Fabricated bridge and culvert node mechanical property analysis method and system based on neural network

Abstract

The invention provides an assembly type bridge and culvert node mechanical property analysis method and system based on a neural network, and relates to the technical field of data processing, and the method comprises the steps: obtaining real-time monitoring data of an assembly type bridge and culvert node, the real-time monitoring data comprising strain data, vibration data and load data; performing feature extraction and optimization on the real-time monitoring data of the fabricated bridge and culvert nodes to obtain a nonlinear feature set; performing mechanical property evaluation on the nonlinear feature set to obtain a real-time mechanical property evaluation result; performing model construction and optimization based on the real-time mechanical property evaluation result and preset historical monitoring data to obtain an optimized mechanical property change trend prediction model; and sending the real-time monitoring data to the optimized mechanical property change trend prediction model for prediction to obtain the mechanical property of the fabricated bridge and culvert node in a preset time period. The precision, timeliness and reliability of mechanical property monitoring of the fabricated bridge and culvert joint are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to a method and system for analyzing the mechanical properties of assembled bridge and culvert nodes based on a neural network. Background Art

[0002] With the rapid development of urban infrastructure, assembled bridge and culvert, as an important underground structure form, is widely used in bridge engineering. As a key component of it, the assembled bridge and culvert node bears important mechanical functions and is easily affected by factors such as geological changes, traffic loads, and construction quality. Therefore, accurately evaluating the mechanical properties of assembled bridge and culvert nodes and timely carrying out reinforcement and repair have become important issues to ensure project safety and extend the service life of the structure.

[0003] Currently, many engineering projects adopt traditional monitoring methods, such as manual inspections and conventional instrument measurements. These methods usually rely on manual experience and lack the ability of real-time monitoring and dynamic evaluation. Although some sensor-based monitoring systems have been proposed, the existing technologies mostly rely on simple data collection and analysis methods, and cannot effectively process complex mechanical property data and non-linear characteristics, nor can they timely and accurately evaluate the long-term change trend and damage risk of the structure. This has led to many potential structural problems not being discovered in the early stage, thus increasing the maintenance cost and safety hazards.

[0004] Therefore, there is an urgent need for a method and system for analyzing the mechanical properties of assembled bridge and culvert nodes based on a neural network to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for analyzing the mechanical properties of assembled bridge and culvert nodes based on a neural network to improve the above problems. To achieve the above purpose, the technical solutions adopted by the present invention are as follows:

[0006] In a first aspect, the present application provides a method for analyzing the mechanical properties of assembled bridge and culvert nodes based on a neural network, including:

[0007] Obtaining real-time monitoring data of the assembled bridge and culvert nodes, where the real-time monitoring data includes strain data, vibration data, and load data;

[0008] Performing feature extraction and optimization on the real-time monitoring data of the assembled bridge and culvert nodes to obtain a non-linear feature set;

[0009] Evaluating the mechanical properties of the non-linear feature set to obtain a real-time mechanical property evaluation result;

[0010] Based on the real-time mechanical property evaluation result and preset historical monitoring data, constructing and optimizing a model to obtain an optimized prediction model for the change trend of mechanical properties;

[0011] Send the real-time monitoring data to the optimized prediction model of the mechanical property change trend for prediction, and obtain the mechanical properties of the assembled bridge and culvert node in a preset time period.

[0012] In a second aspect, the present application also provides a mechanical property analysis system for assembled bridge and culvert nodes based on a neural network, including:

[0013] An acquisition unit, configured to acquire real-time monitoring data of the assembled bridge and culvert node, where the real-time monitoring data includes strain data, vibration data, and load data;

[0014] A processing unit, configured to perform feature extraction and optimization on the real-time monitoring data of the assembled bridge and culvert node to obtain a non-linear feature set;

[0015] An evaluation unit, configured to perform mechanical property evaluation on the non-linear feature set to obtain a real-time mechanical property evaluation result;

[0016] A construction unit, configured to perform model construction and optimization based on the real-time mechanical property evaluation result and preset historical monitoring data to obtain an optimized prediction model of the mechanical property change trend;

[0017] A prediction unit, configured to send the real-time monitoring data to the optimized prediction model of the mechanical property change trend for prediction, and obtain the mechanical properties of the assembled bridge and culvert node in a preset time period.

[0018] The beneficial effects of the present invention are as follows:

[0019] By collecting real-time strain, vibration, and load data and combining algorithms such as deep belief network, gradient boosting machine, and long short-term memory network, the present invention realizes the efficient extraction and prediction of the mechanical properties of assembled bridge and culvert nodes. Specifically, the present invention uses a feature extraction model to optimize the monitoring data, performs mechanical property evaluation using a non-linear feature set, constructs a prediction model of the mechanical property change trend based on historical data and real-time monitoring results, and finally realizes the accurate evaluation of the long-term health status of assembled bridge and culvert nodes. In addition, the present invention also combines adversarial network, isolation forest algorithm, and variational autoencoder for anomaly detection and damage analysis, and optimizes the reinforcement strategy based on fuzzy rule reasoning and multi-objective optimization function to ensure the maximum utilization of reinforcement resources. The present invention can effectively make up for the deficiencies of the prior art and improve the accuracy, timeliness, and reliability of the mechanical property monitoring of assembled bridge and culvert nodes.

