A hydraulic tunnel safety monitoring and early warning method

Abstract

The application relates to the technical field of safety monitoring and early warning of hydraulic tunnel structures, and discloses a safety monitoring and early warning method for a hydraulic tunnel, which comprises the following steps: deploying a multi-modal sensor network and an edge computing node inside the tunnel to collect structural parameter data such as displacement, stress and temperature in real time. Through data preprocessing, the data quality is improved, and a multi-scale spatiotemporal graph model is constructed based on the preprocessing result to extract dynamic spatiotemporal correlation features. The model is optimized by using a graph neural network, and the model parameters are trained and aggregated in combination with a federated learning framework, so that the overall performance is improved while privacy is protected. The method performs anomaly detection through a physical constraint equation, triggers a multi-level early warning mechanism, and ensures the safety and stability of the tunnel. According to the application, a multi-scale spatiotemporal graph model is constructed, and a graph neural network is used to optimize the model state, so that dynamic spatiotemporal correlation features can be accurately extracted. This makes it possible to identify changes in the tunnel structure earlier.

Description

Technical Field

[0001] This invention relates to the field of structural safety monitoring and early warning technology for hydraulic tunnels, specifically a method for safety monitoring and early warning of hydraulic tunnels. Background Technology

[0002] The safety of hydraulic tunnels is directly related to the stability of the project and the safety of personnel. Modern monitoring systems typically rely on multiple sensors to collect structural parameters such as displacement, stress, and temperature inside the tunnel. These sensors improve the accuracy of monitoring to a certain extent.

[0003] First, existing monitoring systems generally suffer from latency in data processing because data transmission often involves multiple stages, including data acquisition, transmission, and storage. This latency leads to insufficient real-time performance, making it difficult to respond promptly in emergencies. Second, many traditional monitoring systems lack effective data feedback and processing mechanisms. Even when an initial alarm is triggered, subsequent responses are often slow. Due to the lack of automated processing and intelligent analysis, monitoring personnel need to operate manually, increasing the risk of misjudgment and missed detection. Third, existing technologies are also inadequate in detecting anomalies in monitoring data. Early warning systems based on simple thresholds struggle to cope with complex situations. This approach is susceptible to noise, and sensor data varies under different environmental conditions, leading to inconsistencies in detection and low signal-to-noise ratios. Furthermore, traditional systems often fail to fully utilize edge computing and federated learning technologies. Edge computing can move data processing tasks closer to the data source, thereby reducing latency. In addition, existing technologies rely heavily on static models for dynamic model updates. This makes it difficult for models to adapt to real-time changes in the environment and monitoring data, thus affecting the accuracy of predictions. Furthermore, some applications lack effective security measures; insufficient encryption is not implemented during data transmission, which may lead to the leakage of sensitive data and affect the overall system security. Current privacy protection technologies cannot adequately protect user data without impacting performance. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for safety monitoring and early warning of hydraulic tunnels, which solves the problems of insufficient real-time performance and accuracy in safety monitoring of hydraulic tunnels.

[0005] To achieve the above objectives, the present invention employs the following technical solution: deploying a multimodal sensor network and edge computing nodes within the tunnel to collect structural parameter data of the tunnel in real time;

[0006] The collected structural parameter data is preprocessed, including filtering to remove noise and outliers.

[0007] Based on the preprocessed structural parameter data, a multi-scale spatiotemporal graph model is constructed to extract dynamic spatiotemporal correlation features, identify potential structural changes, and use the output of this model to provide basic information for subsequent analysis.

[0008] The constructed multi-scale spatiotemporal graph model is trained and its parameters are aggregated using a federated learning framework.

[0009] Anomaly detection is performed on the output results based on the physical constraint equations to accurately identify potential anomalies and calculate the corresponding risk levels.

[0010] Based on the anomaly detection results, a multi-level early warning mechanism is triggered, and the early warning strategy is automatically adjusted.

[0011] After the early warning strategy is implemented, the model parameters are dynamically updated to reflect the new monitoring data and early warning response results.

[0012] Preferably, the multimodal sensor network includes:

[0013] Displacement sensors are used to monitor displacement changes at various locations in the tunnel in order to assess the tunnel's deformation.

[0014] Stress sensors are used to detect the distribution and changes of stress on tunnel structural materials in order to determine the stress state of the structure.

[0015] Temperature sensors are used to collect real-time temperature data inside the tunnel in order to analyze the impact of temperature changes on the tunnel structure.

[0016] Preferably, the data preprocessing includes:

[0017] First, filtering is performed, using a weighted moving average method to process the structural parameter data in order to remove noise and outliers from the data;

[0018] Then, smoothing is performed, using techniques such as exponential smoothing.

