A method and system for identifying abnormal data in bridge construction monitoring

Abstract

The present invention relates to the technical fields of artificial intelligence and data processing, and specifically relates to a method and system for identifying abnormal data in bridge construction monitoring. The method is as follows: collect bridge monitoring vibration signal data by setting sensors and devices; perform cleaning, conversion, and standardization processing on the collected data; construct an abnormal data identification model for bridge construction monitoring based on a graph neural network model, input the processed data into the abnormal data identification model for bridge construction monitoring for training to identify abnormal situations. The system analyzes based on the collected bridge structure information, hydrological and meteorological information, real-time monitoring information, as well as the abnormal identification results of maintenance and handover information and methods, and transmits the analysis results to the front end. The present invention constructs an abnormal data identification model for bridge construction monitoring based on a graph neural network model and trains the model, which can monitor complex vibration signal data during bridge construction and better identify abnormal situations.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and data processing, and particularly to a method and system for identifying abnormal data in bridge construction monitoring. Background Art

[0002] During the bridge construction process, the accuracy and reliability of monitoring data are crucial for ensuring construction safety and preventing potential risks. However, traditional monitoring data processing methods often have many limitations when faced with complex bridge construction environments. For example, traditional machine learning methods rely on manually designed features and are difficult to fully express the non-linear relationships hidden in the data. Especially when the data volume is huge and has high-dimensional features, they are easily restricted by overfitting and computational efficiency. In addition, when dealing with data with complex similarity and heterogeneity, existing technologies often assume that the relationships between data are uniform, ignoring the local structure and spatio-temporal similarity between nodes, resulting in inaccurate classification effects and being easily interfered by noise. At the same time, traditional neural network methods cannot fully consider the graph structure information between nodes when processing structured data, ignoring the potential spatio-temporal relationships in the data and being difficult to capture complex data relationships. These deficiencies make it difficult for existing technologies to effectively identify abnormal situations when faced with complex vibration signal data in bridge construction monitoring, and cannot meet the requirements of high precision, high efficiency, and high robustness for monitoring data processing in actual projects.

[0003] The Chinese invention patent with the publication number CN119535441A proposes a method and system for detecting damage to the bridge pavement layer based on ground penetrating radar, obtaining clear radar echo data, judging whether there are damage signs, extracting waveform coding features of the detected damage signs to identify damage features; comparing and analyzing the extracted damage features with the database in the normal state to determine the existence and type of damage; according to the analysis results, using adaptive frequency modulation technology to accurately detect the depth and range of damage; predicting the damage development trend, and then obtaining the detection results.

[0004] The Chinese invention patent with the publication number CN119475989A proposes a method for predicting the vortex-induced vibration response of a double-deck steel truss bridge based on machine learning algorithms. By combining wind tunnel tests, numerical simulations, and machine learning methods, the prediction of the vortex-induced vibration response of the double-deck steel truss bridge is realized, with good stability and high prediction accuracy, thus providing a reference for bridge design, operation, and maintenance.

[0005] The Chinese invention patent with the publication number CN119442849A proposes a method and device for predicting the spatio-temporal evolution of the service performance of a bridge structure based on a graph network, which relates to the technical field of predicting the service performance of a bridge structure. A bridge graph model is constructed, and a basic prediction model for the evolution of the service performance of the bridge structure is constructed. By embedding a parametric graph learning module that adaptively updates the spatio-temporal dependence relationship of bridge structure components, a trained prediction model for the evolution of the service performance of the bridge structure is obtained, thereby capturing the spatio-temporal dependence relationship between the same type of structural components and different types of structural components in the bridge.

[0006] In summary, the prior art has the following objective drawbacks:

[0007] (1) The traditional particle swarm optimization method is prone to falling into local optima, resulting in the model being unable to effectively explore the global optimal solution, affecting the accuracy and stability of anomaly detection;

[0008] (2) Traditional machine learning methods (such as support vector machines and decision trees) rely on manually designed features and are difficult to fully express the latent non-linear relationships in the data. Especially when the data volume is huge and has high-dimensional features, they are easily restricted by overfitting and computational efficiency;

[0009] (3) Conventional convolutional neural networks or fully connected neural networks assume that the relationships between data are uniform, ignoring the local structure and spatio-temporal similarity between nodes, and are difficult to process data with complex similarity and heterogeneity. The classification effect is not precise enough and is easily interfered by noise;

[0010] (4) Traditional neural network methods (such as fully connected neural networks) cannot fully consider the graph structure information between nodes when processing structured data, ignoring the potential spatio-temporal relationships in the data and being difficult to capture complex data relationships;

[0011] (5) The prior art has poor adaptability in different construction stages and environments and cannot automatically adjust parameters according to the actual situation, resulting in unstable performance of the model in complex environments.

[0012] Therefore, in view of the above problems, the present invention proposes a method and system for identifying abnormal data in bridge construction monitoring. Summary of the Invention

[0013] In view of the deficiencies of the prior art, the present invention has developed a method and system for identifying abnormal data in bridge construction monitoring. By constructing a model for identifying abnormal data in bridge construction monitoring, the present invention can monitor complex vibration signal data during bridge construction and better identify abnormal situations.

