An intelligent continuous rigid frame bridge construction period deformation prediction system

Abstract

The present application relates to a kind of intelligent continuous rigid frame bridge construction period deformation prediction system, the system includes construction data acquisition module, data preliminary processing module, feature extraction model construction module, mode special model construction module and deformation prediction module.The original data is obtained by data acquisition;Using data standardization, weighted time alignment, effect decoupling and data set segmentation data preliminary processing method;Using multi-branch deep learning model as feature extraction model, the time dynamic of bridge structure response, spatial topological relationship and physical mechanics priori are integrated, the completeness of feature expression and discriminability are significantly improved;Using the progressive strategy of feature extraction and clustering first, then grouping independent modeling, not only fully consider the specificity of different construction stages, but also improve the prediction accuracy through targeted model optimization, solve the mode diversity problem existing in complex construction system.

Description

Technical Field

[0001] This invention relates to the field of intelligent continuous rigid frame bridge deformation prediction technology, specifically to an intelligent continuous rigid frame bridge deformation prediction system during construction. Background Technology

[0002] Deformation prediction during the construction period of continuous rigid frame bridges utilizes advanced technologies such as machine learning, deep learning, and big data analysis. It combines multi-dimensional data on the bridge's construction progress, material properties, and environmental factors to establish an intelligent prediction model. This model can predict the deformation that may occur during the construction process in real time. Through intelligent analysis, it can not only optimize construction plans and reduce construction risks, but also effectively reduce the costs in bridge design and construction and extend the service life of the bridge.

[0003] However, traditional deformation prediction during the construction period of continuous rigid frame bridges suffers from several technical problems. Firstly, it lacks a dedicated feature extraction mechanism, failing to fully exploit the spatiotemporal correlation patterns and physical laws inherent in the construction data. This results in insufficient capture of key features in the evolution of bridge structural behavior, and the prediction model is susceptible to noise interference. Secondly, traditional deformation prediction during the construction period of continuous rigid frame bridges uses a uniform global model to process the deformation prediction of all segments, ignoring the heterogeneity of working conditions and the diversity of patterns during construction. When there are significant differences in construction conditions and structural responses between segments, the uniform model parameters cannot simultaneously meet the prediction requirements of all situations, leading to increased prediction bias. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent deformation prediction system for continuous rigid frame bridges during construction. Traditional deformation prediction systems for continuous rigid frame bridges lack a dedicated feature extraction mechanism, failing to fully exploit the spatiotemporal correlation patterns and physical laws inherent in construction data. This results in insufficient capture of key features in the evolution of bridge structural behavior and susceptibility of the prediction model to noise interference. This solution creatively employs a multi-branch deep learning model as the feature extraction model. Through multi-perspective feature fusion, it integrates the temporal dynamics of the bridge structural response, spatial topological relationships, and prior physical and mechanical knowledge, significantly improving the completeness and discriminative power of feature representation. This lays a reliable data foundation for subsequent accurate deformation prediction. The fundamental problem is that traditional deformation prediction for continuous rigid frame bridges during construction uses a unified global model to process the deformation prediction of all segments, ignoring the heterogeneity of working conditions and the diversity of patterns during construction. When there are significant differences in construction conditions and structural responses between segments, the unified model parameters cannot simultaneously meet the prediction requirements of all situations, resulting in increased prediction deviations. This solution creatively adopts a progressive strategy of first extracting and clustering features, and then grouping and independently modeling. By intelligently grouping segments with similar features and designing a dedicated predictor for each group, it not only fully considers the specificity of different construction stages, but also improves the prediction accuracy through targeted model optimization, thus solving the problem of pattern diversity in complex construction systems.

[0005] The technical solution adopted by the present invention is as follows: The present invention provides an intelligent deformation prediction system for continuous rigid frame bridges during construction, including a construction data acquisition module, a preliminary data processing module, a feature extraction model construction module, a pattern-specific model construction module, and a deformation prediction module;

[0006] The construction data acquisition module obtains the original dataset for deformation prediction through data collection and sends the original dataset for deformation prediction to the data preliminary processing module.

[0007] The preliminary data processing module employs data standardization, weighted time-series alignment, effect decoupling, and dataset segmentation to obtain a dataset to be predicted, a feature extraction training set, and a feature extraction test set. The dataset to be predicted is then sent to the deformation prediction module, while the feature extraction training set and the feature extraction test set are sent to the feature extraction model construction module.

[0008] The feature extraction model construction module is used to construct the model required for extracting global feature representations of segments. By constructing a multi-branch deep learning model as the feature extraction model, feature extraction is then performed to obtain a global feature set of historical segments. The feature extraction model is then sent to the deformation prediction module, and the global feature set of historical segments is sent to the pattern-specific model construction module.