[0019] Finally, normalization is performed to unify data from different sources to the same standard range.

[0020] Preferably, the multi-scale spatiotemporal graph model includes:

[0021] Define a graph structure G(V,E), where V is the set of nodes and E is the set of edges. Nodes V represent different monitoring points, and edges E represent the dependencies between nodes.

[0022] For each node i∈V, define a node feature vector Hi, denoted as Hi=f(Xi,Ti), where Xi is the structural parameter data of the node and Ti is the temporal feature of the node;

[0023]

[0024] in, Let i represent the feature vector of node i in layer l+1, and let l represent the node features after being updated through layer l.

[0025] σ: is the activation function, used to introduce nonlinearity and make the model more robust;

[0026] ∑ j∈N(i) : Represents the calculation for all adjacent nodes j of node i, where N(i) is the set of adjacent nodes of node i;

[0027] The normalization constant is used to represent the relative influence between nodes, so that the connection weights can be properly handled.

[0028] W (l) : Represents the weight matrix of the l-th layer, which is responsible for mapping the features of adjacent nodes to the new feature space of node i;

[0029] This represents the feature vector of adjacent node j in layer l;

[0030] b (l) : Represents the bias term of the l-th layer, which is added to the node feature calculation to improve model performance;

[0031] Using the time feature T, a Long Short-Term Memory (LSTM) network is used for modeling to obtain the temporal output O, which is expressed by the following formula:

[0032] O t =LSTM(H,T);

[0033] Where H represents node features and T represents the time series;

[0034] Combining the GNN output H and the LSTM output O yields the comprehensive feature Z:

[0035] Z = W·H + U·O;

[0036] Here, W and U are the corresponding weight matrices, respectively;

[0037] Structural changes are identified based on comprehensive feature Z. Potential structural anomalies are predicted using a classification model, and the output category label Y is represented as follows:

[0038] Y = Softmax(W Y ·Z+b Y );

[0039] Among them, W Y and b Y These are the weights and biases for the classification model.

[0040] Preferably, the physical constraint equations include:

[0041] Definition of physical constraint equations: For the monitored tunnel structure, the existing physical constraint relationships are assumed to be represented by the following equations:

[0042]

[0043] Among them, F total For the total force, F i Let be the force acting under the i-th monitoring state.

[0044] Threshold determination criteria: A safety threshold is set based on the yield strength of the material, expressed as:

[0045] threshold = σ allowable ·(1+□);

[0046] Where, σ allowable □ represents the allowable yield strength of the material, and □ represents the safety margin.

[0047] Output comparison with threshold: The output is compared to identify potential anomalies, as shown below:

[0048] if|F o utput|>threshold,thendetectanomaly;

[0049] The if statement is used to check whether a given condition is true.

[0050] |F output |: Represents the absolute value of the output. Using absolute values ​​is to account for the possibility of the output being positive or negative, enabling the detection mechanism to effectively identify any cases outside the normal range;

[0051] threshold: represents the threshold defined earlier;

[0052] The `thendetectanomaly` operation indicates the action to be performed when the condition is true, i.e., an anomaly is detected. In practical applications, this may trigger an alarm system, log the event, or take other related measures to address potential security risks.

[0053] Risk level calculation: The risk level R is defined by calculating the degree to which the safety threshold is exceeded, and is expressed as:

[0054]

[0055] Among them, |F output-threshold|: Indicates the absolute difference between the output result and the threshold. This value is used to measure the degree of deviation between the actual output and the set safety range.

[0056] threshold: Used as the denominator, it obtains the relative degree of change by comparing with the threshold, ensuring that the calculated risk level is based on the proportional relationship of the threshold;

[0057] • 100%: Converts the relative risk calculation results into a percentage form, making them more intuitive and easier to understand and assess the risk level;

[0058] Provides quantitative basis for risk assessment;

[0059] Risk level classification: Based on the calculated risk level R, potential anomalies are classified into different levels, as follows:

[0060]

[0061] in, A risk level of less than 10% indicates a low risk, typically within an acceptable range.

[0062] A risk level between 10% and 30% indicates a medium risk.

[0063] A risk level of 30% or higher indicates a high risk, requiring urgent measures to prevent potential harm.

