Safety early warning system for water conservancy construction site

Abstract

The invention discloses a water conservancy construction site safety early warning system which comprises the components of S1, a multi-source heterogeneous sensor module which is used for outputting an original data stream with a timestamp; s2, an edge processing module which is used for generating a time sequence feature vector sequence; s3, a graph modeling module which is used for mapping each node in the weighted graph into a low-dimensional fusion feature graph; s4, a risk identification module which is used for realizing risk level inference through side attention weight adjustment; s5, a self-adaptive early warning module which is used for outputting a multi-level early warning signal according to a risk level inference result; s6, a response module which is used for controlling an audible and visual alarm, a field terminal and a remote platform to execute response operation according to the multi-stage early warning signal; and S7, a cooperative control module which is used for realizing data acquisition interval adjustment through sliding mean deviation control. The method has the advantages of being comprehensive in sensing dimension, low in response delay and high in risk adaptability.

Description

Technical Field

[0001] This invention relates to the field of safety early warning technology, and in particular to a safety early warning system for water conservancy construction sites. Background Technology

[0002] With the acceleration of infrastructure intelligence, water conservancy construction scenarios are gradually evolving into complex systems involving multiple regions and disciplines. Given the uncertainties and dynamic nature of the water conservancy construction site environment, how to achieve early identification and dynamic intervention of safety risks in damp, vibrating, and high-risk working environments has become a key factor restricting project progress and personnel safety.

[0003] In existing technologies, safety management at water conservancy construction sites mainly relies on single sensors or manual inspections for environmental monitoring and hazard identification. Common methods include fixed-point video surveillance, environmental parameter collection, and structural load monitoring. However, these data exhibit strong heterogeneity and low correlation, making it difficult to construct a unified information fusion model. This results in insufficient dimensions for risk factor identification and ineffective utilization of data redundancy. Furthermore, existing early warning mechanisms often rely on a central server as the core node for data processing and risk assessment. Limited by unstable communication conditions and uneven distribution of computing resources in construction areas, remote central processing models suffer from response lag and delayed risk identification, failing to meet the high-timeliness requirements of construction safety management scenarios.

[0004] Regarding the fusion and modeling of multi-source monitoring data, some existing methods have initially introduced strategies such as multimodal data analysis and graph network modeling. However, most of them still rely on static graph structures and preset parameters to construct risk models, making it difficult to incorporate the dynamic changes in the spatiotemporal relationships between nodes during actual construction. Furthermore, they fail to incorporate local inference with edge nodes, resulting in limited real-time performance and availability of the overall system. In addition, current early warning methods mostly use fixed thresholds or expert experience rules to set levels, lacking the ability to adaptively adjust based on actual construction conditions. This makes it difficult to cope with the dynamic evolution of risk factors in different stages and regions during the construction cycle, leading to problems such as generalized early warning levels and homogenized response measures.

[0005] In summary, there is an urgent need for a safety early warning system that integrates multi-source heterogeneous sensing and edge intelligence for water conservancy construction sites. This system should be able to achieve high-precision modeling of multi-dimensional risk factors, dynamic response adjustment to time-varying communication and drift states, and adaptive risk classification in response to changes in the construction environment. This would solve the bottleneck problems of existing methods in terms of data fusion, response timeliness, and risk identification. Summary of the Invention

[0006] One objective of this invention is to propose a safety early warning system for water conservancy construction sites based on multi-source heterogeneous perception fusion and edge intelligent collaboration. This invention integrates data from multiple types of sensors, including video surveillance, environmental perception, structural status, and personnel positioning, to construct a cross-modal collaborative mapping and graph convolution risk identification mechanism. By deploying edge nodes, it achieves localized risk modeling and dynamic response control at the construction site. Combined with real-time working condition-driven attention adjustment and multi-level early warning discrimination logic, it enables efficient identification and hierarchical early warning control of risk factors in complex construction environments. It has the advantages of comprehensive perception dimensions, low response latency, and strong risk adaptability.

[0007] A safety early warning system for water conservancy construction sites according to an embodiment of the present invention includes: S1, Multi-source heterogeneous sensor module, used to collect visible light monitoring images, environmental parameters, structural status and personnel location information, and encapsulate the above-mentioned sensing data into a raw data stream with timestamps and output it; S2, Edge processing module, used to perform synchronization and standardization processing on the original data stream based on a sliding time window to generate a time-series feature vector sequence; S3, Graph Modeling Module, is used to calculate the covariance matrix between time-series features based on the time-series feature vector sequence, construct a weighted graph by combining the spatial distance of each sensor, and use the spectral embedding method to map each node in the weighted graph into a low-dimensional fusion feature map; S4, Risk Identification Module, is used to input low-dimensional fused feature maps into a graph convolutional network with an attention gating mechanism, and to infer risk level by adjusting edge attention weights; S5, Adaptive early warning module, is used to dynamically adjust the edge attention weights of the graph convolutional network based on real-time construction status indicators, and output multi-level early warning signals based on the risk level inference results; S6, Response module, used to control the audible and visual alarm, field terminal and remote platform to perform response operations according to the multi-level early warning signals; S7, the collaborative control module, is used to monitor the communication delay and data drift between each sensor and the edge processing module, and to adjust the data acquisition interval through sliding mean deviation control.

[0008] Optionally, S3 specifically includes: S31, Covariance Calculation Submodule, is used to construct a feature matrix for each type of sensor data based on the temporal feature vector sequence output by the edge processing module. ,in, Indicates the length of the time window. Representing feature dimension, Represent the real number field and compute its covariance matrix. ; S32, Spatial Distance Calculation Submodule, used to calculate the Euclidean distance between any two sensor nodes based on the spatial deployment location of each sensor on the construction site. and Spatial distance And generate the spatial distance matrix between nodes. ,in, This represents the total number of sensor nodes. Indicates the first The and the first The physical distance between the sensors ; S33, Weighted graph construction submodule, used to fuse the covariance matrix. Spatial distance matrix Construct a weighted adjacency matrix The fusion method is based on the following function:

[0009] in, Indicates the first The and the first The connection weights between nodes To map the covariance in the feature space to the similarity measure of node pairs, , , The fusion coefficient is... Represents an exponential function; S34, Graph embedding generation submodule, used to generate graphs based on the weighted adjacency matrix. and the feature vector corresponding to each sensor node ,in, For nodes In the characteristic matrix The eigenvectors in the graph are used to construct the Laplacian matrix of the weighted graph. ,in, For the degree matrix, and for Perform feature decomposition, before extraction The eigenvectors corresponding to the smallest eigenvalues ​​are used as low-dimensional embedding representations of the nodes to form a low-dimensional fused feature map.