[0014] The technical solution for the present invention to solve the technical problem is a method for identifying abnormal data in bridge construction monitoring, which is specifically as follows:

[0015] S1. Data collection: Collect bridge monitoring vibration signal data by setting sensors and devices;

[0016] S2. Data processing: Clean, transform, and standardize the collected data;

[0017] S3. Build an abnormal data recognition model for bridge construction monitoring: Build an abnormal data recognition model for bridge construction monitoring based on the graph neural network model, input the processed data into the abnormal data recognition model for bridge construction monitoring for training, and identify abnormal situations.

[0018] In the specific implementation manner, the abnormal data recognition model for bridge construction monitoring includes the following structural components:

[0019] An input layer for receiving analysis results, a chaotic particle swarm optimization initial module for dynamically generating initial parameters, a multi-layer graph convolutional layer for aggregating neighborhood information through the attention mechanism and degree normalization, a topological sensitive kernel function weighting layer for strengthening local similarity features, and a neuron using the Sigmoid activation function as the output layer.

[0020] In the specific implementation manner, the training process of the abnormal data recognition model for bridge construction monitoring is as follows:

[0021] (1) Initialize the parameters of the graph neural network;

[0022] (2) Perform particle update and fitness evaluation;

[0023] (3) Perform node information propagation and update of the graph neural network;

[0024] (4) Perform topological structure sensitive kernel function weighting and feature strengthening;

[0025] (5) Perform iterative training and convergence judgment.

[0026] In the specific implementation manner, the initialization of the parameters of the graph neural network is as follows:

[0027] Use the graph neural network algorithm based on chaotic particle swarm optimization to process the vibration signals of the bridge. When initializing the parameters of the graph neural network, set the state of the particle swarm, and initialize the particle swarm through chaotic mapping on the basis of the particle swarm optimization method, that is, perform chaotic particle swarm optimization. Among them, the chaotic mapping function uses a sine perturbation term to enhance the chaotic characteristics, and the chaotic perturbation intensity coefficient in the chaotic mapping function is set manually according to the number of iterations;

[0028] The calculation formula for chaotic particle swarm optimization is as follows:

[0029] ,

[0030] ,

[0031] Among them, represents the chaotic mapping function, represents the modulo operation that makes the result fall within the interval, represents the th particle's state value at the th chaotic iteration, represents the th particle's state value at the th chaotic iteration, represents the chaotic perturbation intensity coefficient.

[0032] In the specific implementation, the particle update and fitness evaluation are as follows:

[0033] Based on the prediction results of the graph neural network model, calculate the fitness of the particles. By minimizing the fitness function, the model distinguishes normal data from abnormal data. The fitness function is calculated using the cross-entropy loss calculation method. The formula for calculating the fitness is as follows:

[0034] ,

[0035] Among them, represents the th particle 's fitness, represents the number of samples participating in the training, represents the th sample's true label, represents the logarithmic function, with the default base being 10, represents the prediction result of the graph neural network model when the input is the th sample's input features and the parameter is the parameter corresponding to the th particle, is the th sample's input features;

[0036] Use the topology-sensitive kernel function to weight according to the local structural differences between the nodes of the graph neural network, update the parameters of the particles, calculate the update speed of the particles, dynamically adjust the weight of the weighting term in the topology-sensitive kernel function according to the overall fitness distribution of the particle swarm, and then update the position of the particles based on the fitness of the particles, thereby updating the neural network parameters. The formula for updating the position of the particles is as follows:

[0037] ,

[0038] Among them, represents the th particle's position at the th iteration, Indicates the position of the th particle at the th iteration, Indicates the th particle at the th iteration of the velocity, Indicates the learning rate for updating the particle position, Indicates the th particle at the th iteration of the fitness, The item adopts a fitness feedback mechanism.

[0039] In the specific implementation manner, the particle velocity is updated based on the topology structure sensitive kernel function, and the calculation formula is as follows:

[0040] ,

[0041] Among them, Indicates the th particle at the th iteration of the velocity, Indicates the th particle at the th iteration of the velocity, Indicates the th particle at the th iteration of the position, Indicates the inertia weight that controls the proportion of the particle inheriting the original velocity, Indicates the first acceleration constant that reflects the degree of dependence of the particle on its own experience, Indicates the first random number, Indicates the th historical best position of the particle, Indicates the second acceleration constant that reflects the degree of dependence of the particle on its own experience, Indicates the second random number, Indicates the th iteration of the global best particle position, Indicates the weight of the acceleration term, and the weight of the acceleration term is adjusted according to the fitness, Indicates the th neighbor particle set of the particle, Indicates the th particle in the neighbor particle set at the th iteration of the position, is the L2 norm.