[0020] Other features and advantages of the present invention will be described in the subsequent specification, and in part will be obvious from the specification, or will be understood by implementing the embodiments of the present invention. The objectives and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in the written specification, claims, and drawings. Brief Description of the Drawings

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0022] Figure 1 Schematic flow diagram of the method for analyzing the mechanical properties of assembled bridge and culvert nodes based on neural network described in the embodiments of the present invention;

[0023] Figure 2 Schematic structural diagram of the system for analyzing the mechanical properties of assembled bridge and culvert nodes based on neural network described in the embodiments of the present invention.

[0024] In the figure: 701, acquisition unit; 702, processing unit; 703, evaluation unit; 704, construction unit; 705, prediction unit. Detailed Embodiments

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. Usually, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but only represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0026] It should be noted that: similar reference numerals and letters indicate similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of the present invention, terms such as "first" and "second" are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0027] Embodiment 1:

[0028] This embodiment provides a method for analyzing the mechanical properties of assembled bridge and culvert joints based on a neural network.

[0029] Refer to Figure 1 , in the figure, it is shown that this method includes step S1, step S2, step S3, step S4 and step S5.

[0030] Step S1: Obtain the real-time monitoring data of the assembled bridge and culvert joint, where the real-time monitoring data includes strain data, vibration data and load data;

[0031] It can be understood that in this step, multi-dimensional real-time monitoring information including strain data, vibration data and load data of the assembled bridge and culvert joint is first collected through sensors arranged on the bridge and culvert joint. Among them, the assembled bridge and culvert joint usually refers to the joint part between different parts of the assembled bridge and culvert structure. In the construction of assembled bridges and culverts, bridges and culverts are composed of multiple precast components (such as boxes, covers, side walls, etc.) combined at the site through connection joints. These connection parts (i.e., joints) are very critical parts of the structure, and they are responsible for transmitting forces, stresses and loads between different parts. These data are the physical responses of the assembled bridge and culvert joint under different working conditions, reflecting the dynamic changes of the structure under external loads and internal forces. Strain data can help us monitor the deformation of materials and then infer the stress state of the structure; vibration data can reveal possible abnormal vibration modes during the operation of the structure, such as the expansion of cracks or the loosening of joint connections; load data mainly reflects the magnitude and direction of the external forces acting on the bridge and culvert joint.

[0032] Step S2: Extract and optimize the features of the real-time monitoring data of the assembled bridge and culvert joint to obtain a non-linear feature set;

[0033] It can be understood that in this step, representative non-linear features are extracted from the real-time monitoring data of the assembled bridge and culvert joint, and these features can accurately reflect the changes in the mechanical properties of the structure. In order to deal with the complex mechanical behaviors and responses of bridge and culvert joints, appropriate algorithms need to be used for feature extraction and these features need to be optimized to ensure the accuracy and predictability of subsequent analysis. In this step, step S2 includes step S21, step S22, step S23 and step S24.

[0034] Step S21: Perform factor analysis on the real-time monitoring data of the assembled bridge and culvert joint to obtain a set of mechanical feature factors for extracting the assembled bridge and culvert joint;

[0035] It is understandable that this step extracts a set of fewer and representative mechanical characteristic factors from the real-time monitoring data of the assembled bridge and culvert joints. These factors can reveal the main mechanical behaviors of the assembled bridge and culvert joints, such as the potential relationships between multi-dimensional data like strain, vibration mode, and load. Since the real-time monitoring data of the assembled bridge and culvert joints often includes various complex interaction relationships, factor analysis can effectively reduce the data dimension and reveal the main influencing factors in this data.

[0036] First, based on the standardized monitoring data, calculate the covariance matrix between the data of each dimension. The covariance matrix describes the linear relationship between each data dimension. For the monitoring data of the assembled bridge and culvert joints, the covariance matrix helps us identify the relationships between variables such as strain, vibration, and load;

[0037] Next, perform eigenvalue decomposition to decompose the covariance matrix into eigenvalues and eigenvectors. The eigenvalues reflect the variance of the data in the direction of the corresponding eigenvectors, and the eigenvectors represent the directions of each factor. Larger eigenvalues correspond to the main factors, and smaller eigenvalues correspond to noise or unimportant factors; for example, factors related to the mechanical properties of the assembled bridge and culvert joints extracted through eigenvalue decomposition may include stress, deformation, and vibration mode, etc. Each factor corresponds to an eigenvector, representing a combination of different monitoring data dimensions.

[0038] Then, use the rotation method of variance maximization rotation to rotate the factors to obtain the mechanical characteristic factors. Through factor analysis, the original data is transformed from multi-dimensional to a few key factors, simplifying the complexity of subsequent analysis.

[0039] Step S22: Perform unsupervised learning on the set of mechanical characteristic factors through the autoencoder of the deep belief network to obtain the non-linear characteristics of the set of mechanical characteristic factors;

[0040] It is understandable that in this step, the input data is the set of mechanical characteristic factors obtained by factor analysis. These factors include various mechanical characteristics such as stress, deformation, and vibration mode. The dimension of the data is usually high, and there may be complex non-linear relationships. The autoencoder will gradually compress and learn the low-dimensional representation of these characteristics through multiple hidden layers;

[0041] The main task of the encoder part is to map the input data to a low-dimensional space. In this process, the encoder performs non-linear mapping on the data through multiple neural network layers, and the output is the compressed hidden layer feature, called "encoding". The goal of the encoder is to learn a low-dimensional feature that can effectively represent the original data by minimizing the reconstruction error between the input and the output. For example: Assume that the input data is the characteristics of strain and vibration mode. Through the encoder network, these data are compressed into a low-dimensional representation containing important information about the node characteristics.