[0009] The pattern-specific model construction module is used to cluster the global features of segments and construct multiple pattern-specific models based on the clustering results for segment deformation prediction. An adaptive clustering model and an improved gradient boosting tree model are constructed as segment clustering models and pattern-specific models, respectively, and the segment clustering models and the pattern-specific models are sent to the deformation prediction module.

[0010] The deformation prediction module is used for segmental deformation prediction during the construction period of a continuous rigid frame bridge. Specifically, it combines the feature extraction model, the segment clustering model, and the pattern-specific model to perform segmental deformation prediction and obtain segmental deformation prediction results.

[0011] Furthermore, in the construction data acquisition module, the deformation prediction raw dataset specifically includes a past deformation prediction raw dataset and a real-time deformation prediction raw dataset. The past deformation prediction raw dataset specifically includes the overall construction data of historically completed bridges and the corresponding data related to each completed bridge segment, the data related to each completed bridge segment of the currently under construction bridge, the data related to the short-term construction segments of the currently under construction bridge segment, and deformation pattern labels. The real-time deformation prediction raw dataset specifically includes the overall construction data of the currently under construction bridge and the real-time segment-related data of the currently under construction bridge segment.

[0012] Furthermore, in the preliminary data processing module, the data standardization is used to ensure the quality and consistency of the construction data, specifically by handling missing values ​​through linear interpolation. The rule-based detection and correction of outliers and the Z-score standardization method eliminate the influence of dimensions, resulting in a clean and scale-uniform data sequence.

[0013] The weighted time alignment is used to solve the problem of time sequence misalignment caused by differences in construction speed among different bridge segments. Specifically, it introduces the importance weight of the process through a weighted dynamic time warping method to obtain a data sequence with time axis alignment.

[0014] The effect decoupling is used to separate the pure load response component from the total data. Specifically, it decomposes the temperature effect and the time-dependent creep effect through a regression model to obtain the load response time series data mainly caused by the construction load.

[0015] The dataset segmentation is used to obtain training data and test data, specifically by segmenting the original dataset of past deformation predictions.

[0016] The real-time deformation prediction raw dataset is pre-processed through data standardization, weighted temporal alignment, and effect decoupling to obtain a dataset to be predicted. The past deformation prediction raw dataset is pre-processed through data standardization, weighted temporal alignment, effect decoupling, and dataset segmentation to obtain a feature extraction training set and a feature extraction test set.

[0017] Furthermore, in the feature extraction model construction module, a model is used to construct the global feature representation of the segment, providing a reliable data foundation for subsequent segment deformation prediction. Specifically, a multi-branch deep learning model is constructed as the feature extraction model. The multi-branch deep learning model outputs global segment features through a multi-branch feature extraction and fusion mechanism.

[0018] The feature extraction model construction module specifically includes data preparation, multi-branch feature extraction, feature fusion, feature extraction and output, and model construction and training.

[0019] The data preparation is used to convert the preprocessed data into a format that the model can process, and includes:

[0020] Time window partitioning is used to divide continuous time series data into segments of fixed length. Specifically, it involves using a sliding window method to extract data segments of fixed time steps, resulting in multiple time window samples.

[0021] Batch normalization is used to stabilize the training process and accelerate convergence. Specifically, it combines samples from multiple time windows into batch tensors, and performs mean and variance normalization on the data of each batch to obtain a stable input tensor.

[0022] The multi-branch feature extraction is used to extract feature representations of bridge deformation from different perspectives, and includes:

[0023] The spatiotemporal convolution branch is used to capture the local spatiotemporal patterns of deformed data. Specifically, it extracts features by sliding along the time dimension through a one-dimensional convolution operation to obtain spatiotemporal convolution features.

[0024] The graph neural network branch is used to model the spatial topological relationship between bridge monitoring points. Specifically, the layout of bridge segment monitoring points is modeled as a graph structure, and the connection relationship between monitoring points is processed by a graph convolutional network to obtain graph structure features.

[0025] The physical prior branch is used to embed prior knowledge of bridge mechanics into the model. Specifically, it obtains physical guidance features by combining physical basis functions and learnable parameters.

[0026] The temporal attention branch is used to focus on deformation changes at key time points during construction. Specifically, it calculates the importance weights of different time points through a self-attention mechanism to obtain temporal weighted features.

[0027] The feature fusion is used to integrate feature information extracted from multiple branches, and includes:

[0028] Feature dimension alignment is used to unify the dimensions of features from different branches for fusion. Specifically, it maps spatiotemporal convolutional features, graph structure features, physical guidance features, and temporal weighted features to a space of the same dimension through linear transformation, resulting in aligned spatiotemporal convolutional features, aligned graph structure features, aligned physical guidance features, and aligned temporal weighted features.