[0010] Optionally, S4 specifically includes: S41, Graph Node Initialization Submodule, is used to receive the low-dimensional fused feature map output by the graph modeling module and initialize the embedding vector of each graph node. The corresponding node type labels and sensor category codes are concatenated to form an enhanced feature representation. ,in, ; S42, Multi-head attention weight calculation submodule, used to calculate weights based on each head during graph convolution. Calculate adjacent nodes and Attention coefficient The calculation method is as follows:

[0011] in, For the number of attention heads, Represents a non-linear activation function. Indicates the first in the figure Enhanced feature representation of each node, Indicates the previous number The node's Enhanced feature representation of each neighboring node, Represents the set of adjacent nodes of a node. Indicates attention head The weight vector Indicates attention head The linear transformation matrix; S43, Risk Feature Aggregation Submodule, used to perform multi-scale aggregation of each neighbor node based on the multi-head attention coefficient to obtain the node update vector. The vector is then fused with the original vector through residual connections and input into the next convolutional layer. S44, Risk Level Determination Submodule, is used to input the node vector output by the multi-layer graph convolution into the fused construction environment state vector, and output the risk level label based on the risk scenario mapping function. ,in, This indicates the total number of risk level categories.

[0012] Optionally, S44 specifically includes: S441, the classification input construction submodule, is used to receive the updated representation vector of each node output by the graph convolution aggregation submodule. And obtain the temporal feature vector of the current construction environment parameters in the edge processing module. The two are concatenated to form an extended classification input vector. ,in, , Indicates the output dimension of the convolution. Represents the dimension of state features. This represents a vector concatenation operation; S442, Risk Level Classification Submodule, used to preset the risk level set. ,in, This indicates the total number of risk level categories, with each category corresponding to a different level of construction risk. S443, Discriminant network structure submodule, used to construct a classification neural network composed of fully connected layers, and to receive the extended classification input vector. Intermediate representations are generated by sequentially applying a set of linear transformations and nonlinear activation functions. The output layer uses the softmax activation function to calculate the probability distribution of each risk level. ,in,:

[0013] in, Indicates the output layer number Weight vector of risk level; S444, Risk Label Output Submodule, used to determine the node based on the category with the highest probability in the softmax output as the final classification result. The corresponding risk level label .

[0014] Optionally, S5 specifically includes: S51, Construction Status Acquisition Submodule, used to obtain a set of real-time construction status indicators for the current construction area from the time-series feature vector sequence. ,in, Indicates the first Item status indicators, Represents the dimension of state features; S52, Edge Weight Adjustment Submodule, is used to input the set of construction status indicators into the edge weight dynamic mapping function. And the attention coefficient corresponding to each edge in the graph convolutional network. Make adjustments; S53, Dynamic Multi-Level Discrimination Submodule, is used to determine the risk level label of each node in the final output of the graph convolutional network. and its corresponding softmax probability distribution Combined with the dynamic early warning strategy table specified in the construction status Determine the node Warning level ,in, Indicates the total number of warning level categories; S54, Warning Signal Output Submodule, used to output warning signals based on the warning level of each node. Generate corresponding graded early warning signals and transmit the signals to the response module.

[0015] Optionally, S7 specifically includes: S71, Communication Status Monitoring Submodule, is used to record the communication delay value between each sensor node and the edge processing module in each acquisition cycle. ,in, , Indicates the sampling time index. Indicates the total number of sensors; S72, Data Drift Estimation Submodule, is used to extract the numerical changes in data collected by each sensor in adjacent sampling periods and construct a drift sequence. ,in, For the first Each node at time... eigenvectors, For feature dimension, Represents the vector norm; S73, Sliding mean deviation calculation submodule, used to calculate the deviation within a set sliding window length. Within this context, calculate the moving average deviation of the communication delay and data drift for each node, assuming the th node... Each node at time step The communication delay value is Then their moving averages are defined as follows:

[0016] in, Represents a node In time forward The average communication delay within each time step Indicates the first Each node at time step The communication delay value at any given time. Represents a node In time forward The mean drift over each time step For sliding index, Represents a node At time step The amount of drift within a given time period; S74, Data Interval Adjustment Submodule, is used to construct the sampling interval update function. Update the result based on the current mean deviation. Sampling period of each sensor The calculation method is as follows:

[0017] in, For adjustment coefficients, The joint deviation function of delay and drift depends on and .

[0018] The beneficial effects of this invention are: (1) This invention integrates heterogeneous sensor data such as images, environment, structure and positioning to construct a unified temporal feature vector and spatial covariance matrix expression framework, and constructs a weighted graph structure based on the physical distance between nodes, thereby realizing spatiotemporal collaborative modeling of multi-source sensor data in the graph model. This breaks through the problem of single dimension of risk factor expression and difficulty in data integration under the traditional homogeneous sensor mode, and improves the system's overall perception ability of multi-dimensional risk factors in complex construction scenarios.

[0019] (2) This invention introduces a graph convolutional network structure with attention gating mechanism in edge computing nodes, realizes adaptive modeling of feature differences between nodes through multi-head attention weighting, and dynamically adjusts the weight distribution in the graph structure based on the edge attention adjustment function driven by construction status, thereby realizing localized modeling and timely identification of on-site risks, effectively avoiding the problem of risk inference lag caused by network delay and data congestion in traditional cloud processing methods, and improving the distributed intelligent response capability of the construction area.