[0042] In the specific implementation manner, the node information propagation and update of the graph neural network are specifically as follows:

[0043] In the graph neural network, an information propagation mechanism based on graph convolution operation is adopted. Based on the influence of the features of the current node and its neighbor nodes, a self-enhancing propagation effect is formed, and then iterative hierarchical propagation is carried out to strengthen the local features in the graph structure, thereby capturing the complex relationships in the graph data. Among them, the weight of feature propagation is automatically adjusted according to the similarity between nodes;

[0044] The calculation formula for the node features of the graph neural network is as follows:

[0045] ,

[0046] where, represents the feature of the -th layer and the -th node of the graph neural network, represents the attention coefficient vector, is the transpose of the attention coefficient vector, represents vector concatenation, represents the feature of the -th layer and the -th node of the graph neural network, represents the feature of the -th layer and the -th node of the graph neural network, represents the feature of the -th layer and the -th node of the graph neural network, is the LeakyReLU activation function, represents the set of adjacent nodes of the -th node, represents the degree of the -th node, represents the degree of the -th node, represents a positive integer, represents a positive integer, represents the weight of the graph neural network, represents the bias of the graph neural network, represents the Sigmoid activation function.

[0047] In the specific implementation, topological structure-sensitive kernel function weighting and feature enhancement are performed:

[0048] Through the adoption of the topological structure-sensitive kernel function for feature enhancement, the topological structure-sensitive kernel function is calculated based on the norm between node features, and then the neighborhood feature aggregation weighted by the kernel function is used to enhance the feature response between locally similar nodes and suppress the interference of noise nodes. The enhanced features of the -th layer of the graph neural network are input into the next layer of graph convolution or fully connected layer;

[0049] The calculation formula for feature enhancement is as follows:

[0050] ,

[0051] Wherein, represents the enhancement coefficient, represents the enhanced feature of the -th layer and the -th node of the graph neural network, represents the topology-sensitive kernel function between the -th layer and the -th node and the -th node of the graph neural network.

[0052] In the specific implementation manner, iterative training and convergence judgment are performed:

[0053] Repeat the training process of the abnormal data recognition model for bridge construction monitoring, and calculate the F1 score and the value of the cross-entropy loss function after each iteration;

[0054] After continuously iterating 10 times, if the decrease amplitude of the loss function value is less than 0.001 or the fluctuation range of the F1 score is less than 0.5%, it is determined that the abnormal data recognition model for bridge construction monitoring converges, and the training is terminated; otherwise, continue to iterate until the set maximum number of iterations is reached.

[0055] The present invention also provides an abnormal data recognition system for bridge construction monitoring, which executes an abnormal data recognition method for bridge construction monitoring, specifically as follows:

[0056] Collection and monitoring system: Monitor and collect bridge structure information, hydrological and meteorological information, real-time monitoring information, and maintenance and handover information through a variety of sensors and devices;

[0057] Data resource layer: Integrate the collected data to form a unified data source, and then process and analyze the data source;

[0058] Business application layer: Based on the graph neural network, identify abnormal data in bridge construction monitoring, combine the abnormal monitoring results with the data in the data resource layer for analysis, and display the analysis results in the form of charts, reports, etc.;

[0059] The business application layer internally includes an application support layer, which extracts, transforms, loads, and processes the data in the data resource layer through the application support layer;

[0060] Front-end access layer: The system provides a user-friendly interface and transmits the analysis results of the business application layer to the front end.

[0061] The effects provided in the invention content are only the effects of the embodiments, rather than all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0062] (1) The present invention uses chaotic mapping to initialize the particle swarm, ensuring a uniform distribution of particle states, avoiding the local optimum problem in traditional particle swarm optimization, improving the diversity of parameter initialization and the global exploration ability, enabling the model to better capture spatio-temporal patterns in complex vibration signal data, and significantly enhancing the stability and convergence speed of the training process;

[0063] (2) The present invention uses a topology-sensitive kernel function in particle update and feature enhancement, not only considering the degree relationship of nodes, but also modeling the similarity of nodes, strengthening local similar features, suppressing the interference of noise nodes, improving the sensitivity to abnormal signals, enabling the model to more accurately identify abnormal patterns, especially effectively suppressing false alarm signals during low signal-to-noise ratio periods;

[0064] (3) Through graph convolution operations and attention mechanisms in the present invention, the features of nodes not only depend on themselves, but are also affected by neighboring nodes, gradually strengthening local features, achieving the capture of complex data relationships, enhancing the model's ability to process non-linear and complex structures in bridge monitoring vibration signal data, and improving the accuracy and robustness of abnormal monitoring;

[0065] (4) In the present invention, the chaotic perturbation intensity coefficient and acceleration term weight are dynamically adjusted according to the iterative process, enhancing the adaptability and robustness of the model, automatically adjusting parameters during different construction stages (such as high-risk periods), strengthening the particle aggregation in key areas, avoiding the oscillation problem in traditional methods, and improving the stability of the model in complex environments.

[0066] In summary, the present invention can monitor complex vibration signal data during bridge construction, better identify abnormal situations, and improve the accuracy, efficiency, and robustness of processing monitoring data in actual engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention.

[0068] Figure 1 It is a schematic flow chart of the method of the present invention.

[0069] Figure 2 It is an example diagram of bridge detection vibration signal data.