[0029] Cross-attention calculation is used to capture the correlation between different branch features after alignment. Specifically, from the alignment spatiotemporal convolution features, alignment graph structure features, alignment physical guidance features, and alignment temporal weighted features, one feature is selected as the query and the other feature is selected as the key and value. Cross-attention is calculated to perform pairwise fusion between features to obtain multiple cross-fused features.

[0030] Weighted feature fusion is used to determine the fusion ratio of each branch feature and generate the final comprehensive feature representation. Specifically, it involves concatenating all cross-fusion features, performing linear mapping, generating the weights of each cross-fusion feature using the softmax activation function, and then weighted fusion of all cross-fusion features to obtain the fused comprehensive feature.

[0031] The feature extraction and output are used to extract and output segmental global feature representations from the fused comprehensive features, and include the following:

[0032] Global average pooling is used to reduce the feature dimensionality while retaining global information. Specifically, it involves performing global average pooling on the fused composite features along the time dimension to obtain pooled composite features.

[0033] The activation function is designed to introduce nonlinearity and adapt to the bounded output characteristics of deformation prediction. Specifically, the deformation activation function is designed by combining the Swish function and the tanh function.

[0034] The global feature output is used to generate the final global feature representation. Specifically, the pooled comprehensive features are mapped to a low-dimensional space through a fully connected layer with a deformable activation function to obtain the segment global features.

[0035] The model construction and training specifically involves constructing a multi-branch deep learning model by integrating the data preparation, multi-branch feature extraction, feature fusion, and feature extraction and output. The model is then trained and its performance is verified based on the feature extraction training set and the feature extraction test set, resulting in the trained multi-branch deep learning model as the feature extraction model.

[0036] Using the dataset to be predicted and the feature extraction training set as input to the feature extraction model, the global feature set of historical segments is extracted based on the overall construction data of the bridge under construction in the dataset to be predicted, the overall construction data of the historical completed bridges in the feature extraction training set, the relevant data of each completed bridge segment, and the relevant data of each completed bridge segment of the bridge under construction.

[0037] Furthermore, in the pattern-specific model construction module, it is used to cluster the global features of segments and construct multiple pattern-specific models based on the clustering results to perform segment deformation prediction. Specifically, it constructs an adaptive clustering model and a gradient boosting tree model as the segment clustering model and the pattern-specific model, respectively.

[0038] The pattern-specific model construction module specifically includes adaptive clustering analysis, pattern-specific model construction, input segment identification and matching, and model fine-tuning and prediction.

[0039] The adaptive clustering analysis is used to cluster the global feature set of historical segments to discover bridge segments with similar construction conditions. The content includes:

[0040] Feature space similarity calculation is used to quantify the degree of similarity between different segments in the feature space. Specifically, it is to obtain the standard feature distance by calculating the normalized Euclidean distance between global features of historical segments.

[0041] The temporal shape similarity metric is used to evaluate the shape similarity of the deformation development curves of different segments caused by construction loads. Specifically, it is calculated by introducing time weights into the Fréchet distance calculation, emphasizing the shape matching of key construction nodes, and obtaining the temporal shape distance.

[0042] The composite distance metric is used to generate a composite distance that comprehensively considers feature space similarity and temporal shape similarity. Specifically, it adjusts the contribution ratio of feature space similarity and temporal shape similarity by weighting coefficients that change over time to obtain the composite distance.

[0043] Local density calculation is used to identify density peak points in the feature space. Specifically, it obtains the local density by combining a kernel density estimation method based on local scale adjustment of K nearest neighbors.

[0044] Relative distance calculation is used to determine the distance between each segment and higher local density segments. Specifically, it involves adjusting the original distance calculation by introducing time consistency constraints to obtain the relative distance under the constraints.

[0045] The decision value calculation and center selection are used to automatically identify suitable cluster centers. Specifically, the decision value is calculated by combining local density and relative distance under constraints, and the segment with the decision value greater than the preset decision value threshold is selected as the cluster center.

[0046] Clustering assignment strategy design is used to assign segments to appropriate clusters. Specifically, by combining a composite distance metric and a time interval penalty assignment method, segments are assigned to the corresponding clusters to obtain segment clustering results.

[0047] The pattern-specific model construction is used to train a lightweight prediction model for each cluster. Specifically, based on the global features of historical segments in each cluster, a corresponding gradient boosting tree model is constructed and trained as a pattern-specific model. The pattern-specific models corresponding to each cluster are independent of each other and are used to output the deformation prediction results of continuous rigid frame bridge segments during the construction period. The deformation prediction results are specifically the predicted deformation pattern labels of bridge segments.