[0020] (3) Based on the results of multi-level risk classification, this invention constructs a dynamic early warning strategy table by combining real-time construction indicators, and introduces a sliding mean deviation mechanism to periodically evaluate the data drift and communication delay of each sensor. The sampling interval is adaptively adjusted by the joint deviation function, and the triggering frequency of the early warning signal and the node feedback rhythm are controlled at the same time. The system has high robustness, low resource dependence and self-calibration capability, effectively breaking through the problem of the existing system's serious dependence on static thresholds and manual rules. The safety assurance stability in complex water conservancy construction scenarios is significantly enhanced. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a safety early warning system for water conservancy construction sites proposed in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figure 1 A safety early warning system for water conservancy construction sites, the system comprising: S1, Multi-source heterogeneous sensor module, used to collect visible light monitoring images, environmental parameters, structural status and personnel location information, and encapsulate the above-mentioned sensing data into a raw data stream with timestamps and output it; In this embodiment, the multi-source heterogeneous sensor module adopts a multi-channel parallel acquisition architecture to access and acquire data from various types of monitoring equipment (sensors) at the water conservancy construction site in real time. The sensor types include, but are not limited to, video surveillance equipment, environmental parameter sensors, structural condition monitoring instruments, and personnel positioning tag systems. To achieve timeliness consistency and format compatibility of data from different types of sensors, the module deploys a unified acquisition gateway and a timestamp synchronization mechanism, and configures independent cache queues, parsing templates, and sampling frequency parameters for each type of sensor acquisition channel.

[0024] Among them, image sensors collect video frame stream data and label them at the frame level with frame timestamps; environmental sensors collect real-time environmental parameters such as temperature, humidity, concentration of harmful gases, and wind speed, and output structured numerical streams at preset frequencies; structural condition monitoring equipment collects state variables such as displacement, vibration, and stress of concrete structures or support components, and constructs a two-dimensional feature array with sensor identifiers and sampling times as keys; the personnel positioning module outputs the spatial coordinates and identification of construction personnel in real time based on UWB or GPS, along with tag numbers and sampling timestamps.

[0025] All collected data is converted into a timestamp-driven raw data stream format via a unified data encapsulation interface. Each data record contains sensor type, sampling time, spatial location, raw value vector, and device number information, forming input data with a parsable structure and unified timing sequence. This provides stable and complete basic sensing input support for data fusion and mapping analysis in subsequent edge processing modules.

[0026] S2, Edge processing module, used to perform synchronization and standardization processing on the original data stream based on a sliding time window to generate a time-series feature vector sequence; In this embodiment, the edge processing module is deployed on edge computing nodes at the water conservancy construction site. It is used to perform synchronous preprocessing and standardized feature extraction operations within a sliding time window on the raw data stream output by the multi-source heterogeneous sensor module. To improve the temporal alignment and spatial representation consistency among different types of data, the edge processing module constructs a time window mechanism with a fixed step size and performs buffer reordering and delay tolerance matching for each type of data stream according to the timestamp, ensuring a unified temporal basis for data input.

[0027] In its implementation, the edge processing module extracts frame-level statistical features (such as target quantity, average grayscale, and edge density) from image stream data. For environmental parameter streams, it uses normalization and trend calculation methods to generate moving averages and fluctuation features. For structural state data, it performs denoising and first-order derivative transformation to form dynamic response feature vectors. For personnel positioning data, it generates behavior vectors based on trajectory length and dwell time. All processing steps use a unified time window as the processing unit, reconstructing different types of data streams into aligned feature representations within the same time period.

[0028] The edge processing module presets standardized parameter templates for each type of sensor and uses processing strategies such as Z-score standardization and min-max normalization to perform unified dimension mapping on the feature vectors, outputting a sequence of temporal feature vectors with the same time granularity and fixed dimension structure.

[0029] S3, Graph Modeling Module, is used to calculate the covariance matrix between time-series features based on the time-series feature vector sequence, construct a weighted graph by combining the spatial distance of each sensor, and use the spectral embedding method to map each node in the weighted graph into a low-dimensional fusion feature map; S4, Risk Identification Module, is used to input low-dimensional fused feature maps into a graph convolutional network with an attention gating mechanism, and to infer risk level by adjusting edge attention weights; S5, Adaptive early warning module, is used to dynamically adjust the edge attention weights of the graph convolutional network based on real-time construction status indicators, and output multi-level early warning signals based on the risk level inference results; S6, Response module, used to control the audible and visual alarm, field terminal and remote platform to perform response operations according to the multi-level early warning signals; In this embodiment, the response module is used to perform hierarchical control and issue linkage commands to various types of response devices at the construction site based on the multi-level early warning signals generated by the adaptive early warning module. The response devices include, but are not limited to, audible and visual alarms, visual display terminals, mobile operation panels, and remote command platforms. In order to realize dynamic mapping and rapid linkage between early warning levels and response behaviors, the response module constructs an event-driven execution framework based on a rule mapping table and deploys local response agents according to a node distribution structure.

[0030] In practical applications, the system divides the early warning level into several risk levels. The response module generates a response instruction set, including the response object, execution method, and duration, based on the response strategy parameters corresponding to each level. The corresponding mapping structure is stored locally on the edge node in the form of a response rule table, and supports remote updates as the construction scenario is configured.

[0031] When a multi-level warning signal is triggered, the response module first synchronizes the warning instructions corresponding to the medium- and high-level risk nodes to the local audible and visual alarm control interface, and activates the buzzer and warning light signals of the corresponding intensity. At the same time, the risk location and level information are graphically marked on the on-site operation terminal to prompt personnel to enter the control state. If the level exceeds the set threshold, the system uploads the risk node data and scene status to the remote management platform through the communication interface, and links the emergency plan module to push dispatch suggestions, initiate task issuance and remote broadcasting, etc.

[0032] S7, the collaborative control module, is used to monitor the communication delay and data drift between each sensor and the edge processing module, and to adjust the data acquisition interval through sliding mean deviation control.