[0070] Figure 3 It is a three-dimensional dynamic convergence surface diagram of the present invention and the existing method, where Figure 3In (a) is the convergence surface diagram of the method in the present invention, Figure 3 In (b) is the convergence surface diagram of the traditional PSO-GNN, Figure 3 In (c) is the convergence surface diagram of the random forest, Figure 3 In (d) is the convergence surface diagram of the SVM.

[0071] Figure 4 Is the spatio-temporal heat map of the present invention.

[0072] Figure 5 Is the spatio-temporal heat map of the LSTM long short-term memory network warning system.

[0073] Figure 6 Is the spatio-temporal heat map of the traditional vibration analysis technology.

[0074] Figure 7 Is the comparison diagram of the training loss convergence curves of the method in the present invention and the existing methods. Detailed implementation manners

[0075] In order to clearly illustrate the technical features of the present solution, the present invention will be described in detail below through specific implementation manners and in conjunction with its accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the present invention.

[0076] Embodiment 1

[0077] The present invention provides a method for identifying abnormal data in bridge construction monitoring, which is as follows:

[0078] S1. Data collection: Collect bridge monitoring vibration signal data by setting sensors and devices. The collected bridge detection vibration signals are as Figure 2 shown;

[0079] S2. Data processing: Clean, transform, and standardize the collected data;

[0080] S3. Construct an abnormal data identification model for bridge construction monitoring: Based on the graph neural network model, construct an abnormal data identification model for bridge construction monitoring, and input the processed data into the abnormal data identification model for bridge construction monitoring for training to identify abnormal situations.

[0081] In the specific implementation manner, the abnormal data identification model for bridge construction monitoring includes the following structural components:

[0082] The input layer for receiving analysis results, the chaotic particle swarm optimization initial module for dynamically generating initial parameters, the multi-layer graph convolutional layer for aggregating neighborhood information through the attention mechanism and degree normalization, the topological sensitive kernel function weighting layer for strengthening local similarity features, and the neuron using the Sigmoid activation function as the output layer.

[0083] In the specific implementation manner, the training process of the abnormal data recognition model for bridge construction monitoring is as follows:

[0084] (1) Initialize the parameters of the graph neural network;

[0085] (2) Perform particle update and fitness evaluation;

[0086] (3) Perform node information propagation and update of the graph neural network;

[0087] (4) Perform topological structure sensitive kernel function weighting and feature enhancement;

[0088] (5) Perform iterative training and convergence judgment.

[0089] In the specific implementation manner, the initialization of the parameters of the graph neural network is as follows:

[0090] Use the graph neural network algorithm based on chaotic particle swarm optimization to process the vibration signals of the bridge. When initializing the parameters of the graph neural network, set the state of the particle swarm, and initialize the particle swarm through chaotic mapping on the basis of the particle swarm optimization method, that is, perform chaotic particle swarm optimization. Among them, the chaotic mapping function uses a sine perturbation term to enhance the chaotic characteristics, and the chaotic perturbation intensity coefficient in the chaotic mapping function is set manually according to the number of iterations;

[0091] Set the state representation of the particle swarm as:

[0092] ,

[0093] The position of the particle is initialized as:

[0094] ,

[0095] Among them, represents the state of the th particle, including the weights and biases of the graph neural network; represents the number of the particle swarm, represents the initial position of the th particle, represents the chaotic mapping function, and its output non-uniform distribution simulates the time-varying characteristics of the bridge monitoring vibration signal;

[0096] The calculation formula of chaotic particle swarm optimization is as follows:

[0097] ,

[0098] ,

[0099] Among them, represents the chaotic mapping function, Indicates that the result is located in The modulo operation of the interval, Indicates the th particle at the state value during the Indicates the th particle at the state value during the Indicates the chaos perturbation intensity coefficient;

[0100] ,

[0101] Among them, Indicates the maximum number of iterations of the chaos iteration, Indicates the index of the current chaos iteration number, Indicates the chaos perturbation intensity coefficient during the

[0102] In particle swarm optimization, each particle represents a solution, which explores the problem space by adjusting its own position and is influenced by its own historical experience (personal best position) and group experience (global best position).

[0103] The chaotic particle swarm optimization uses the method of chaotic mapping on the basis of traditional particle swarm optimization. The chaotic sequence has good ergodicity and irregularity, so it can effectively avoid the local optimum problem in the traditional particle swarm optimization algorithm. By using chaotic mapping during the initialization of the particle position, it can ensure that the state distribution of the particle swarm is more uniform, thereby improving the diversity and global exploration ability of the search process;

[0104] The initialization method of the graph neural network based on chaotic particle swarm optimization can ensure that the state distribution of particles is more uniform, thus avoiding the local optimum problem in traditional particle swarm optimization, and further better capturing spatio-temporal patterns from complex vibration signal data.