[0064] Preferably, the abnormality detection results are divided into three warning levels, including:

[0065] Input for anomaly detection results: Based on the detected risk level R, determine the warning trigger conditions, as shown below:

[0066]

[0067] Among them, threshold alert Risk thresholds that trigger warnings;

[0068] trigger alertb This indicates that an alarm has been triggered;

[0069] Setting up a multi-level early warning mechanism: Define the response measures for the multi-level early warning mechanism based on different risk levels R, as shown below:

[0070]

[0071] Dynamic adjustment of early warning strategy: The early warning strategy is adjusted based on the anomaly detection results, as shown below:

[0072]

[0073] Early warning strategy feedback: Record and provide feedback on the adjusted early warning strategy, represented as:

[0074] S (t+1) =S (t) +ΔS;

[0075] Among them, S (t) Let ΔS be the current strategy, and ΔS be the strategy adjustment amount.

[0076] Preferably, the dynamic updating of the model parameters includes:

[0077] Inputting new monitoring data: Inputting new monitoring datasets

[0078]

[0079] Where x i For the input features, y i For the corresponding output, m represents the total number of samples in the dataset. It is an integer from 1 to m, defining the number of samples in the dataset, thus indicating the size of the dataset.

[0080] Represented as:

[0081] D new = [X,Y];

[0082] Among them, D new This represents the monitoring data, where X is the feature matrix and Y is the output vector.

[0083] The goal of model parameter updates: Let W be the model parameter that needs to be updated. The goal is to minimize the new loss function, defined as follows:

[0084]

[0085] Among them, L(W;D) new ): This indicates that given the model parameters W and the new monitoring data D new The loss function;

[0086] m: Represents the total number of samples in the dataset. It is the basis for calculating the average in the loss function to obtain the contribution of each sample to the total loss;

[0087] This represents summing over all samples from 1 to m;

[0088] This represents the loss function itself, which is typically used to measure the difference between predicted and actual values ​​(e.g., mean squared error, cross entropy, etc.).

[0089] f(x i ;W): indicates that the model is applied to the i-th sample x. i The predicted output depends on the sample features and the current model parameters W;

[0090] y i : indicates that it corresponds to x i The true label or target value is usually the result that the model hopes to predict.

[0091] Dynamic update method for model parameters: The model parameters are updated using gradient descent, specifically as follows:

[0092]

[0093] Where η is the learning rate. Let L be the gradient of the loss function L with respect to the parameter W;

[0094] Adjustments based on early warning response results: The early warning response results are incorporated into the update process. The modified parameter updates can be represented by a weighted coefficient λ:

[0095] W adjusted ←W+λΔW;

[0096] Among them, W adjusted This indicates the adjusted parameters;

[0097] ←: Indicates an assignment operation, which means assigning the calculation result on the right to the variable on the left;

[0098] W+λΔW: This part indicates that the new parameter value is the current parameter W plus the learning rate λ multiplied by the parameter change ΔW. This achieves parameter adjustment, enabling the model to respond to detected alerts;

[0099] Model evaluation: After each parameter update, the model performance is evaluated using the validation dataset D. val The expression is as follows:

[0100]

[0101] in, This indicates that the number of samples in the validation set is normalized, |D val | represents the total number of samples in the validation set, used to calculate the average performance;

[0102] This indicates that for the validation set D val Each sample (x) j ,y j Summation is performed;

[0103] f(x j W adjusted): Indicates the use of the adjusted model parameters W adjusted For input feature x j The prediction results;

[0104] This indicates the calculation of the prediction result f(x) j W adjusted ) and the real label y j The loss between the two samples. This is the main metric used to evaluate the model's performance on that sample.

[0105] Preferably, the functions of the edge computing node include:

[0106] Edge computing nodes can perform real-time data storage and historical data management, ensuring that data integrity and reliability are maintained even in the event of network interruption or other abnormal situations.

[0107] This invention provides a method for safety monitoring and early warning of hydraulic tunnels. It has the following beneficial effects:

[0108] 1. This invention employs a multimodal sensor network and edge computing nodes, which not only efficiently collects data such as displacement, stress, and temperature, but also processes this information in real time. This technical solution improves the timeliness and accuracy of data acquisition and solves the problem of incomplete data sources compared to the application of single sensors in existing technologies.

[0109] 2. This invention also introduces a data preprocessing procedure, including filtering and smoothing. This significantly improves data quality and effectively reduces errors caused by noise and outliers. Compared with existing technologies, the use of a simple sensor output method solves the problem of insufficient quality control before data analysis and enhances the reliability of early warning.

[0110] 3. By constructing a multi-scale spatiotemporal graph model and utilizing graph neural networks to optimize the model state, this invention can accurately extract dynamic spatiotemporal correlation features. This enables earlier identification of changes in tunnel structure. Compared to existing static monitoring schemes, the dynamic monitoring capabilities and real-time response of this invention overcome the problem of slow response to changes in environmental variables.