[0033] This implementation integrates multiple types of sensors, including image monitoring, environmental perception, structural status monitoring, and personnel positioning, to construct a unified data acquisition framework. This enables comprehensive perception of multi-dimensional risk factors at water conservancy construction sites. By employing a raw data stream format with timestamps, it effectively solves the problem of unifying the temporal representation and structural compatibility of heterogeneous sensor data. Through fine-tuning the caching mechanism and frequency parameters of various acquisition channels, the system achieves asynchronous acquisition and synchronous alignment of data. While improving the perception coverage, it ensures the coordination and consistency of multiple types of data on the time axis, providing a structurally stable and temporally complete basic data support for subsequent data fusion, graph modeling, and risk identification. This enhances the system's adaptability to complex construction scenarios and its data expression capabilities.

[0034] In this embodiment, S3 specifically includes: S31, Covariance Calculation Submodule, is used to construct a feature matrix for each type of sensor data based on the temporal feature vector sequence output by the edge processing module. ,in, Indicates the length of the time window. Representing feature dimension, Represent the real number field and compute its covariance matrix. ; The covariance calculation submodule constructs a feature matrix based on the temporal feature vector sequence output by the edge processing module. ,in, Indicates the length of the time window. Representing feature dimension, Representing the real number field, each row of the matrix corresponds to a multidimensional feature observation at one time step. To characterize the statistical dependencies between features, the module... Perform column mean normalization and calculate the zero-mean matrix. And construct the feature covariance matrix based on the following formula. :

[0035] in, This represents the mean vector for each column, used for bias removal. If the original feature data has been standardized (e.g., by Z-score), the covariance matrix is ​​transformed into a dimensionless form. The final generated covariance matrix... As a structural input to the graph modeling module, it supports subsequent graph edge weight construction and node similarity modeling operations, meeting the requirements of graph structure learning for numerical stability and relevance expression.

[0036] S32, Spatial Distance Calculation Submodule, used to calculate the Euclidean distance between any two sensor nodes based on the spatial deployment location of each sensor on the construction site. and Spatial distance And generate the spatial distance matrix between nodes. ,in, This represents the total number of sensor nodes. Indicates the first The and the first The physical distance between the sensors ; S33, Weighted graph construction submodule, used to fuse the covariance matrix. Spatial distance matrix Construct a weighted adjacency matrix The fusion method is based on the following function:

[0037] in, Indicates the first The and the first The connection weights between nodes To map the covariance in the feature space to the similarity measure of node pairs, , , The fusion coefficient is... Represents an exponential function; The weighted graph construction submodule is based on the covariance matrix. Spatial distance matrix Perform joint fusion to generate a weighted adjacency matrix between nodes. ,in, This represents the total number of sensor nodes. The feature dimension is used; the fusion method adopts a linear-exponential mixture function, as shown in the following expression:

[0038] in, Indicates the first The and the first The connection weights between nodes; The similarity measure that maps the covariance matrix to the node pairs is derived from the statistical correlation modeling of the feature layer. The dimensions are usually dimensionless after feature normalization. Represents a node and The physical distance between them is usually expressed in meters (m) or centimeters (cm), depending on the site layout. It is a natural exponential function used to model the decay effect of spatial distance on the probability of edge connection; This represents the spatial distance attenuation coefficient, and its unit is . This is used to adjust the strength of the influence of the spatial factor on the connection weights; This is a dimensionless fusion scaling factor used to balance the contribution weights of feature similarity and spatial distance terms. This fusion function introduces spatial topological constraints while ensuring the dominance of feature relevance, achieving joint spatial-feature modeling of edge weights between nodes. It provides a flexible and adjustable weighting strategy for graph structure construction, meeting the modeling requirements of graph neural networks for the differences in edge weight representation and structural distribution.

[0039] S34, Graph embedding generation submodule, used to generate graphs based on the weighted adjacency matrix. and the feature vector corresponding to each sensor node ,in, For nodes In the characteristic matrix The eigenvectors in the graph are used to construct the Laplacian matrix of the weighted graph. ,in, For the degree matrix, and for Perform feature decomposition, before extraction The eigenvectors corresponding to the smallest eigenvalues ​​are used as low-dimensional embedding representations of the nodes to form a low-dimensional fused feature map.

[0040] This implementation constructs a covariance matrix based on time-series features and a spatial distance matrix between sensor nodes to model and measure spatiotemporal two-dimensional feature relationships. A dynamic weighted graph structure is generated using a linear weighting and exponential decay fusion function, effectively characterizing the feature correlations and spatial dependencies among heterogeneous sensors at the construction site. Furthermore, the system introduces a spectral embedding method to perform feature decomposition on the weighted graph's Laplacian matrix, extracting the dominant structure and local variability features in the graph. This compresses and maps high-dimensional sensor data into a low-dimensional embedded representation, forming a stable and structured fusion feature graph. While retaining key correlation information, this significantly improves data modeling efficiency and representational compactness, providing an accurate and topology-aware input foundation for subsequent risk identification using graph convolutional networks.

[0041] In this embodiment, S4 specifically includes: S41, Graph Node Initialization Submodule, is used to receive the low-dimensional fused feature map output by the graph modeling module and initialize the embedding vector of each graph node. The corresponding node type labels and sensor category codes are concatenated to form an enhanced feature representation. ,in, ; S42, Multi-head attention weight calculation submodule, used to calculate weights based on each head during graph convolution. Calculate adjacent nodes and Attention coefficient The calculation method is as follows:

[0042] in, For the number of attention heads, Represents a non-linear activation function. Indicates the first in the figure Enhanced feature representation of each node, Indicates the previous number The node's Enhanced feature representation of each neighboring node, Represents the set of adjacent nodes of a node. Indicates attention head The weight vector Indicates attention head The linear transformation matrix; The multi-head attention weight calculation submodule constructs a multi-head attention weighting mechanism based on the structural characteristics of graph neural networks. It uses a scoring function with leaky activation to calculate the attention coefficients between adjacent node pairs. Its neighboring nodes Between the first Attention weights under each attention head The calculation formula is as follows:

[0043] in, Represents a node The input feature vector, For the first The transformation matrix of each attention head embeds and maps the original input into the intermediate representation space; For the first The weight vector corresponding to each head is applied to the concatenated neighbor node pair representation. This represents a vector concatenation operation; the activation function LeakyReLU is a nonlinear mapping operation with leakage correction, and the exponential function... Used to normalize the score; final attention coefficient It has dimensionless properties and is used to represent nodes. For nodes This indicates the importance of the update; all weight resides in the set of adjacent nodes. Internal normalization gives the attention distribution a probabilistic constraint characteristic, which meets the modeling requirement of dynamically adjustable feature aggregation weights during graph convolution.