[0042] The task of the decoder is to recover the input data from the encoded low-dimensional representation. Through the decoder, the information loss during the encoding process can be examined, and optimization adjustments can be made to make the reconstructed data as close as possible to the original input. The output of the decoder is the non-linear features learned through backpropagation.

[0043] The training objective of the autoencoder is to minimize the error between the input data and the reconstructed data. The commonly used error metric is the mean squared error, that is, to calculate the difference between the reconstructed data and the input data, and continuously adjust the network parameters through the backpropagation algorithm to optimize the representation of the data; in the deep belief network, the ReLU activation function of the hidden layer can effectively capture the non-linear relationships in the input data. For example, when dealing with strain and vibration data, these activation functions help the network learn the stress distribution patterns of nodes under high-pressure conditions, or the deformation trends that may occur during long-term use;

[0044] Moreover, assume that under a certain working condition, the strain data of the assembled bridge and culvert nodes shows a non-linear growth trend, which cannot be captured by traditional linear analysis methods. Through the autoencoder, the model can learn the non-linear features in the strain data, and then provide more accurate input data for subsequent mechanical property evaluation. For example, local plastic deformation may occur in the nodes under certain ultimate loads, and these changes may be ignored in traditional analysis methods. However, through the non-linear feature extraction of the autoencoder, these changes can be effectively captured and used for performance evaluation.

[0045] Step S23: Obtain an optimized mechanical feature set according to the non-linear feature processing of the gradient boosting machine on the mechanical feature factor set;

[0046] It can be understood that the gradient boosting machine first constructs a simple initial model (usually a relatively shallow decision tree) to predict the mechanical feature factors. The prediction results of this model will generate residuals (that is, the difference between the predicted value and the true value). Next, the gradient boosting machine trains a new decision tree by calculating the residuals of the current prediction. The new decision tree will specifically learn the residuals, aiming to correct the errors in the previous prediction through a new round of learning. This process will be repeated multiple times, and a new decision tree will be trained in each round to reduce the residuals of the previous step. The weights of each tree are updated through the gradient descent method, and the final model is the weighted sum of the prediction results of all decision trees. Through the gradient boosting machine, we can further process and optimize these non-linear features. Identify which features (such as specific strain patterns or vibration frequencies) are most critical to the structural performance of the nodes, and gradually adjust the model according to the residuals to make the prediction of the mechanical features more accurate.

[0047] Step S24: Aggregate all the decision trees in the gradient boosting machine based on the ensemble learning method to obtain the weighted average of multiple decision trees, and obtain an optimized non-linear feature set.

[0048] It can be understood that in the gradient boosting machine, each decision tree is trained based on the residuals predicted in the previous step. Therefore, each tree can learn different error patterns. All the decision trees together form an ensemble model, and each tree contributes to the final prediction result. Ensemble learning combines the outputs of multiple decision trees through weighted averaging to generate the final prediction result. The weight of each tree is usually determined according to its performance on the training set, and the tree with better performance will obtain a higher weight. By integrating the prediction results of multiple decision trees, ensemble learning can effectively reduce the overfitting risk of a single decision tree and enhance the generalization ability of the model on new data. The bias learned by each decision tree is corrected by the learning results of other trees, and finally a more stable and reliable prediction result is obtained. In a specific example, data such as the strain, vibration, and load of the prefabricated bridge and culvert nodes are used as input features, and multiple decision trees are generated through the gradient boosting machine. At some nodes, due to excessive load, there may be a relatively complex relationship between the vibration mode and the strain characteristics. Through the ensemble learning method, we weight and integrate the prediction results of all the decision trees to obtain the final optimized feature set. The advantage of this process is that ensemble learning can extract the most valuable features by weighted averaging multiple decision trees, thereby helping the subsequent mechanical property evaluation and prediction model to provide more accurate input data.

[0049] Step S3: Perform a mechanical property evaluation on the non-linear feature set to obtain a real-time mechanical property evaluation result;

[0050] It can be understood that this step uses the non-linear feature set for mechanical property evaluation, which can provide accurate and real-time mechanical property prediction. Compared with traditional evaluation methods, this method based on machine learning and deep learning can better capture the mechanical response of prefabricated bridge and culvert nodes under complex working conditions, improving the accuracy and real-time performance of the evaluation. In addition, the model based on dynamic update can adapt to environmental changes in real time, so that the evaluation result always remains in the latest state, providing a reliable basis for structural health monitoring and fault prediction. In this step, Step S3 includes Step S31, Step S32, and Step S33.

[0051] Step S31: Map the non-linear feature set to a preset high-dimensional space based on the kernel principal component analysis method to obtain the non-linear feature set in the high-dimensional space;

[0052] It can be understood that in this step, the non - linear feature set of the assembled bridge and culvert node is first input into the kernel principal component analysis model. This feature set usually contains non - linear features extracted from node monitoring data (such as strain, vibration, and load). These features reflect the mechanical behavior of the node under complex working conditions. Through the radial basis kernel function, the non - linear features are mapped into a high - dimensional space. At this time, the non - linear structure of the data in the original space will be transformed into a linear structure in the high - dimensional space. Then, by calculating the covariance matrix of the mapped data, eigenvalues and eigenvectors are obtained. The eigenvalues represent the variances of the data in different directions, while the eigenvectors represent the principal components of the data. Finally, by selecting the eigenvectors with larger eigenvalues, most of the variances and key information are retained, thus obtaining the mapped non - linear feature set. The data direction corresponding to the selected eigenvectors is the main feature direction in the high - dimensional space.