[0111] 4. The federated learning framework of this invention ensures data privacy and improves the overall performance of the model. This strategy optimizes the model training process while ensuring distributed operation. Compared with existing methods that rely on centralized data processing, this solution effectively addresses the privacy risks associated with data sharing, making the monitoring system more secure and reliable. Attached Figure Description

[0112] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0113] The technical solutions in 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.

[0114] Please see the appendix Figure 1 This invention provides a method for safety monitoring and early warning of hydraulic tunnels, comprising:

[0115] A multimodal sensor network and edge computing nodes are deployed inside the tunnel to collect structural parameter data of the tunnel in real time, including displacement, stress and temperature information;

[0116] The multimodal sensor network deployed within the tunnel consists of multiple sensors, specifically displacement sensors, stress sensors, and temperature sensors. These sensors possess good environmental adaptability and high sensitivity, enabling them to acquire relevant structural parameters in real time within the complex tunnel environment. Displacement sensors can be deployed at key structural points in the tunnel to monitor displacement changes in various parts of the tunnel. Stress sensors can be deployed using strain gauges or pressure sensors to monitor the stress distribution on the structure in real time. Temperature sensors employ thermocouples or thermistors to acquire temperature data within the tunnel in real time. Edge computing nodes connect to the aforementioned sensor network and are responsible for processing and analyzing the real-time acquired data. First, the edge computing nodes receive digital signals from each sensor and store them in a local database. This process effectively reduces data transmission latency and improves the speed and efficiency of data processing.

[0117] The collected structural parameter data is preprocessed, including filtering to remove noise and outliers.

[0118] In real-time data processing, edge computing nodes first perform data preprocessing. This preprocessing includes filtering to remove noise and outliers from sensor signals. A weighted moving average method is used here, and its calculation is as follows:

[0119]

[0120] Where y[n] is the filtered data, x[n] is the original sensor data, N is the size of the applied filtering window, and w k Using the corresponding weighting coefficients, filtering can significantly improve the accuracy of the data. Subsequently, edge computing nodes undergo smoothing, typically using exponential smoothing, with the calculation relationship as follows:

[0121] S[n] = αx[n] + (1-α)S[n-1];

[0122] Where S[n] is the smoothed data, x[n] is the current sampled value, and α is the smoothing factor, which ranges from 0 to 1. Furthermore, the smoothing process reduces the random fluctuations of the signal and ensures the reliability of subsequent analysis.

[0123] The collected structural parameter data is preprocessed, including filtering to remove noise and outliers.

[0124] After smoothing, normalization is performed. This process unifies data from different sources to the same standard range, as follows:

[0125]

[0126] Where z represents the normalized data, and x represents the original data. min and x max These are the minimum and maximum values ​​of the dataset, respectively. This normalization step ensures data consistency, making subsequent analysis and comparison simpler and faster. After data processing, the edge computing nodes can transmit the processing results to the central server for further analysis. This process first establishes a data transmission channel, then compresses the data using a specific encoding scheme to reduce network transmission time, and finally outputs cleaned monitoring data for advanced analysis and decision support.

[0127] Based on the preprocessed structural parameter data, a multi-scale spatiotemporal graph model is constructed to extract dynamic spatiotemporal correlation features, identify potential structural changes, and use the output of this model to provide basic information for subsequent analysis.

[0128] First, based on the structural parameter data obtained in the preprocessing stage, including displacement, stress, and temperature information, a multi-scale spatiotemporal graph model is constructed. The core of this model lies in forming a spatiotemporal graph through the representation of nodes and edges. Specifically, the nodes of the graph represent the monitoring parameters collected by various sensors, while the edges connect the spatiotemporal relationships between different parameters. For example, an adjacency matrix is ​​constructed for the monitoring data of the tunnel to represent the connection relationships between nodes;

[0129] Modeling a graph using its adjacency matrix can be represented as follows:

[0130]

[0131] Among them, A ij The elements of the adjacency matrix represent the connection relationships between nodes. Through these connections, the spatiotemporal correlation of sensor data can be effectively captured, and important structural deformation and stress distribution characteristics can be identified.

[0132] After constructing a multi-scale spatiotemporal graph model, a graph neural network (GNN) is further used to optimize it. Graph neural networks can capture complex spatiotemporal relationships between data by iteratively updating the feature states of nodes. The specific update formula is as follows:

[0133]

[0134] in, Let N(i) represent the node characteristics at layer l, and W represent the neighboring nodes of node i. (l) It is the learned weight matrix, b (l) Let σ be the bias term and σ be the activation function. After multiple iterations, the feature vectors of a node can effectively aggregate information from its neighborhood, thereby identifying potential structural changes.