[0044] S44, Risk Level Determination Submodule, is used to input the node vector output by the multi-layer graph convolution into the fused construction environment state vector, and output the risk level label based on the risk scenario mapping function. ,in, This indicates the total number of risk level categories.

[0045] S44, Risk Level Determination Submodule, is used to input the node vectors output by the multi-layer graph convolution into a discrimination network that integrates the construction environment state vector and stage information, and outputs risk level labels based on the risk scenario mapping function. ,in, This indicates the total number of risk level categories.

[0046] This implementation enhances feature representation by introducing node type and sensor category encoding into the low-dimensional embedding vectors output by the graph modeling module, giving the original graph structure stronger semantic discriminative ability. A multi-head attention mechanism is introduced during graph convolution to dynamically adjust the edge weight distribution between adjacent nodes, strengthening the directionality of feature aggregation and the adaptability of the receptive domain. During the aggregation stage, residual connections preserve the original embedding information, avoiding oversmoothing in the deep propagation process of the graph neural network. Finally, a discriminant network based on the fusion of construction status and stage features outputs risk level labels. The system achieves a structure-aware, state-driven risk classification path, effectively improving the accuracy of multi-dimensional risk factor identification and the generalization ability of graph modeling in complex water conservancy construction scenarios. This provides a learnable and scalable intelligent recognition foundation for early warning grading and response strategies.

[0047] In this embodiment, S44 specifically includes: S441, the classification input construction submodule, is used to receive the updated representation vector of each node output by the graph convolution aggregation submodule. And obtain the temporal feature vector of the current construction environment parameters in the edge processing module. The two are concatenated to form an extended classification input vector. ,in, , Indicates the output dimension of the convolution. Represents the dimension of state features. This represents a vector concatenation operation; S442, Risk Level Classification Submodule, used to preset the risk level set. ,in, This indicates the total number of risk level categories, with each category corresponding to a different level of construction risk. S443, Discriminant network structure submodule, used to construct a classification neural network composed of fully connected layers, and to receive the extended classification input vector. Intermediate representations are generated by sequentially applying a set of linear transformations and nonlinear activation functions. The output layer uses the softmax activation function to calculate the probability distribution of each risk level. ,in,:

[0048] in, Indicates the output layer number Weight vector of risk level; The discriminant network structure submodule receives the expanded node input vector. The hidden layer intermediate representations are then formed sequentially through linear transformations and activation functions. ,in, This represents the intermediate feature dimension. The output layer uses the softmax activation function to compute nodes. Probability distributions belonging to each risk level The specific calculation formula is as follows:

[0049] in, Indicates the total number of risk level categories. For the output layer The weight vector of risk level, in dimensionless form, represents the linear mapping parameters to the intermediate representation; For nodes The discriminative feature vector, the input dimension and Matching; the softmax function guarantees the probability values ​​corresponding to all risk levels. And it satisfies the normalization constraint. This mechanism enables probabilistic discriminative modeling of multi-category labels in risk classification tasks, meeting the system's stability and interpretability requirements for identifying multiple risk levels in complex construction environments.

[0050] S444, Risk Label Output Submodule, used to determine the node based on the category with the highest probability in the softmax output as the final classification result. The corresponding risk level label .

[0051] This implementation constructs a dynamic context-aware classification input representation by concatenating the structural risk features output by graph convolution with the current state vector of the construction site, enhancing the sensitivity of the risk discrimination process to environmental changes. A configurable multi-level risk set is introduced for risk level classification, enabling the system to flexibly map according to actual engineering safety standards. The discrimination network adopts a deep classification structure composed of fully connected layers, outputting multi-class probability distributions through a softmax layer, possessing excellent nonlinear discrimination and risk level differentiation capabilities. Finally, risk labels are output based on the maximum probability principle, realizing a risk identification scheme that combines state-driven, structure-aware, and level-refined approaches, effectively improving the intelligent classification and accurate level determination capabilities for construction safety hazards under complex working conditions.

[0052] In this embodiment, S5 specifically includes: S51, Construction Status Acquisition Submodule, used to obtain a set of real-time construction status indicators for the current construction area from the time-series feature vector sequence. ,in, Indicates the first Item status indicators, Represents the dimension of state features; S52, Edge Weight Adjustment Submodule, is used to input the set of construction status indicators into the edge weight dynamic mapping function. And the attention coefficient corresponding to each edge in the graph convolutional network. Make adjustments; The edge weight adjustment submodule receives the set of construction status indicators. And construct a state-driven edge weight adjustment function. Used to sense nodes in the graph , The degree of difference between the current construction states; based on this, the edges in the graph convolutional network... In the Attention weights under each attention head After adjustments, the edge attention coefficients after state correction are obtained. Its expression is as follows:

[0053] in, Represents the original attention weights, derived from the first element in the graph structure. The local perception mechanism of the head has dimensionless properties; Represents a node , The state vector, whose units may include, depending on the definition of the construction state index; As a dimensionless function, the output value is usually normalized to a finite interval (e.g., ); The state adjustment coefficient controls the intensity of the influence of state differences on edge weight adjustment; the final result is... The adjusted edge weight coefficients, representing the differences between the fusion structure relationships and states, preserve attention normalization and support the adaptive propagation operation of subsequent graph neural networks, thereby enhancing the system's response sensitivity to dynamic construction environments and its edge weight control capabilities.