[0053] By using the kernel principal component analysis method to map the non - linear feature set into a high - dimensional space, we can effectively process the complex non - linear data of the assembled bridge and culvert node, thereby extracting a more meaningful high - dimensional feature set. This process helps to reveal the potential mechanical characteristics of the node, provide more accurate evaluation results, and thus provide strong data support for fault prediction and health monitoring.

[0054] Step S32: Establish a regression model based on the support vector machine regression algorithm and the non - linear feature set in the high - dimensional space, and calculate the error between the predicted value and the actual observed value of the support vector regression model to obtain the feature prediction value after minimizing the error.

[0055] It can be understood that this step is based on the support vector machine regression algorithm and the non-linear feature set in the high-dimensional space to establish a regression model, calculate the error between the predicted value and the actual observed value of the support vector regression model, and obtain the feature prediction value after minimizing the error. The purpose of this step is to use the support vector machine regression algorithm to perform regression analysis on the non-linear features and obtain more accurate mechanical property prediction results. Using these non-linear features mapped to the high-dimensional space as the input of the support vector machine model, the goal is to establish a regression model to predict the mechanical properties of the assembled bridge and culvert nodes. Among them, the radial basis kernel function is used as the kernel function, so that the feature set in the high-dimensional space can be better mapped into the regression model, thereby capturing the complex non-linear relationship in the node behavior. Then, the model will calculate the optimal regression hyperplane (i.e., the regression function) according to the training data. Among them, the prediction error is the error between the predicted value and the actual observed value, and the regression error is minimized by adjusting the kernel function parameters; in the actual application of the present invention, the strain, vibration and load data of the node in the past period of time are collected through sensors. These data have undergone feature extraction and optimization in the foregoing steps to obtain a non-linear feature set. Next, the support vector machine model is used to perform regression analysis on these features, and the goal is to predict the future mechanical properties of the node. During the training process, by comparing the historical data with the actual observed values, the support vector machine model can learn the response law of the node under different load conditions. Suppose we hope to predict the stress response of the node at a certain future time point. The support vector machine model will use the training data for modeling and output the predicted value, and then obtain the predicted value with the minimized error.

[0056] Step S33: Generate a knowledge graph from the feature prediction value after minimizing the error, and perform an evaluation based on the knowledge graph and a preset feature evaluation threshold to obtain a real-time mechanical property evaluation result.

[0057] It can be understood that in this step, first, based on the predicted values of mechanical characteristics after minimizing errors obtained in the previous step, a knowledge graph is constructed. The key to this process is to structure all non-linear characteristics, prediction results, and actual performance indicators of nodes (such as stress and vibration), and construct a graph according to the relationships between them. Nodes represent different parts or mechanical states of prefabricated bridges and culverts (such as stress, deformation, and vibration frequency), while edges represent causal relationships or similarities between these parts or states (such as the possible functional relationship between stress and deformation, and the correlation between vibration and load). Then, real-time evaluation is carried out through the constructed knowledge graph and preset feature evaluation thresholds. These evaluation thresholds are critical values obtained according to engineering standards, historical experience, or through simulation, and are used to measure whether a node is within the normal working range, or whether potential failures or structural problems may occur. In a specific example of the present invention, the nodes of the knowledge graph may represent different structural parts of the bridge and culvert (such as the base, pipe joints, etc.), and the edges represent the relationships between them (such as stress conduction, vibration mode transmission, etc.). The predicted values of the characteristics of each node will be mapped to the relevant nodes in the graph, such as the predicted value of stress, the predicted value of vibration frequency, etc. Assuming that the safety threshold of stress is 50 MPa, exceeding this threshold is considered to indicate a structural problem, and then the evaluation result of "normal" or "abnormal" is obtained through the method in this step.

[0058] Step S4: Based on the real-time mechanical performance evaluation results and preset historical monitoring data, construct and optimize a model to obtain an optimized prediction model for the changing trend of mechanical performance.

[0059] It can be understood that in this step, by combining real-time evaluation results with historical data, a model is established that can accurately predict the changing trend of the future mechanical performance of the nodes of prefabricated bridges and culverts. This process not only considers the current real-time monitoring data, but also effectively integrates historical data, and uses data-driven methods to scientifically predict the future changes in mechanical performance. In this step, step S4 includes step S41, step S42, and step S43.

[0060] Step S41: Standardize the real-time mechanical performance evaluation results and preset historical monitoring data to obtain fused data with consistent dimensions.

[0061] It is understandable that in this step, for real-time monitoring data and historical monitoring data, a standardization method is adopted to unify the dimensions. For example, the unit of real-time stress data is MPa, and the unit of historical temperature data is °C. By applying the Z-score standardization method, the stress data and temperature data can be standardized into the form of unit standard deviation, so that these two types of data can be compared on the same scale. After the standardization process, both the real-time monitoring data and the historical monitoring data have been converted into data with the same dimension. Next, these data are aligned according to the timestamps and data fusion is performed. The fused dataset includes real-time evaluation results and historical data, providing consistent inputs for subsequent modeling and prediction.