[0135] Furthermore, in order to extract dynamic spatiotemporal correlation features from the constructed multi-scale spatiotemporal graph model, feature aggregation is first performed. The global feature representation of each node is calculated based on the relationship between the node and its neighboring nodes. These features include the clustering coefficient and betweenness centrality of the node, which can reflect the importance of the node in the entire spatiotemporal structure.

[0136] After feature extraction, potential structural changes are identified based on the dynamic features output by the model. Specifically, different thresholds can be set to determine whether the model output exceeds the normal range, thereby identifying potential abnormal events. This process requires real-time analysis of the output results. First, the extracted dynamic spatiotemporal features are compared with the preset normal range. Then, by determining the threshold, it is determined whether there are significant structural changes. Finally, the judgment result provides the basis for subsequent analysis, such as whether to initiate further security assessment procedures.

[0137] The constructed multi-scale spatiotemporal graph model is trained and its parameters are aggregated using a federated learning framework.

[0138] First, federated learning is a distributed learning approach that improves overall model performance by aggregating model updates from participating nodes without sharing local data. It begins by training a multi-scale spatiotemporal graph model on each edge computing node. The training optimization objective of this model is to minimize the loss function:

[0139]

[0140] in, Here, is the global loss function, and N is the number of nodes participating in the learning process. Let y be the loss function for the i-th edge node. i For the true value, These are the model predictions. Each edge node updates its model parameters locally and calculates the parameter update amount.

[0141] Then, the edge nodes upload the calculated local update quantities to the central server. This step ensures that each node only uploads the updated model parameters and does not transmit the original data. Specifically, the nodes calculate the update quantities using the following formula:

[0142]

[0143] Where, Δθ i Let be the parameter update amount for the i-th node. For the updated model parameters, These are the original model parameters. This process effectively protects data privacy and reduces data transmission costs. After receiving the parameter updates uploaded by all nodes, the central server performs an aggregation operation. The aggregation process uses a weighted average strategy to reasonably reflect the contribution of each edge node. Specifically, the aggregation process is represented as follows:

[0144]

[0145] Where, θ global n represents the global model parameters. i Let N be the number of samples at node i and N be the total number of samples. By weighted aggregation of the update amounts of all nodes, a more accurate global model can be obtained.

[0146] After parameter aggregation, the central server sends the updated global model to all participating edge nodes so that the nodes can continue to perform the next round of local training. This process first establishes the global model, and then iteratively trains the model of each node through the corresponding update algorithm, and finally outputs the optimized multi-scale spatiotemporal graph model.

[0147] Throughout the implementation process, the edge computing nodes and the central server establish a connection through a secure encrypted communication protocol to ensure the security of parameter updates and model aggregation. This design can effectively reduce potential security risks and guarantee the reliability of data transmission.

[0148] This federated learning framework enables efficient training and updating of multi-scale spatiotemporal graph models while preserving data privacy at each edge node as much as possible. The implementation scheme clarifies the logical connections and physical relationships between various technical elements, ensuring that those skilled in the art can implement it directly, thereby meeting the legal requirement of full disclosure. In addition, the clear formula derivation and variable definition ensure the integrity and readability of the technical features.

[0149] Anomaly detection is performed on the output results based on the physical constraint equations to accurately identify potential anomalies and calculate the corresponding risk levels.

[0150] In the deployed multimodal sensor network, each sensor collects real-time structural parameter data of the tunnel, including displacement, stress, and temperature information. This data is initially processed by edge computing nodes and then sent to the central computing system for further analysis. Based on this, a physical constraint model is first established to describe the relationship between stress and strain. This model preferably uses the constitutive equations for linear elastic materials, the expression of which is:

[0151] σ=E·ε

[0152] Where σ is stress, E is the elastic modulus of the material, representing the material's stiffness, and ε is strain, representing the degree of deformation of the material. Strain can be calculated using the following formula:

[0153]

[0154] Where ΔL represents the deformation and L0 represents the original length of the material, the current stress and strain values ​​can be obtained by analyzing the real-time monitoring data.

[0155] Subsequently, the monitoring data will be substituted into the physical constraint model for anomaly detection. First, a threshold is set to determine whether an anomaly exists. For example, when the detected stress exceeds the set safety threshold, the system marks the data point as an anomaly. The risk level can be calculated using the following formula:

[0156]

[0157] Where R is the risk level, σ actual The actual monitored stress value, σ threshold For the safety threshold, σ max This calculation method, which uses the maximum load-bearing stress of the material, enables accurate identification of potential anomalies.

[0158] In the implementation process, the physical connection of the data flow involves data collection by multimodal sensors, initial processing by edge computing nodes, and then transmission of the data to the central computing system for analysis. The establishment of the physical constraint model and data input ensure the logical connection between the various technical elements.