[0105] In the specific implementation manner, the particle update and fitness evaluation are as follows:

[0106] Based on the prediction result of the graph neural network model, the fitness of the particle is calculated. By minimizing the fitness function, the model can distinguish normal data and abnormal data. The fitness function is calculated by using the calculation method of cross-entropy loss. In the present invention, a topology structure-sensitive kernel function is used to weight according to the local structure difference between the nodes of the graph neural network, so that the position update of each particle not only depends on the historical optimal solution and the global optimal solution, but also considers the local similarity relationship between particles for parameter update. The calculation formula of the fitness is as follows:

[0107] ,

[0108] Among them, represents the th particle 's fitness, represents the number of samples participating in the training, represents the th sample's true label, represents the logarithmic function, with a default base of 10, represents the prediction result of the graph neural network model when the input is the input feature of the th sample and the parameter is the parameter corresponding to the th particle, is the input feature of the th sample;

[0109] By minimizing the fitness function, it is ensured that the model can accurately distinguish normal and abnormal vibration signals, and can effectively evaluate the model's ability to identify abnormal signals; in addition, when the particles are updating parameters, they can more precisely capture the characteristics of the graph structure data, enhance the diversity and accuracy of the particle swarm search, and avoid the local optimum problem encountered in the traditional particle swarm optimization method;

[0110] The topological structure-sensitive kernel function is used to weight according to the local structure differences between the nodes of the graph neural network, update the parameters of the particles, calculate the update speed of the particles, dynamically adjust the weight of the weighting term in the topological structure-sensitive kernel function according to the overall fitness distribution of the particle swarm, and then update the position of the particles based on the fitness of the particles, and further update the neural network parameters. The calculation formula for updating the position of the particles is as follows:

[0111] ,

[0112] Among them, represents the position of the th particle at the th iteration, represents the position of the th particle at the th iteration, represents the speed of the th particle at the th iteration, represents the learning rate for updating the particle position, represents the fitness of the th particle at the th iteration, The

[0113] The topological structure-sensitive kernel function is a function that can weight node features according to the topological relationship between nodes. It not only considers the degree relationship of nodes but also can evaluate the similarity between adjacent nodes, realizing the enhancement of local features in the graph neural network and being able to better capture complex structured data patterns.

[0114] In the specific implementation, the particle velocity is updated based on the topological structure-sensitive kernel function, and the calculation formula is as follows:

[0115] ,

[0116] where, represents the velocity of the -th particle at the -th iteration, represents the velocity of the -th particle at the -th iteration, represents the position of the -th particle at the -th iteration, represents the inertia weight that controls the proportion of the particle inheriting the original velocity, represents the first acceleration constant that reflects the degree of dependence of the particle on its own experience, represents the first random number, represents the historical best position of the -th particle, represents the second acceleration constant that reflects the degree of dependence of the particle on its own experience, represents the second random number, represents the global best particle position at the -th iteration, represents the weight of the acceleration term, and the weight of the acceleration term is adjusted according to the fitness, represents the -th particle's neighbor particle set, and the neighborhood particle set is selected according to the physical topology of the sensor, represents the position of the particle in the -th neighbor particle set at the -th iteration, is the L2 norm, is set to 0.2, is set to 0.3, is set to 0.2, is set to 0.5, is set to 0.2;

[0117] The weight of the acceleration term is dynamically adjusted according to the overall fitness distribution of the particle swarm. When the fitness of the particle swarm varies greatly, that is, when there are obvious sub-optimal regions, the weight of the acceleration term is increased to strengthen local search. For example, during the casting stage of the closure section (high-risk period), the weight of the acceleration term is automatically increased to 0.8 to strengthen the particle aggregation in the key regions (such as the cantilever end). Conversely, the weight of the acceleration term is decreased to avoid oscillations. The calculation formula is as follows:

[0118] ,

[0119] where, represents the dynamic adjustment factor, represents the fitness of the th particle in the th iteration, represents the fitness of the th particle in the th iteration, represents the maximum fitness value of all particles in the th iteration, is set to 0.8.

[0120] In the specific implementation manner, the node information propagation and update of the graph neural network are as follows:

[0121] In the graph neural network, an information propagation mechanism based on graph convolution operations is adopted. Based on the influence of the features of the current node and its neighbor nodes, a self-enhancing propagation effect is formed, and then iterative hierarchical propagation is carried out to strengthen the local features in the graph structure, thereby capturing the complex relationships in the graph data. Among them, the weight of feature propagation is automatically adjusted according to the similarity between nodes;

[0122] When performing the node information propagation and update of the graph neural network, the weight of feature propagation can be automatically adjusted according to the similarity between nodes, enhancing the modeling ability of the graph neural network model for complex data relationships and improving the accuracy and robustness of anomaly detection;

[0123] The calculation formula for the node features of the graph neural network is as follows:

[0124] ,

[0125] where, represents the feature of the th node in the th layer of the graph neural network, represents the attention coefficient vector, is the transpose of the attention coefficient vector, represents vector concatenation, represents the feature of the th node in the th layer of the graph neural network, Represents the feature of the -th layer and the -th node of the graph neural network, Represents the feature of the -th layer and the -th node of the graph neural network, is the LeakyReLU activation function, Represents the set of adjacent nodes of the -th node, Represents the degree of the -th node, Represents the degree of the -th node, Represents a positive integer, Represents a positive integer, Represents the weight of the graph neural network, Represents the bias of the graph neural network, Represents the Sigmoid activation function;