[0062] Step S42: Divide the fused data with consistent dimensions according to the windows of a preset time series, and input the fused data of multiple preset time periods obtained by the division into a preset long short-term memory network for learning to obtain a trained long short-term memory network model;

[0063] It is understandable that in this step, the time series data that has been standardized and fused is divided according to the preset window size. Here, the "window" refers to the monitoring data collected within each period of time. Among them, the size of the window is determined according to the sampling frequency of the data and the actual application requirements. Then, the divided time window data is input into the long short-term memory network and converted into the input format of the long short-term memory network (number of samples, number of time steps, number of features). The divided time series data is used as the input and sent into the preset long short-term memory network for training. The long short-term memory network includes an input gate, a forget gate, and an output gate, which are used to determine which information needs to be retained and which information needs to be discarded, thereby effectively capturing the long-term dependencies in the time series. Among them, the long short-term memory network includes a long short-term memory network unit layer and can be combined with other layers such as a fully connected layer and a Dropout layer to capture more complex time series features. During training, the network adjusts the parameters in the network through the backpropagation algorithm so that the model can minimize the prediction error. Among them, the present invention uses the mean square error as the loss function.

[0064] Step S43: Optimize the hyperparameters based on the simulated annealing algorithm, and substitute the optimized hyperparameters into the trained long short-term memory network model to obtain an optimized long short-term memory network model.

[0065] It can be understood that in this step, by selecting the learning rate, the number of hidden layer units, the batch size, and the number of time steps as hyperparameters, and then randomly selecting a set of hyperparameters as the initial solution. For example, the learning rate is 0.01, the number of hidden layer units is 50, the batch size is 32, and the number of time steps is 30. Search within the neighborhood of the current solution to generate new hyperparameter combinations. For example, neighborhood solutions can be generated by increasing or decreasing the learning rate, adjusting the number of hidden layer units, etc. Then, for each new hyperparameter combination, train the LSTM model and calculate its prediction error on the validation set. If the error of the new solution is less than the preset threshold, accept this solution; if the error is greater than the preset threshold, accept this solution with a preset probability. As the simulated annealing process progresses, gradually reduce the temperature. The decrease in temperature means that the acceptance probability of the inferior solution by the algorithm decreases, and finally converges to the global optimal solution. Until when the temperature reaches the set minimum value, or when no better solution is found for several consecutive times, the algorithm terminates. Through the global optimization characteristics of the simulated annealing algorithm, it is possible to find the optimal or near-optimal hyperparameter combination within a large search space, avoiding the computational burden of traditional grid search or random search methods.

[0066] Step S5: Send the real-time monitoring data to the optimized mechanical property change trend prediction model for prediction to obtain the mechanical properties of the precast bridge and culvert node in a preset time period.

[0067] It can be understood that in this step, by using the optimized mechanical property change trend prediction model and combining with the real-time monitoring data, the future mechanical properties of the node can be accurately predicted, especially the change trend of the node's force can be captured. Among them, after step S5, there are also steps S51 and S52.

[0068] Step S51: Construct an adversarial network based on the optimized mechanical property change trend prediction model and the real-time monitoring data, generate final simulation data with an error less than the threshold from the real-time monitoring data based on the adversarial network, and perform anomaly identification on the final simulation data based on the isolation forest algorithm to obtain the real-time monitoring data corresponding to the abnormal nodes;

[0069] It can be understood that in this step, based on the optimized mechanical property change trend prediction model and real-time monitoring data, simulated data similar to the actual data is generated. Among them, the adversarial network includes a generator and a discriminator. The goal of the generator is to generate simulated data as similar as possible to the real data, ensuring that the error between the simulated data and the actual monitoring data is less than a preset threshold. The generator can reduce the error by continuously optimizing its parameters. Then, a discriminator is constructed based on the data of the generator. The goal of the discriminator is to improve the generation ability of the generator by distinguishing between real data and generated data. Through repeated training, the generator and the discriminator play against each other, and the generator can finally generate data with the smallest error from the real data. In each round of training, the generator will try to generate simulated data close to the actual data, while the discriminator will judge the authenticity of the generated data and feedback the error information to the generator to optimize the generator parameters.

[0070] It can be understood that the Isolation Forest algorithm is an efficient anomaly detection algorithm. It judges whether a data point is abnormal by constructing multiple trees and using the "isolation" degree of the data point from the trees. First, in this step, the finally generated simulated data is input into the Isolation Forest model for processing;

[0071] The Isolation Forest "isolates" the data by randomly selecting features and constructing multiple trees. For each data point, it calculates the degree of its isolation. The data point with a higher isolation degree is more likely to be an outlier. For each data point, the Isolation Forest algorithm generates an anomaly score. If the score exceeds a certain set threshold, the data point is considered abnormal data. In a specific example of the present invention, the vibration data in a certain period showed a sudden change. The generator generated simulated data based on historical data, but the Isolation Forest algorithm recognized that the vibration data in this period was quite different from the patterns of most data and had a high anomaly score. Therefore, the Isolation Forest will mark this data as abnormal, and then provide possible warning information for engineers to prevent potential failures that have not been detected.

[0072] Step S52: Based on the variational autoencoder, perform damage analysis on the real-time monitoring data corresponding to the abnormal node to obtain the damage data in the real-time monitoring data corresponding to the abnormal node.