[0054] S53, Dynamic Multi-Level Discrimination Submodule, is used to determine the risk level label of each node in the final output of the graph convolutional network. and its corresponding softmax probability distribution Combined with the dynamic early warning strategy table specified in the construction status Determine the node Warning level ,in, Indicates the total number of warning level categories; S54, Warning Signal Output Submodule, used to output warning signals based on the warning level of each node. Generate corresponding graded early warning signals and transmit the signals to the response module.

[0055] This implementation introduces real-time status indicators from the construction site as driving factors to construct a state-aware dynamic adjustment mechanism for edge weights. This enables the graph convolutional network to adjust the influence intensity between adjacent nodes based on actual working conditions during risk propagation, enhancing the model's ability to flexibly control risk propagation paths. Combined with soft classification output and a dynamic early warning strategy table, the system achieves adaptive discrimination of multi-level early warning levels, overcoming the limitations of fixed threshold rules that are difficult to adapt to the evolution of the construction cycle. Simultaneously, by controlling the graded response of each node's early warning level, it outputs structured and timely early warning signals, effectively supporting the differentiated processing logic of subsequent response modules. Overall, this improves the system's risk perception depth and early warning decision accuracy in complex environments, providing a highly flexible and responsive key support capability for intelligent safety management during water conservancy construction.

[0056] In this embodiment, S5 specifically includes: S71, Communication Status Monitoring Submodule, is used to record the communication delay value between each sensor node and the edge processing module in each acquisition cycle. ,in, , Indicates the sampling time index. Indicates the total number of sensors; S72, Data Drift Estimation Submodule, is used to extract the numerical changes in data collected by each sensor in adjacent sampling periods and construct a drift sequence. ,in, For the first Each node at time... eigenvectors, For feature dimension, Represents the vector norm; S73, Sliding mean deviation calculation submodule, used to calculate the deviation within a set sliding window length. Within this context, calculate the moving average deviation of the communication delay and data drift for each node, assuming the th node... Each node at time step The communication delay value is Then their moving averages are defined as follows:

[0057] in, Represents a node In time forward The average communication delay within each time step Indicates the first Each node at time step The communication delay value at any given time. Represents a node In time forward The mean drift over each time step For sliding index, Represents a node At time step The amount of drift within a given time period; The sliding mean deviation calculation submodule constructs a sliding time window mechanism to estimate the average communication delay and data drift of each sensor node within consecutive time steps, and calculates the mean deviation of the first time step. Each node at time step The sliding statistical deviation value. Let the sliding window length be... ,node At any moment The communication delay is Data drift is The moving average deviation is defined as follows:

[0058] in, Represents a node In time forward The moving average of communication delay over a time step, typically in seconds (s) or milliseconds (ms), depending on the accuracy of the actual communication measurement; Represents a node The average value of data drift within the same time range, with units consistent with the change in the eigenvector, depends on the physical quantity being monitored; A sliding index is used to traverse historical sampling points within a time window. This calculation method eliminates short-term noise while preserving the state trend, providing a stable and adjustable estimation input for subsequent adaptive sampling control.

[0059] S74, Data Interval Adjustment Submodule, is used to construct the sampling interval update function. Update the result based on the current mean deviation. Sampling period of each sensor The calculation method is as follows:

[0060] in, For adjustment coefficients, The joint deviation function of delay and drift depends on and .

[0061] The data interval control submodule constructs a dynamic update function to adjust the sensor's sampling period, and adjusts the node based on the moving average deviation calculation result. The sampling interval is adaptively adjusted. Let the current sampling period be... The delay-drift joint bias function is The adjustment coefficient is The sampling period update formula is:

[0062] in, The unit is time (e.g., seconds s, milliseconds ms), indicating the sensor's first... Node at time The length of the data sampling period; It is a dimensionless joint bias function whose value depends on the mean communication delay of the node within the current sliding window. and drift mean The overall trend of change is used to characterize the stability of the node's operating state; A configurable adjustment factor is used to control the sensitivity of the deviation to the sampling interval adjustment; its unit is... The reciprocal of the value is usually dimensionless. This function can dynamically increase or decrease the sampling interval, enabling on-demand allocation of sensing resources and balanced scheduling of node loads, thereby improving the system's response flexibility and resource utilization in unstable communication environments.

[0063] This implementation introduces a real-time monitoring mechanism for communication latency and data drift between sensor nodes and edge processing modules, constructing a latency-drift dual-dimensional sliding evaluation model to dynamically assess system operational stability while ensuring the integrity of the perceived data. A sliding window mean is used to smoothly model historical communication states and data fluctuation trends, effectively suppressing the impact of short-term abnormal interference on system scheduling. Based on this, a joint deviation-driven sampling interval self-adjustment function is constructed to intelligently adjust the sampling period according to the node's operating status. This ensures high-frequency anomaly detection requirements while reducing resource load under communication-constrained conditions, significantly improving the overall robustness, adaptability, and energy efficiency of the system, providing effective support for a highly stable sensing system in complex water conservancy construction sites.

[0064] Example 1: To verify the effectiveness and practicality of the water conservancy construction site safety early warning system of this invention in a real engineering environment, the project team chose to deploy the system in the "XW Irrigation District Backbone Water Conveyance Canal Construction Project" located in a tributary section of the Three Gorges Reservoir in Yichang City, Hubei Province. This project is a typical mountainous water conservancy construction project, with a total line length of over 9 kilometers. The construction area has a large elevation difference, significant changes in environmental humidity, and a high degree of overlap and density of personnel, equipment, and support structures, making it a typical water conservancy construction environment characterized by "high hidden dangers + multiple work sites + strong interference".