[0159] Based on the model's output, the system assesses and classifies abnormal situations, generating corresponding alarm information. After removing noise and abnormal data, the system will provide real-time feedback on the current risk level and transmit alarms to monitoring operators, thereby achieving dynamic monitoring and timely response.

[0160] Based on the anomaly detection results, a multi-level early warning mechanism is triggered, and the early warning strategy is automatically adjusted.

[0161] In the safety monitoring and early warning method for hydraulic tunnels, the result of anomaly detection is key to invoking a multi-level early warning mechanism. First, edge computing nodes perform anomaly detection through physical constraint equations and calculate the risk level based on real-time collected structural parameter data. The formula for calculating the risk level is as follows:

[0162]

[0163] Where R represents the risk level, S is the stress state of the current monitoring data, and S0 is the safety threshold of the structure. This formula is used to judge the stability of the tunnel and effectively identify potential anomalies.

[0164] After anomaly detection is completed, the system compares the risk level with the preset threshold to determine the triggered warning level. The specific process is as follows: first, a warning level classification standard is established, including three states: normal, attention, and danger. Then, the detection results are analyzed through a judgment model, and the current warning instruction is output according to the risk level.

[0165] If the risk level is less than the safety threshold, the system will confirm the status as normal and no additional measures are required. If the risk level is within a certain range, the system will enter an alert state, send a warning to the monitoring personnel, and remind them to conduct regular checks and strengthen observation. If the risk level exceeds the set safety threshold, the system will automatically trigger a dangerous state and take emergency response measures, including initiating evacuation procedures, notifying the emergency team, and implementing necessary protective measures, thereby ensuring the safety of tunnel monitoring.

[0166] In implementing the early warning mechanism, the system possesses self-learning capabilities, enabling it to dynamically adjust the early warning strategy based on historical monitoring data and current anomaly detection results. Specifically, when monitoring data and risk levels change, the model parameters will be updated accordingly. This update mechanism consists of the following algorithms:

[0167]

[0168] Where, θ t θ represents the current model parameters. t-1 This represents the model parameters at the previous time step, where η is the learning rate. This represents the gradient of the loss function with respect to the model parameters. This process ensures that the model is automatically adjusted after each data acquisition to maintain a rapid response to changes in tunnel conditions.

[0169] Furthermore, the implementation of the early warning mechanism also involves data feedback and visualization. Monitoring personnel will obtain real-time monitoring data and early warning information through a visual interface. The graphical display enhances the timeliness and understandability of the alarms, thereby helping operators make quick judgments and decisions. The interface will update the monitoring data in real time and provide corresponding warnings and countermeasures based on the current status, ensuring that monitoring personnel can immediately grasp the overall safety status of the tunnel.

[0170] After the early warning strategy is implemented, the model parameters are dynamically updated to reflect the new monitoring data and early warning response results;

[0171] After the early warning strategy is implemented, the new monitoring data collected by the edge computing nodes will be used to dynamically update the parameters of the multi-scale spatiotemporal graph model. This process first includes collecting real-time monitoring data from various sensor networks, which covers new displacement, stress and temperature information. This information is transmitted to the central data integration system to ensure the timeliness and accuracy of data updates.

[0172] During the dynamic updating of model parameters, an incremental learning algorithm is used to update the model parameters based on new monitoring data. The basic principle is to adjust the previous model using newly collected data, thereby reflecting environmental changes and new monitoring information. An example of the mathematical model updating process is as follows:

[0173] First, the loss function is defined as follows:

[0174]

[0175] Where L(θ) is the loss function, N is the number of samples, and y i For the real labels, f(x) i ;θ) is the predicted value, θ is the model parameter, and the purpose of the loss function is to minimize the difference between the predicted value and the actual value;

[0176] Next, after each new data acquisition, the parameters are updated using the following formula:

[0177]

[0178] Where, θ ′ Here are the updated parameters, and η is the learning rate. The gradient of the loss function with respect to the model parameters is used in this process, which enables the model to adaptively adjust after new data confirms the detection and early warning response results.

[0179] The dynamic update of the model needs to take into account the quality and relevance of the new monitoring data. Therefore, the system first performs data preprocessing, including filtering, smoothing and normalization, to ensure that high-quality data without noise is input into the model update process. This can improve the reliability of the updated model parameters and enhance the ability to predict future data.

[0180] After the parameters are updated, the new model parameters will be used to regenerate the multi-scale spatiotemporal graph model in order to extract new dynamic spatiotemporal correlation features. This process first establishes a new spatiotemporal graph model, then updates the adjacency matrix of the graph by calculating the node states and connections in the network, and finally optimizes the model through a graph neural network.