[0126] The degree of a node refers to the number of neighbor nodes connected to this node. In a graph neural network, the degree of a node reflects the importance of this node in the graph. Nodes with a larger degree usually have more information transmission channels. Characterizes the degree normalization term, whose function is to suppress the dominant effect of high-degree nodes. For example, the weight ratio between the mid-span sensor (degree = 8) and the side-span sensor ( = 3) is reduced from 2.67 to 1.0, avoiding the normal vibration in the mid-span from masking the abnormality at the side support;

[0127] Characterizes the attention mechanism, whose function is to learn the vibration transmission path between nodes. For example, at the tower-beam node, the attention weight automatically enhances the correlation strength between the top of the tower and the anchor node (the measured weight ratio reaches 1:0.3), reflecting the main path of cable force transmission.

[0128] In the specific implementation manner, topological structure-sensitive kernel function weighting and feature enhancement are performed:

[0129] Through the use of a topological structure-sensitive kernel function for feature enhancement, the topological structure-sensitive kernel function is calculated based on the norm between node features, and then through the kernel function-weighted neighborhood feature aggregation, the feature response between locally similar nodes is enhanced, and the interference of noise nodes is suppressed. The enhanced feature of the -th layer of the graph neural network is input to the next layer of graph convolution or fully connected layer;

[0130] The method of weighting the topology - sensitive kernel function not only considers the degree relationship between nodes, but also models the similarity of nodes, improving the processing ability of graph neural networks for the non - linear and complex structures in bridge monitoring vibration signal data. It enables the graph neural network model to pay more attention to local features during information propagation. For the local structural defects and non - linear fluctuations in bridge monitoring vibration signals, it can strengthen the connection between similar signals and improve the sensitivity to abnormal signals. The calculation formula of the topology - sensitive kernel function is as follows:

[0131] ,

[0132] where, represents the topology - sensitive kernel function between the -th node and the -th node in the -th layer of the graph neural network, represents the scale parameter of the kernel function, represents the feature of the -th node in the -th layer of the graph neural network, represents the feature of the -th node in the -th layer of the graph neural network, is set to 0.1; Explicitly implement degree normalization to balance the dominant effect of high - degree nodes and achieve joint modeling of degree relationship and feature similarity; Characterizes the degree of feature difference and is used to detect local vibration anomalies. For example, when the feature difference between adjacent nodes exceeds a preset threshold (such as 0.3), the model may determine it as an anomaly (corresponding to non - uniform settlement above 0.2 mm).

[0133] The calculation formula for feature enhancement is as follows:

[0134] ,

[0135] where, represents the enhancement coefficient, is set to 0.3, represents the enhanced feature of the -th node in the -th layer of the graph neural network, which represents the local features aggregated by weighting through the topology - sensitive kernel function. It enhances the response to local abnormal patterns (such as local vibration anomalies of bridges) by fusing the similar features of neighboring nodes and suppresses the influence of noise nodes at the same time.

[0136] In the specific implementation, iterative training and convergence judgment are carried out:

[0137] Repeat the training process of the abnormal data recognition model for bridge construction monitoring, and calculate the F1 score and the value of the cross-entropy loss function after each iteration;

[0138] After 10 consecutive iterations, if the decrease in the value of the loss function is less than 0.001 or the fluctuation range of the F1 score is less than 0.5%, it is determined that the abnormal data recognition model for bridge construction monitoring converges and the training is terminated; otherwise, continue the iteration until the set maximum number of iterations is reached. The maximum number of iterations is set to 200.

[0139] Embodiment 2

[0140] A system for identifying abnormal data in bridge construction monitoring executes a method for identifying abnormal data in bridge construction monitoring, which is as follows:

[0141] Data acquisition and monitoring system: Monitor and collect bridge structure information, hydrological and meteorological information, real-time monitoring information, and maintenance and handover information through various sensors and devices;

[0142] The monitoring of bridge structure information includes strain detection (using distributed fiber optic grating sensors, arranged along the key sections of the main girder and piers to capture concrete cracking or steel structure fatigue signals in real time), displacement monitoring (deploying Beidou / GNSS high-precision positioning terminals, combined with inclinometers, to monitor bridge deflection, bearing slip, and settlement), and vibration monitoring (installing a triaxial accelerometer array, with a sampling frequency of 200 Hz to identify abnormal vibration modes);

[0143] The monitoring of hydrological and meteorological information includes water level and flow velocity monitoring, water quality monitoring (monitoring chloride ion concentration and hydrogen ion concentration in water), wind speed monitoring, temperature monitoring, humidity monitoring, and lightning monitoring;

[0144] Real-time monitoring includes vehicle load detection and emergency monitoring;

[0145] Maintenance and handover information includes structural data, environmental data, and operation and maintenance records obtained from the spatio-temporal database;

[0146] Data resource layer: Integrate the collected data to form a unified data source, and then process and analyze the data source;

[0147] Business application layer: Identify abnormal data in bridge construction monitoring based on graph neural networks, combine the abnormal monitoring results with the data in the data resource layer for analysis, and display the analysis results in the form of charts, reports, etc.;

[0148] The business application layer internally includes an application support layer, which extracts, transforms, loads, and processes the data in the data resource layer;

[0149] Front-end access layer: The system provides a user-friendly interface and transmits the analysis results of the business application layer to the front end.