[0065] During the early stages of construction, the project management unit encountered several safety issues, such as scaffold displacement, temporary road slippage, and personnel leaving their posts during nighttime operations. Because traditional monitoring methods mostly rely on a single approach of video and temperature and humidity sensors, they lack the ability to collaboratively analyze multi-source data on-site, making it impossible to predict and respond to potential hazards in real time. This resulted in delayed responses to some risk events and even minor accidents.

[0066] To address the aforementioned issues, the construction team officially implemented the "Water Conservancy Construction Site Safety Early Warning System" proposed in this invention in September 2024. The system deployment includes: 12 video monitoring nodes, 8 sets of environmental sensors, 11 structural stress and support displacement monitoring points, and 7 sets of UWB personnel positioning equipment. Combined with self-developed edge computing terminals deployed across 3 construction zones, it forms a multi-source heterogeneous sensor fusion network. Additionally, 2 mobile edge computing node vehicle-mounted terminals are provided for supplementary data processing at night and in mobile construction areas.

[0067] After system startup, the first step is to use a multi-source heterogeneous sensing module to uniformly timestamp and synchronously sample real-time image data, wind speed, humidity, soil temperature, steel support axial force, and personnel distribution status collected at the construction site. This data is then aggregated and normalized via a sliding window by the edge processing module to form a high-density temporal feature vector. Subsequently, the graph modeling module fuses the spatial distribution of different types of sensor nodes with the statistical covariance matrix to construct a graph-structured adjacency matrix. Finally, a spectral embedding method is used to generate a fused node representation usable by a graph neural network.

[0068] During the inference phase of the early warning model, the system uses a graph convolutional network model combined with a multi-head attention mechanism to infer the node state and dynamically generate the risk level probability distribution of each node. At the same time, the adaptive early warning module inputs construction status indicators (such as current task progress, day and night operation time, and equipment start-up and shutdown frequency) into the edge weight adjustment function, so that the attention propagation path of the model can be dynamically adjusted according to the state at different operation stages, improving the robustness to sudden state changes. Finally, based on the risk judgment output, the system synchronously transmits the corresponding multi-level early warning signals to the on-site LED indicator screen, audible and visual alarm devices, and remote dispatch and command platform to achieve real-time early warning and response linkage.

[0069] To evaluate the effectiveness of the system implementation, the project team conducted a two-month comparative experiment before and after system deployment. The specific statistics are shown in Table 1 below.

[0070] Table 1: Statistical Comparison of Early Warning System Deployment Effects Before and After Deployment at Water Conservancy Construction Sites

[0071] The results show that after deploying this system, the average risk identification lead time was 16.4 minutes, ahead of the actual on-site intervention time, significantly improving accident prevention capabilities. The system successfully identified 18.3 instances of personnel violations per month, more than three times that of traditional manual inspections, and can display violation trajectories in real time through a visual terminal, improving management transparency. For scenarios involving abnormal structural support displacement, the system's alarm success rate jumped from 65.2% to 96.7%, and no actual accident events were missed by the system during the entire assessment period.

[0072] Furthermore, in terms of risk assessment efficiency, the alarm process, which originally had an average response time of over 20 minutes, has been compressed to within 3 minutes, effectively ensuring a real-time closed loop between command and dispatch and frontline response. Statistics show that among the high-risk warning events generated by the system, the actual proportion triggering emergency response is as high as 78.4%, indicating that this level of assessment mechanism has high accuracy and action value. Regarding the concern of construction workers about "false alarms causing disturbance," the system incorporates multiple state constraints and redundant data cross-validation mechanisms in model training, keeping the overall false alarm rate below 2.3%, far lower than the industry average.

[0073] In a typical incident, at 21:37 on October 12, 2024, the system continuously detected structural displacement exceeding the empirical threshold and accompanied by abnormal micro-vibration in the southwest corner support section of Zone D. An early warning was generated in just 47 seconds, and on-site audio-visual linkage and main control platform push were completed within 92 seconds. Subsequent investigation revealed that ground seepage and collapse in the support structure caused bottom support imbalance. Without the early warning, subsequent concrete pouring could very likely have triggered structural collapse. Fortunately, the early response allowed the construction team to immediately suspend operations and reinforce the structure, averting a major construction accident.

[0074] In summary, the water conservancy construction site safety early warning system proposed in this invention forms a complete closed loop in terms of multi-source heterogeneous data fusion, edge computing risk modeling, graph neural network risk inference, and adaptive state control, significantly improving the foresight of risk identification and the efficiency of response. Its deployment results show that the system is not only suitable for static structure monitoring but also possesses deep adaptability to dynamic construction scenarios, truly realizing the implementation and effectiveness of intelligent safety management in water conservancy projects. It provides a replicable and scalable technical path for the safety assurance of subsequent complex water conservancy projects such as dams, power stations, and water diversion tunnels.

[0075] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A safety early warning system for water conservancy construction sites, characterized in that, The system includes: S1, Multi-source heterogeneous sensor module, used to collect visible light monitoring images, environmental parameters, structural status and personnel location information, and encapsulate the above-mentioned sensing data into a raw data stream with timestamps and output it; S2, Edge processing module, used to perform synchronization and standardization processing on the original data stream based on a sliding time window to generate a time-series feature vector sequence; S3, Graph Modeling Module, is used to calculate the covariance matrix between time-series features based on the time-series feature vector sequence, construct a weighted graph by combining the spatial distance of each sensor, and use the spectral embedding method to map each node in the weighted graph into a low-dimensional fusion feature map; S4, Risk Identification Module, is used to input low-dimensional fused feature maps into a graph convolutional network with an attention gating mechanism, and to infer risk level by adjusting edge attention weights; S5, Adaptive early warning module, is used to dynamically adjust the edge attention weights of the graph convolutional network based on real-time construction status indicators, and output multi-level early warning signals based on the risk level inference results; S6, Response module, used to control the audible and visual alarm, field terminal and remote platform to perform response operations according to the multi-level early warning signals; S7, the collaborative control module, is used to monitor the communication delay and data drift between each sensor and the edge processing module, and to adjust the data acquisition interval through sliding mean deviation control.