[0181] The model update process between edge computing nodes and the central server is executed in the following steps: first, new monitoring data is collected; then, data processing is performed to filter out anomalies; and finally, incremental learning algorithms are used to update the model parameters, thereby reflecting the new monitoring data and early warning response results. This process ensures that the model parameters are dynamically adjusted as the data changes, maintaining the ability to respond quickly to potential anomalies.

[0182] 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 monitoring and early warning of hydraulic tunnels, characterized in that, Includes the following steps: A multimodal sensor network and edge computing nodes are deployed inside the tunnel to collect structural parameter data of the tunnel in real time; The collected structural parameter data is preprocessed, including filtering to remove noise and outliers. Based on the preprocessed structural parameter data, a multi-scale spatiotemporal graph model is constructed to extract dynamic spatiotemporal correlation features, identify potential structural changes, and use the output of this model to provide basic information for subsequent analysis. The constructed multi-scale spatiotemporal graph model is trained and its parameters are aggregated using a federated learning framework. Anomaly detection is performed on the output results based on the physical constraint equations to accurately identify potential anomalies and calculate the corresponding risk levels. Based on the anomaly detection results, a multi-level early warning mechanism is triggered, and the early warning strategy is automatically adjusted. After the early warning strategy is implemented, the model parameters are dynamically updated to reflect the new monitoring data and early warning response results.

2. The method for safety monitoring and early warning of hydraulic tunnels according to claim 1, characterized in that, The multimodal sensor network includes: Displacement sensors are used to monitor displacement changes at various locations in the tunnel in order to assess the tunnel's deformation. Stress sensors are used to detect the distribution and changes of stress on tunnel structural materials in order to determine the stress state of the structure. Temperature sensors are used to collect real-time temperature data inside the tunnel in order to analyze the impact of temperature changes on the tunnel structure.

3. The method for safety monitoring and early warning of hydraulic tunnels according to claim 1, characterized in that, The data preprocessing includes: First, filtering is performed, using a weighted moving average method to process the structural parameter data in order to remove noise and outliers from the data; Then, smoothing is performed using exponential smoothing techniques. Finally, normalization is performed to unify data from different sources to the same standard range.

4. The method for safety monitoring and early warning of hydraulic tunnels according to claim 1, characterized in that, The multi-scale spatiotemporal graph model includes: Define a graph structure G(V,E), where V is the set of nodes and E is the set of edges. Nodes V represent different monitoring points, and edges E represent the dependencies between nodes. For each node i∈V, define a node feature vector Hi, denoted as Hi=f(Xi,Ti), where Xi is the structural parameter data of the node and Ti is the temporal feature of the node; in, Let i represent the feature vector of node i in layer l+1, and let l represent the node features after being updated through layer l. σ: is the activation function, used to introduce nonlinearity and make the model more robust; ∑ j∈N(i) : Represents the calculation for all adjacent nodes j of node i, where N(i) is the set of adjacent nodes of node i; The normalization constant is used to represent the relative influence between nodes, so that the connection weights can be properly handled. W (l) : Represents the weight matrix of the l-th layer, which is responsible for mapping the features of adjacent nodes to the new feature space of node i; This represents the feature vector of adjacent node j in layer l; b (l) : Represents the bias term of the l-th layer, which is added to the node feature calculation to improve model performance; Using the time feature T, a Long Short-Term Memory (LSTM) network is used for modeling to obtain the temporal output O, which is expressed by the following formula: O t =LSTM(H,T); Where H represents node features and T represents the time series; Combining the GNN output H and the LSTM output O yields the comprehensive feature Z: Z = W·H + U·O; Here, W and U are the corresponding weight matrices, respectively; Structural changes are identified based on comprehensive feature Z. Potential structural anomalies are predicted using a classification model, and the output category label Y is represented as follows: Y =Softmax(W Y ·Z+b Y ): Among them, W Y and b Y These are the weights and biases for the classification model.