[0150] In addition, to ensure the security of the system, a unified identity authentication and management mechanism is adopted to control users' access rights to the system.

[0151] Example 3

[0152] As Figure 3 shown, by comparing the search processes of different optimization algorithms in the high-dimensional parameter space through three-dimensional dynamic convergence surfaces, the improvement effect of chaotic particle swarm optimization initialization on the training stability of neural networks is verified. The technology of the present invention is compared with traditional particle swarm optimization, random forest, and support vector machine. The experimental results show that the convergence surface of the technology of the present invention is smoother and reaches a lower loss value, indicating that the chaotic initialization strategy can effectively avoid the parameter search from falling into local optima and at the same time reduce the oscillation amplitude during the training process, making the model show stronger robustness and stability in the complex parameter environment of actual engineering.

[0153] Example 4

[0154] As Figures 4 to 6 shown, by using the method of combining spatio-temporal heat maps with contour line analysis to analyze the spatio-temporal distribution characteristics of the anomaly probability of different algorithms in bridge monitoring, and by comparing the capabilities of various methods to capture the abnormal evolution during the construction process, the performance of the technology of the present invention is verified with that of the long short-term memory network warning system and traditional vibration analysis technology. The experimental data of the heat map show that the technology of the present invention can more accurately identify abnormal patterns and effectively suppress false alarm signals during low signal-to-noise ratio periods, indicating the strengthening ability of the topology-sensitive kernel function for spatio-temporal correlation features, enabling the model to distinguish real anomalies from environmental noise.

[0155] Example 5

[0156] As Figure 7 shown, to verify the influence of chaotic particle swarm optimization initialization on the training process of graph neural networks, by comparing the convergence characteristics of methods such as traditional particle swarm optimization, random forest, and support vector machine, the experimental results show that the training loss curve of the technology of the present invention not only converges faster but also reaches a lower loss value in the end, and the curve fluctuation amplitude is significantly smaller than that of traditional methods, indicating that the chaotic initialization strategy can effectively avoid the parameter optimization from falling into local optima and at the same time significantly improve the stability of the training process, making the model show stronger adaptability to complex bridge monitoring data.

[0157] Although the specific implementation modes of the invention are described above in conjunction with the drawings, it is not a limitation to the protection scope of the present invention. Based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A method for identifying abnormal data in bridge construction monitoring, characterized in that, It includes the following steps: S1. Data acquisition: Collect bridge monitoring vibration signal data by setting sensors and devices; S2. Data processing: Clean, transform, and standardize the collected data; S3. Construct an abnormal data recognition model for bridge construction monitoring: Based on the graph neural network model, construct an abnormal data recognition model for bridge construction monitoring, input the processed data into the abnormal data recognition model for bridge construction monitoring for training, and identify abnormal situations; The training process of the abnormal data recognition model for bridge construction monitoring is as follows: (1) Initialize the parameters of the graph neural network; (2) Perform particle update and fitness evaluation; (3) Perform node information propagation and update of the graph neural network; (4) Perform topological structure sensitive kernel function weighting and feature enhancement; (5) Perform iterative training and convergence judgment; The specific initialization of the parameters of the graph neural network is as follows: Use the graph neural network algorithm based on chaotic particle swarm optimization to process the vibration signals of the bridge. When initializing the parameters of the graph neural network, set the state of the particle swarm, and initialize the particle swarm through chaotic mapping on the basis of the particle swarm optimization method, that is, perform chaotic particle swarm optimization. Among them, the chaotic mapping function uses a sine perturbation term to enhance the chaotic characteristics, and the chaotic perturbation intensity coefficient in the chaotic mapping function is set manually according to the number of iterations; The calculation formula of chaotic particle swarm optimization is as follows: ， ， Among them, represents the chaotic mapping function, represents the modulo operation that makes the result fall within the interval, represents the state value of the th particle at the th chaotic iteration, represents the state value of the th particle at the th chaotic iteration, represents the chaotic perturbation intensity coefficient.

2. The method for identifying abnormal data in bridge construction monitoring according to claim 1, wherein The abnormal data recognition model for bridge construction monitoring consists of the following structural components: The input layer for receiving analysis results, the initial module of chaotic particle swarm optimization for dynamically generating initial parameters, the multi-layer graph convolution layer for aggregating neighborhood information through the attention mechanism and degree normalization, the topological structure sensitive kernel function weighting layer for strengthening local similarity features, and the output layer using neurons with Sigmoid activation functions.

