A sliding time window is used to divide the continuous data into segments, and the time domain statistics and frequency domain characteristics within each window are calculated. Time-domain statistics and frequency-domain features are categorized and integrated according to sensor type to form a dynamic feature sequence containing spatial location labels.

3. The method for safety detection and early warning of water conservancy projects based on artificial intelligence according to claim 2, characterized in that, Step S2 includes the following: A deep neural network architecture integrating spatial attention and channel attention is constructed, comprising a feature encoding layer and an association inference layer. The feature encoding layer uses a convolutional neural network to extract local spatiotemporal features, strengthens the feature weights of key parts through a spatial attention module, and suppresses redundant signal interference through a channel attention module. The association inference layer models the topological relationship between different sensor nodes through a graph convolutional network, dynamically adjusts the node connection strength using an attention mechanism, and fuses the encoded local features with global association features to generate a high-dimensional representation of the engineering structure state.

4. The method for safety detection and early warning of water conservancy projects based on artificial intelligence according to claim 3, characterized in that, Step S3 includes the following: We collect monitoring data and failure mode labels from historical engineering accident cases and classify them into task types such as structural cracks, abnormal seepage, and slope instability. We construct a hybrid expert model, which includes multiple sub-expert networks and a gating network. The sub-expert networks are trained for different task types, and the gating network dynamically allocates task weights based on the feature distribution of the input high-dimensional representation. The model parameters are optimized using the probability of accident occurrence as the loss function. The gradient contribution of each monitoring indicator in different tasks is calculated through backpropagation, and the safety risk contribution weight of each indicator is obtained after normalization.

5. The method for safety detection and early warning of water conservancy projects based on artificial intelligence according to claim 4, characterized in that, Step S4 includes the following: Real-time environmental parameters include rainfall, temperature change rate, and upstream and downstream water level difference; structural response characteristics include vibration dominant frequency offset and strain accumulation rate.

6. The method for safety detection and early warning of water conservancy projects based on artificial intelligence according to claim 5, characterized in that, Step S4 also includes the following: To quantify the dynamic impact of environmental factors on engineering structures, the time-varying curve of rainfall and the gradient characteristics of temperature change rate are first extracted from real-time environmental parameters. A long short-term memory network is used to predict the trend of the time-varying curve and calculate the predicted rainfall value for the next 24 hours. The predicted value is compared with historical data for the same period and converted into an environmental risk coefficient through a membership function. Combined with the gradient characteristics of temperature change rate, a comprehensive environmental impact index is generated by weighted summation.

7. The method for safety detection and early warning of water conservancy projects based on artificial intelligence according to claim 5, characterized in that, Step S4 also includes the following: To characterize the degree of anomaly in the structural response, the power spectral density of the vibration signal and the cumulative curve of the strain data are first extracted from the structural response characteristics. Wavelet transform is used to perform time-frequency decomposition of the power spectral density to calculate the offset of the dominant frequency component. The cumulative curve is linearly fitted, and the strain accumulation rate is obtained through residual analysis. After normalizing the dominant frequency offset and the strain accumulation rate, the comprehensive anomaly value of the structural response is calculated using Euclidean distance.

8. The method for safety detection and early warning of water conservancy projects based on artificial intelligence according to claim 5, characterized in that, Step S4 also includes the following: The comprehensive environmental impact index, the comprehensive structural response anomaly value, and the contribution weights of each monitoring indicator are combined into a risk assessment vector, which is then input into a pre-trained gradient boosting tree model. The model output is then labeled as the real-time risk probability. The product of the basic early warning threshold and the real-time risk probability is adjusted by a correction factor and then added together, where the correction factor is dynamically updated according to the years of operation of the project. Based on the adjusted warning threshold, indicators that exceed the warning threshold are selected from the contribution weights of each monitoring indicator, and their corresponding feature sequences are marked as key risk factors.

9. The method for safety detection and early warning of water conservancy projects based on artificial intelligence according to claim 8, characterized in that, Step S5 includes the following: Topological relationships are extracted from the spatial location labels of key risk factors to construct a directed graph model of the engineering structure. A graph attention network is used to rank the importance of nodes in the directed graph model and calculate the risk propagation probability of each node. Based on the risk propagation probability, key risk factors are divided into source nodes, intermediate nodes, and terminal nodes. The risk signals of nodes at different levels are aligned in time series, and multi-source information is fused through attention weight allocation to generate a graded early warning result that includes risk level, impact range, and early warning timeliness.

10. The method for safety detection and early warning of water conservancy projects based on artificial intelligence according to claim 9, characterized in that, The step of using a graph attention network to rank the importance of nodes in a directed graph model and calculating the risk propagation probability of each node includes: An engineering structure graph model is constructed, which includes node features and edge weights. The node features are normalized values ​​of key risk factors, and the edge weights are dynamically calculated based on sensor spatial distance and signal correlation. A bidirectional gated recurrent unit is used as the node update function of the graph neural network, and neighbor node information is fused through a neighborhood aggregation mechanism. An attention mechanism is introduced to weight the contributions of different neighboring nodes, and high-order correlation features of nodes are learned through multi-layer graph convolution operations. The learned node features are input into a fully connected layer, and the risk propagation probability of each node is output.

Images