[0073] It can be understood that in this step, damage usually manifests as abnormal changes in the strain, vibration, or load data of the structure. For example, a sharp increase in the strain value may represent a crack or material fatigue of the node; abnormal fluctuations in the vibration amplitude may be related to the loosening or deformation of the node. Therefore, the damage data includes the sudden change data of the strain data, the abnormal fluctuation data of the vibration data, and the sudden change data of the load change.

[0074] In this step, the variational autoencoder maps the input real-time monitoring data (including data such as strain, vibration, and load) to the latent space. At this time, the information in the monitoring data is compressed into low-dimensional latent variables, and the encoder of the variational autoencoder finds these latent variables so that they can effectively represent the input data. Among them, each point in the latent space corresponds to a specific representation of the input data. Then, the latent variables are obtained from the latent space through the decoder and mapped back to the original data space, thereby reconstructing an output as similar as possible to the original data. Furthermore, the variational autoencoder is trained by minimizing the reconstruction error and the KL divergence of the latent space, and then damage data is obtained. In a specific example of the present invention, it is assumed that the strain data at a certain moment fluctuates less under normal conditions, but after being input into the variational autoencoder, the reconstruction error increases significantly. This may indicate that the strain data in this period has been damaged. For example, microcracks or material fatigue may have occurred near the strain sensor, resulting in a large change in the strain value at this node. The variational autoencoder can accurately identify these anomalies and mark them as damaged data through the increase in the reconstruction error. Among them, after step S52, there are also steps S53, S54, S55, and S56.

[0075] Step S53: Fuzzify the damage data of the bridge and culvert node to obtain damage data of multiple damage levels;

[0076] It can be understood that in this step, first, the membership function of each damage level is defined. For example, the membership function of minor damage may be presented as a triangle, indicating that the membership of minor damage data is relatively high within a specific range, and outside this range, the membership drops rapidly. Then, according to the actual damage values (such as strain, vibration, load, etc.) in the monitoring data, these data are mapped to multiple damage levels through the membership function. Finally, by inputting the damage data (such as strain, vibration, load) into the fuzzification processing model, according to the preset membership function, the membership corresponding to each data point is calculated one by one. For each monitoring data point, calculate its membership under each damage level, and generate a fuzzy damage level based on these membership values.

[0077] In this step, three membership functions are constructed, including a minor damage membership function, a moderate damage membership function, and a severe damage membership function. Among them, the minor damage membership function is as follows:

[0078]

[0079] Among them, μ1(x) represents the value of the minor damage membership function, x is the damage value of the node, and both a and b are the interval boundaries of minor damage.

[0080] Among them, the moderate damage membership function is as follows:

[0081]

[0082] Among them, μ2(x) represents the membership function value of moderate damage, x is the damage value of the node, b is the interval boundary of mild damage, and both c and d are the interval boundaries of moderate damage.

[0083]

[0084] Among them, μ3(x) represents the membership function value of severe damage, x is the damage value of the node, and both c and d are the interval boundaries of moderate damage.

[0085] Step S54: Convert the historical reinforcement plan into reinforcement plans for multiple damage levels to obtain reinforcement plans for multiple damage levels.

[0086] It can be understood that in this step, in practical applications, the relationship between the damage level of the structure and the reinforcement plan is non-linear. Generally, the more severe the damage, the stronger the required reinforcement measures. The division of damage levels is usually based on monitoring data such as the strain, vibration, and load of the structure. After fuzzy processing, different damage levels are obtained. According to the experience of historical reinforcement plans, a mapping relationship can be established between these damage levels and the corresponding reinforcement plans. For example: Mild damage: Only local reinforcement is required, and simple repair materials such as carbon fiber patches and adhesives are used. Moderate damage: Reinforcement of nodes or connection parts is required, and relatively complex reinforcement plans such as using steel plates, stiffeners, or concrete pouring are adopted. Severe damage: Comprehensive large-scale reinforcement is required, and measures such as using prestressed steel cables and concrete jacketing are used.

[0087] Step S55: Select the reinforcement plan corresponding to the damage data of each bridge and culvert node based on the multi-objective optimization function to obtain the preliminary reinforcement strategy for all damaged bridge and culvert nodes.

[0088] It can be understood that in this step, according to the damage data of each node obtained in the previous steps, it is input into the multi-objective optimization algorithm. The damage data of each node may include monitoring data such as strain, vibration, and load, as well as damage level and damage characteristic data generated according to damage analysis. Based on these data, the algorithm will evaluate the damage degree of each node and select the corresponding reinforcement plan according to the preset damage level.

[0089] Among them, the multi-objective optimization function is as follows:

[0090]

[0091] Among them, f1(e), f2(e),..., f m(e) represents different objective functions, which are used to measure the optimization objectives in various aspects, such as cost, effectiveness, and resource utilization. e represents a vector of optimization decision variables, which are different values of the reinforcement scheme or design variables. r1(e), r2(e),..., r n (e) represents the constraint condition functions, which represent certain constraint conditions, such as resource limitations and structural limitations. These functions calculate the decision variable e and ensure that they are within the given corresponding upper limits R1, R2,..., R n range, R1, R2,..., R n represents the upper limit of the constraint condition, which represents the maximum allowable value of a specific constraint, such as the maximum limit of resources and the maximum load-bearing capacity of the structure. T start represents the start time of the task or reinforcement process. T finish (e) represents the completion time of the reinforcement process calculated based on the reinforcement scheme of the decision variable e. T max represents the maximum allowable completion time, which represents the time limit for the completion of the reinforcement task. L(e) is the total damage level of the reinforcement scheme calculated according to the decision variable e, where the mild damage level is 1, the moderate damage level is 2, and the severe damage level is 3. L max represents the maximum allowable damage level, which is used to limit the damage control of the reinforcement scheme. n add (e) represents the quantity of additional resources required in the reinforcement scheme generated by the decision variable e. n max represents the maximum acceptable quantity of additional resources. C(e) represents the cost of the reinforcement scheme. C max represents the maximum allowable cost.