3. The abnormal data recognition method for bridge construction monitoring according to claim 2, characterized in that, The particle update and fitness evaluation are specifically as follows: Calculate the fitness of the particles based on the prediction results of the graph neural network model. Minimize the fitness function to enable the model to distinguish normal data and abnormal data. The fitness function is calculated using the calculation method of cross-entropy loss. The calculation formula of the fitness is as follows: ， Among them, represents the th particle 's fitness, represents the number of samples participating in training, represents the th sample's true label, represents the logarithmic function, with a default base of 10, represents the prediction result of the graph neural network model when the input is the input features of the th sample and the parameters are the parameters corresponding to the th particle, is the input feature of the th sample; Use the topological structure sensitive kernel function to weight according to the local structure differences between the nodes of the graph neural network, update the parameters of the particles, calculate the update speed of the particles, dynamically adjust the weight of the weighting term in the topological structure sensitive kernel function according to the overall fitness distribution of the particle swarm, and then update the position of the particles based on the fitness of the particles, thereby updating the neural network parameters. The calculation formula for updating the position of the particles is as follows: ， Among them, represents the position of the th particle at the th iteration, represents the position of the th particle at the th iteration, represents the velocity of the th particle at the th iteration, represents the learning rate for updating the particle position, represents the fitness of the th particle at the th iteration, The term adopts a fitness feedback mechanism.

4. The abnormal data recognition method for bridge construction monitoring according to claim 3, wherein Update the particle velocity based on the topological structure sensitive kernel function. The calculation formula is as follows: ， Among them, represents the velocity of the th particle at the th iteration, represents the velocity of the th particle at the th iteration, represents the position of the th particle at the th iteration, represents the inertia weight that controls the proportion of the particle inheriting the original velocity, represents the first acceleration constant that reflects the degree of dependence of the particle on its own experience, represents the first random number, represents the th particle's historical best position, represents the second acceleration constant that reflects the degree of dependence of the particle on its own experience, represents the second random number, represents the th iteration's global best particle position, represents the weight of the acceleration term, and the weight of the acceleration term is adjusted according to the fitness, represents the th particle's neighbor particle set, represents the th particle in the neighbor particle set at the th iteration's position, is the L2 norm.

5. The method for identifying abnormal data in bridge construction monitoring according to claim 4, wherein The node information propagation and update of the graph neural network are specifically as follows: In the graph neural network, adopt an information propagation mechanism based on graph convolution operations. Based on the influence of the features of this node and the neighbor nodes, form a self-enhancing propagation effect, and then perform iterative hierarchical propagation to strengthen the local features in the graph structure, thereby capturing the complex relationships in the graph data. Among them, the weight of feature propagation is automatically adjusted through the similarity between nodes; The calculation formula of the node features of the graph neural network is as follows: ， Among them, represents the feature of the th node in the th layer of the graph neural network, represents the attention coefficient vector, is the transpose of the attention coefficient vector, represents vector concatenation, represents the feature of the th node in the th layer of the graph neural network, represents the feature of the th node in the th layer of the graph neural network, is the LeakyReLU activation function, represents the set of adjacent nodes of the th node, represents the degree of the th node, represents the degree of the th node, represents a positive integer, represents a positive integer, represents the weight of the graph neural network, represents the bias of the graph neural network, represents the Sigmoid activation function.

6. The abnormal data identification method for bridge construction monitoring according to claim 5, characterized in that, Perform topological structure sensitive kernel function weighting and feature enhancement: By adopting a topology-sensitive kernel function for feature enhancement, calculating the topology-sensitive kernel function based on the norm between node features, and then aggregating neighborhood features weighted by the kernel function to enhance the feature response between locally similar nodes and suppress the interference of noise nodes, the enhanced features of the layer are input into the next layer of graph convolution or fully connected layer; The calculation formula for feature enhancement is as follows: ， Among them, represents the enhancement coefficient, represents the enhanced feature of the -th layer and the -th node of the graph neural network, represents the topology-sensitive kernel function between the -th layer and the -th node and the -th node of the graph neural network.

7. The method for identifying abnormal data in bridge construction monitoring according to claim 6, characterized in that, Perform iterative training and convergence judgment: Repeat the training process of the abnormal data identification model for bridge construction monitoring, and calculate the F1 score and cross-entropy loss function value after each iteration; After 10 consecutive iterations, if the decrease in the loss function value is less than 0.001 or the fluctuation range of the F1 score is less than 0.5%, it is determined that the abnormal data identification model for bridge construction monitoring converges, and the training is terminated. Otherwise, continue to iterate until the set maximum number of iterations is reached.

8. An abnormal data identification system for bridge construction monitoring, which executes an abnormal data identification method for bridge construction monitoring described in any one of claims 1-7, specifically as follows: Collection and monitoring system: Monitor and collect bridge structure information, hydrological and meteorological information, real-time monitoring information, and maintenance and handover information through various sensors and devices; Data resource layer: Integrate the collected data to form a unified data source, and then process and analyze the data source; Business application layer: Identify abnormal data for bridge construction monitoring based on a graph neural network, combine the abnormal monitoring results with the data in the data resource layer for analysis, and display the analysis results in the form of charts and reports; The business application layer internally includes an application support layer, which extracts, transforms, loads, and processes the data in the data resource layer through the application support layer; Front-end access layer: The system provides a user-friendly interface and transmits the analysis results of the business application layer to the front end.

Images