5. The method for safety monitoring and early warning of hydraulic tunnels according to claim 1, characterized in that, The federated learning framework includes: Model initialization: Initialize the global model parameter W0 on the central server, as follows: W0←Initialize(); Local model training: On each edge computing node k, local data D is used. k The model training and update process is as follows: in, This represents the weight update of node k in the t-th iteration, which includes the parameters of the currently trained model; This represents the weight of node k in the (t-1)th iteration; η: Represents the learning rate. This hyperparameter controls the step size of each update, ensuring that the adjustment of model weights is not too large, thereby avoiding oscillations during training. Represents the loss function L k Regarding weight The gradient; L k The loss function is based on the local data D of node k. k To evaluate the accuracy of the model; D k The local dataset on node k includes the input features used for training and their corresponding labels; L_k is the local loss function, expressed as: in, Used for processing from D k The loss of each sample is averaged to obtain an overall loss assessment; This represents the summation of all training samples at node k, where (x, y) are the features and labels of the samples; f(x;W): represents the prediction of the input feature x using the current weight W through the model f, and the calculation of the output result; This represents the loss function used to compare the difference between the model's predicted values ​​f(x; W) and the true labels y; Model update feedback: After each edge node completes training, it uploads the model update data to the central server. in: This indicates the amount of progress or adjustment the model has made during the learning process, and it can be used to evaluate the effectiveness of the update. Model aggregation: In the central server, the parameters are aggregated by combining the update data uploaded by all edge nodes to obtain a new global model. The parameter W^{∧{(t)} is represented as: Among them, W (t) : Represents the weights of the global model obtained after the t-th iteration; W (t-1) : Represents the weights of the global model at the (t-1)th iteration, i.e., the parameters of the global model in the previous iteration; This represents the weight update amount of edge node k in the t-th iteration, reflecting the changes in the local model of node k during this round of training; K: Represents the total number of edge nodes participating in training, used as the denominator to calculate the average update amount, to ensure the uniformity of the model aggregation process; The process of training the local model in the previous step is repeated until the model is aggregated in the next step, until the termination condition is met, such as reaching the set training error threshold or reaching the predetermined number of iterations n.

6. The method for safety monitoring and early warning of hydraulic tunnels according to claim 1, characterized in that, The physical constraint equations include: Definition of physical constraint equations: For the monitored tunnel structure, the existing physical constraint relationships are assumed to be represented by the following equations: Among them, F total For the total force, F i Let be the force acting under the i-th monitoring state. Threshold determination criteria: A safety threshold is set based on the yield strength of the material, expressed as: threshold=σ allowable ·(1+□); Where, σ allowable □ represents the allowable yield strength of the material, and □ represents the safety margin. Output comparison with threshold: The output is compared to identify potential anomalies, as shown below: if|F o utput|>threshold,thendetectanomaly; The if statement is used to check whether a given condition is true. |F output |: Represents the absolute value of the output. Using absolute values ​​is to account for the possibility of the output being positive or negative, enabling the detection mechanism to effectively identify any cases outside the normal range; threshold: represents the threshold defined earlier; The `thendetectanomaly` operation indicates the action to be performed when the condition is true, i.e., an anomaly is detected. In practical applications, this may trigger an alarm system, log the event, or take other related measures to address potential security risks. Risk level calculation: The risk level R is defined by calculating the degree to which the safety threshold is exceeded, and is expressed as: Among them, |F output -threshold|: Indicates the absolute difference between the output result and the threshold. This value is used to measure the degree of deviation between the actual output and the set safety range. threshold: Used as the denominator, it obtains the relative degree of change by comparing with the threshold, ensuring that the calculated risk level is based on the proportional relationship of the threshold; • 100%: Converts the relative risk calculation results into a percentage form, making them more intuitive and easier to understand and assess the risk level; Provides quantitative basis for risk assessment; Risk level classification: Based on the calculated risk level R, potential anomalies are classified into different levels, as follows: in, A risk level of less than 10% indicates a low risk, typically within an acceptable range. A risk level between 10% and 30% indicates a medium risk. A risk level of 30% or higher indicates a high risk, requiring urgent measures to prevent potential harm.

7. The method for safety monitoring and early warning of hydraulic tunnels according to claim 1, characterized in that, Based on the anomaly detection results, three warning levels are defined, including: Input of anomaly detection results: Based on the detected risk level R, determine the conditions for triggering an early warning. Represented as: Among them, threshold alert Risk thresholds that trigger warnings; trigger alertb This indicates that an alarm has been triggered; Setting up a multi-level early warning mechanism: Define the response measures for the multi-level early warning mechanism based on different risk levels R, as shown below: Dynamic adjustment of early warning strategy: The early warning strategy is adjusted based on the anomaly detection results, as shown below: Early warning strategy feedback: Record and provide feedback on the adjusted early warning strategy, represented as: S (t+1) =S (t) +ΔS; Among them, S (t) Let ΔS be the current strategy, and ΔS be the strategy adjustment amount.

8. The method for safety monitoring and early warning of hydraulic tunnels according to claim 1, characterized in that, 。 9. A method for safety monitoring and early warning of hydraulic tunnels according to claim 1, characterized in that, The functions of the edge computing node include: Edge computing nodes can perform real-time data storage and historical data management, ensuring that data integrity and reliability are maintained even in the event of network interruption or other abnormal situations.

Images