[0092] Step S56: Conduct an analysis on maximizing the utilization of resources for the preliminary reinforcement strategies of all damaged bridge and culvert nodes, and reselect all the said preliminary reinforcement strategies based on the analysis results to obtain the optimal reinforcement strategy.

[0093] It can be understood that in this step, the objective function is defined by the ratio between the cost of the reinforcement scheme and the expected benefits (such as the improvement of node bearing capacity). Resource constraints are the core constraint conditions of this model, mainly including budget limitations, construction resource limitations (such as labor and equipment), and time limitations (such as construction period and project delivery date). The decision variable can be the selection of the reinforcement scheme for each node, represented as a 0-1 decision variable, that is, whether a certain reinforcement method is adopted for a certain node. Then, based on the results of the analysis on maximizing the utilization of resources, under the condition of meeting the resource constraint conditions, the system will re-evaluate and select all the preliminary reinforcement strategies. The reinforcement scheme for each node will be re-ranked according to the consumption of the resources required by it and its contribution to the overall reinforcement effect, and the optimal reinforcement strategy will be selected. The finally generated reinforcement strategy will be the reinforcement scheme with the most optimized resource utilization, ensuring the best reinforcement effect of the project under resource limitations.

[0094] Example 2:

[0095] As Figure 2 shown, this embodiment provides a mechanical property analysis system for prefabricated bridge and culvert joints based on a neural network. Refer to Figure 2 The system includes an acquisition unit 701, a processing unit 702, an evaluation unit 703, a construction unit 704, and a prediction unit 705.

[0096] The acquisition unit 701 is used to acquire real-time monitoring data of prefabricated bridge and culvert joints. The real-time monitoring data includes strain data, vibration data, and load data.

[0097] The processing unit 702 is used to extract and optimize features from the real-time monitoring data of the prefabricated bridge and culvert joints to obtain a non-linear feature set.

[0098] The evaluation unit 703 is used to evaluate the mechanical properties of the non-linear feature set to obtain a real-time mechanical property evaluation result.

[0099] The construction unit 704 is used to construct and optimize a model based on the real-time mechanical property evaluation result and preset historical monitoring data to obtain an optimized prediction model for the trend of mechanical property changes.

[0100] The prediction unit 705 is used to send the real-time monitoring data to the optimized prediction model for the trend of mechanical property changes for prediction to obtain the mechanical properties of the prefabricated bridge and culvert joints in a preset time period.

[0101] It should be noted that regarding the system in the above embodiment, the specific manners in which each module performs operations have been described in detail in the embodiment related to the method, and will not be elaborated here.

[0102] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0103] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present invention, and all should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for analyzing mechanical properties of assembled bridge and culvert nodes based on neural network, characterized in that: include: Acquire real-time monitoring data of prefabricated bridge and culvert nodes, wherein the real-time monitoring data includes strain data, vibration data and load data; Extracting and optimizing the real-time monitoring data of the prefabricated bridge and culvert nodes to obtain a nonlinear feature set; Performing mechanical property evaluation on the nonlinear feature set to obtain real-time mechanical property evaluation results; Based on the real-time mechanical property evaluation results and preset historical monitoring data, a model is constructed and optimized to obtain an optimized mechanical property change trend prediction model; The real-time monitoring data is sent to the optimized mechanical property change trend prediction model for prediction to obtain the mechanical properties of the prefabricated bridge and culvert nodes in a preset time period.

2. The neural network-based mechanical performance analysis method for assembled bridge and culvert nodes according to claim 1 is characterized in that , feature extraction and optimization are performed on the real-time monitoring data of the prefabricated bridge and culvert nodes to obtain a nonlinear feature set, including: The real-time monitoring data of prefabricated bridge and culvert nodes are used to perform factor analysis and processing to obtain a set of mechanical characteristic factors for extracting prefabricated bridge and culvert nodes. The mechanical characteristic factor set is unsupervisedly learned through the autoencoder of the deep belief network to obtain the nonlinear characteristics of the mechanical characteristic factor set; According to the nonlinear feature processing of the mechanical feature factor set by the gradient boosting machine, an optimized mechanical feature set is obtained; Based on the ensemble learning method, all decision trees in the gradient boosting machine are aggregated to obtain the weighted average of multiple decision trees and the optimized nonlinear feature set.

3. The neural network-based mechanical performance analysis method for assembled bridge and culvert nodes according to claim 1 is characterized in that , the nonlinear feature set is subjected to mechanical performance evaluation, including: Based on the kernel principal component analysis method, the nonlinear feature set is mapped to the preset high-dimensional space to obtain the nonlinear feature set in the high-dimensional space; A regression model is established based on the support vector machine regression algorithm and the nonlinear feature set in the high-dimensional space, and the error between the predicted value of the support vector regression model and the actual observed value is calculated to obtain the feature prediction value after the error is minimized; The feature prediction values ​​after minimizing the error are used to generate a knowledge graph, and an evaluation is performed based on the knowledge graph and a preset feature evaluation threshold to obtain a real-time mechanical property evaluation result.