2. The water conservancy construction site safety early warning system according to claim 1, characterized in that, S3 specifically includes: S31, Covariance Calculation Submodule, is used to construct a feature matrix for each type of sensor data based on the temporal feature vector sequence output by the edge processing module. ,in, Indicates the length of the time window. Representing feature dimension, Represent the real number field and compute its covariance matrix. ; S32, Spatial Distance Calculation Submodule, used to calculate the Euclidean distance between any two sensor nodes based on the spatial deployment location of each sensor on the construction site. and Spatial distance And generate the spatial distance matrix between nodes. ,in, This represents the total number of sensor nodes. Indicates the first The and the first The physical distance between the sensors ; S33, Weighted graph construction submodule, used to fuse the covariance matrix. Spatial distance matrix Construct a weighted adjacency matrix The fusion method is based on the following function: ； in, Indicates the first The and the first The connection weights between nodes To map the covariance in the feature space to the similarity measure of node pairs, , , The fusion coefficient is... Represents an exponential function; S34, Graph embedding generation submodule, used to generate graphs based on the weighted adjacency matrix. and the feature vector corresponding to each sensor node ,in, For nodes In the characteristic matrix The eigenvectors in the graph are used to construct the Laplacian matrix of the weighted graph. ,in, For the degree matrix, and for Perform feature decomposition, before extraction The eigenvectors corresponding to the smallest eigenvalues ​​are used as low-dimensional embedding representations of the nodes to form a low-dimensional fused feature map.

3. The water conservancy construction site safety early warning system according to claim 2, characterized in that, S4 specifically includes: S41, Graph Node Initialization Submodule, is used to receive the low-dimensional fused feature map output by the graph modeling module and initialize the embedding vector of each graph node. The corresponding node type labels and sensor category codes are concatenated to form an enhanced feature representation. ,in, ; S42, Multi-head attention weight calculation submodule, used to calculate weights based on each head during graph convolution. Calculate adjacent nodes and Attention coefficient The calculation method is as follows: ； in, For the number of attention heads, Represents a non-linear activation function. Indicates the first in the figure Enhanced feature representation of each node, Indicates the previous number The node's Enhanced feature representation of each neighboring node, Represents the set of adjacent nodes of a node. Indicates attention head The weight vector Indicates attention head The linear transformation matrix; S43, Risk Feature Aggregation Submodule, used to perform multi-scale aggregation of each neighbor node based on the multi-head attention coefficient to obtain the node update vector. The vector is then fused with the original vector through residual connections and input into the next convolutional layer. S44, Risk Level Determination Submodule, is used to input the node vector output by the multi-layer graph convolution into the fused construction environment state vector, and output the risk level label based on the risk scenario mapping function. ,in, This indicates the total number of risk level categories.

4. A safety early warning system for water conservancy construction sites according to claim 3, characterized in that, S44 specifically includes: S441, the classification input construction submodule, is used to receive the updated representation vector of each node output by the graph convolution aggregation submodule. And obtain the temporal feature vector of the current construction environment parameters in the edge processing module. The two are concatenated to form an extended classification input vector. ,in, , Indicates the output dimension of the convolution. Represents the dimension of state features. This represents a vector concatenation operation; S442, Risk Level Classification Submodule, used to preset the risk level set. ,in, This indicates the total number of risk level categories, with each category corresponding to a different level of construction risk. S443, Discriminant network structure submodule, used to construct a classification neural network composed of fully connected layers, and to receive the extended classification input vector. Intermediate representations are generated by sequentially applying a set of linear transformations and nonlinear activation functions. The output layer uses the softmax activation function to calculate the probability distribution of each risk level. ,in,: ； in, Indicates the output layer number Weight vector of risk level; S444, Risk Label Output Submodule, used to determine the node based on the category with the highest probability in the softmax output as the final classification result. The corresponding risk level label .

5. A safety early warning system for water conservancy construction sites according to claim 4, characterized in that, S5 specifically includes: S51, Construction Status Acquisition Submodule, used to obtain a set of real-time construction status indicators for the current construction area from the time-series feature vector sequence. ,in, Indicates the first Item status indicators, Represents the dimension of state features; S52, Edge Weight Adjustment Submodule, is used to input the set of construction status indicators into the edge weight dynamic mapping function. And the attention coefficient corresponding to each edge in the graph convolutional network. Make adjustments; S53, Dynamic Multi-Level Discrimination Submodule, is used to determine the risk level label of each node in the final output of the graph convolutional network. and its corresponding softmax probability distribution Combined with the dynamic early warning strategy table specified in the construction status Determine the node Warning level ,in, Indicates the total number of warning level categories; S54, Warning Signal Output Submodule, used to output warning signals based on the warning level of each node. Generate corresponding graded early warning signals and transmit the signals to the response module.

6. A safety early warning system for water conservancy construction sites according to claim 2, characterized in that, Specifically, S7 includes: S71, Communication Status Monitoring Submodule, is used to record the communication delay value between each sensor node and the edge processing module in each acquisition cycle. ,in, , Indicates the sampling time index. Indicates the total number of sensors; S72, Data Drift Estimation Submodule, is used to extract the numerical changes in data collected by each sensor in adjacent sampling periods and construct a drift sequence. ,in, For the first Each node at time... eigenvectors, For feature dimension, Represents the vector norm; S73, Sliding mean deviation calculation submodule, used to calculate the deviation within a set sliding window length. Within this context, calculate the moving average deviation of the communication delay and data drift for each node, assuming the th node... Each node at time step The communication delay value is Then their moving averages are defined as follows: ； in, Represents a node In time forward The average communication delay within each time step Indicates the first Each node at time step The communication delay value at any given time. Represents a node In time forward The mean drift over each time step For sliding index, Represents a node At time step The amount of drift within a given time period; S74, Data Interval Adjustment Submodule, is used to construct the sampling interval update function. Update the result based on the current mean deviation. Sampling period of each sensor The calculation method is as follows: ； in, For adjustment coefficients, The joint deviation function of delay and drift depends on and .

Images