4. The neural network-based mechanical performance analysis method for assembled bridge and culvert nodes according to claim 1 is characterized in that , based on the real-time mechanical performance evaluation results and the preset historical monitoring data, the model is constructed and optimized, including: Standardizing the real-time mechanical property evaluation results and the preset historical monitoring data to obtain fusion data with consistent dimensions; The fused data with consistent dimensions are divided according to the windows of the preset time series, and the fused data of multiple preset time periods obtained by the division are input into the preset long short-term memory network for learning to obtain the trained long short-term memory network model; The hyperparameters are optimized based on a simulated annealing algorithm, and the optimized hyperparameters are substituted into the trained long short-term memory network model to obtain an optimized long short-term memory network model.

5. The neural network-based mechanical performance analysis method for assembled bridge and culvert nodes according to claim 1 is characterized in that ,After obtaining the mechanical properties of the prefabricated bridge and culvert nodes in the preset time period, it also includes: An adversarial network is constructed based on the optimized mechanical property change trend prediction model and real-time monitoring data, and the final simulation data with an error less than a threshold value compared with the real-time monitoring data is generated based on the adversarial network. The final simulation data is anomaly identified based on the isolation forest algorithm to obtain the real-time monitoring data corresponding to the abnormal nodes. Based on the variational autoencoder, the damage data of the real-time monitoring data corresponding to the abnormal node is analyzed to obtain the damage data in the real-time monitoring data corresponding to the abnormal node.

6. A neural network-based mechanical performance analysis system for assembled bridge and culvert nodes, characterized in that: include: An acquisition unit, used to acquire real-time monitoring data of the assembled bridge and culvert nodes, wherein the real-time monitoring data includes strain data, vibration data and load data; A processing unit, used for extracting and optimizing features of the real-time monitoring data of the prefabricated bridge and culvert nodes to obtain a nonlinear feature set; An evaluation unit, used for performing mechanical property evaluation on the nonlinear feature set to obtain a real-time mechanical property evaluation result; A construction unit, used to construct and optimize a model based on the real-time mechanical property evaluation results and preset historical monitoring data to obtain an optimized mechanical property change trend prediction model; The prediction unit is used to send the real-time monitoring data to the optimized mechanical performance change trend prediction model for prediction, so as to obtain the mechanical performance of the prefabricated bridge and culvert nodes in a preset time period.

7. The neural network-based mechanical performance analysis system for assembled bridge and culvert nodes according to claim 6 is characterized in that: The processing unit comprises: The first processing subunit is used to perform factor analysis processing based on the real-time monitoring data of the prefabricated bridge and culvert nodes to obtain a set of mechanical characteristic factors for extracting the prefabricated bridge and culvert nodes; The second processing subunit is used to perform unsupervised learning on the mechanical characteristic factor set through the autoencoder of the deep belief network to obtain the nonlinear characteristics of the mechanical characteristic factor set; The third processing subunit is used for processing the nonlinear characteristics of the mechanical characteristic factor set according to the gradient boosting machine to obtain an optimized mechanical characteristic set; The fourth processing subunit is used to aggregate all decision trees in the gradient boosting machine based on an ensemble learning method to obtain a weighted average of multiple decision trees to obtain an optimized nonlinear feature set.

8. The neural network-based mechanical performance analysis system for assembled bridge and culvert nodes according to claim 6 is characterized in that: The evaluation unit comprises: A first evaluation subunit is used to map the nonlinear feature set to a preset high-dimensional space based on a kernel principal component analysis method to obtain a nonlinear feature set in the high-dimensional space; The second evaluation subunit is used to establish a regression model based on a support vector machine regression algorithm and a nonlinear feature set in a high-dimensional space, and calculate the error between the predicted value of the support vector regression model and the actual observed value to obtain a feature prediction value after the error is minimized; The third evaluation subunit is used to generate a knowledge graph based on the feature prediction value after minimizing the error, and perform evaluation based on the knowledge graph and a preset feature evaluation threshold to obtain a real-time mechanical property evaluation result.

9. The neural network-based mechanical performance analysis system for assembled bridge and culvert nodes according to claim 6 is characterized in that: The construction unit comprises: A first construction subunit is used to standardize the real-time mechanical property evaluation results and preset historical monitoring data to obtain fusion data with consistent dimensions; The second construction subunit is used to divide the fused data with consistent dimensions according to the windows of the preset time series, and input the fused data of multiple preset time periods obtained by division into a preset long short-term memory network for learning, so as to obtain a trained long short-term memory network model; The third construction subunit is used to optimize the hyperparameters based on a simulated annealing algorithm, and substitute the optimized hyperparameters into the trained long short-term memory network model to obtain an optimized long short-term memory network model.

10. The neural network-based mechanical performance analysis system for assembled bridge and culvert nodes according to claim 6 is characterized in that: After the prediction unit, it also includes: The first judgment subunit is used to build an adversarial network based on the optimized mechanical property change trend prediction model and the real-time monitoring data, and generate final simulation data with an error less than a threshold value with the real-time monitoring data based on the adversarial network, and identify anomalies of the final simulation data based on the isolation forest algorithm to obtain real-time monitoring data corresponding to the abnormal nodes; The second judgment subunit is used to perform damage analysis on the real-time monitoring data corresponding to the abnormal node based on the variational autoencoder to obtain damage data in the real-time monitoring data corresponding to the abnormal node.

Images