[0048] The input segment identification and matching is used to match the current bridge segment under construction to an existing cluster category, and includes:

[0049] Global feature extraction is used to obtain the feature representation of the current bridge segment under construction for pattern recognition. Specifically, the feature extraction model is used to process the data of the current bridge segment under construction to obtain the global features of the current segment.

[0050] Fuzzy membership calculation is used for soft classification of the current bridge segment under construction to each cluster category. Specifically, the softmax activation function is used to calculate the fuzzy membership of the current segment's global features to each cluster category based on the composite distance between the current segment's global features and the mean of each cluster, and the current bridge segment under construction is matched to the cluster category with the largest fuzzy membership.

[0051] The model fine-tuning and prediction system is used to quickly adjust and predict the model based on the current bridge segment data under construction. It includes the following:

[0052] Model parameter initialization is used to set the starting point of the fine-tuning process. Specifically, it loads the corresponding pattern-specific model based on the matched clustering category, and uses its model parameters as initial parameters to obtain the initial parameter set.

[0053] The fine-tuning intensity is determined to control the degree of fine-tuning. Specifically, the fine-tuning intensity is adjusted based on the fuzzy membership degree and the data uncertainty value to obtain the actual fine-tuning intensity, which is used as the learning rate to update the initial parameter set in gradient descent to obtain the updated parameter set.

[0054] The output prediction results are used for real-time deformation prediction. Specifically, the updated parameter set is used to build a fine-tuned pattern-specific model, and the deformation prediction results are output based on the global features of the current segment.

[0055] Furthermore, in the deformation prediction module, for segmental deformation prediction during the construction period of a continuous rigid frame bridge, the dataset to be predicted is processed as input to the feature extraction model to obtain the global features of the current segment, which are then input into the segment clustering model. The corresponding pattern-specific model for the clustering category is loaded and fine-tuned. The fine-tuned pattern-specific model uses the global features of the current segment as input to obtain the deformation prediction result of the current bridge segment under construction.

[0056] The beneficial effects achieved by the present invention using the above solution are as follows:

[0057] (1) In view of the technical problems of traditional continuous rigid frame bridge deformation prediction during construction period lacking a dedicated feature extraction mechanism, failing to fully explore the spatiotemporal correlation patterns and physical laws contained in the construction data, resulting in insufficient capture of key features of bridge structural behavior evolution and easy noise interference of the prediction model, this solution creatively adopts a multi-branch deep learning model as the feature extraction model. Through multi-perspective feature fusion, it integrates the temporal dynamics, spatial topological relationships and physical and mechanical priors of bridge structural response, significantly improving the completeness and discriminative power of feature expression, and laying a reliable data foundation for subsequent accurate deformation prediction.

[0058] (2) In view of the technical problem that traditional deformation prediction of continuous rigid frame bridges during construction period uses a unified global model to process the deformation prediction of all segments, ignoring the heterogeneity of working conditions and the diversity of modes during construction, and that when there are significant differences in construction conditions and structural response between segments, the unified model parameters are difficult to meet the prediction requirements of all situations at the same time, resulting in increased prediction deviation, this solution creatively adopts a progressive strategy of first extracting features and clustering, and then grouping and modeling independently. By intelligently grouping segments with similar features and designing a dedicated predictor for each group, it not only fully considers the specificity of different construction stages, but also improves the prediction accuracy through targeted model optimization, thus solving the problem of mode diversity in complex construction systems. Attached Figure Description

[0059] Figure 1 A schematic diagram of a module for an intelligent continuous rigid frame bridge deformation prediction system during construction, provided by the present invention.

[0060] Figure 2 This is a flowchart illustrating the preliminary data processing module.

[0061] Figure 3 A flowchart illustrating the process of building a module for the feature extraction model;

[0062] Figure 4 A flowchart illustrating the process of building a model-specific module for patterns.

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0064] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0065] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0066] Example 1, see Figure 1 The present invention provides an intelligent deformation prediction system for continuous rigid frame bridges during construction, comprising a construction data acquisition module, a preliminary data processing module, a feature extraction model construction module, a pattern-specific model construction module, and a deformation prediction module.

[0067] The construction data acquisition module obtains the original dataset for deformation prediction through data collection and sends the original dataset for deformation prediction to the data preliminary processing module.

[0068] The preliminary data processing module employs data standardization, weighted time-series alignment, effect decoupling, and dataset segmentation to obtain a dataset to be predicted, a feature extraction training set, and a feature extraction test set. The dataset to be predicted is then sent to the deformation prediction module, while the feature extraction training set and the feature extraction test set are sent to the feature extraction model construction module.

[0069] The feature extraction model construction module is used to construct the model required for extracting global feature representations of segments. By constructing a multi-branch deep learning model as the feature extraction model, feature extraction is then performed to obtain a global feature set of historical segments. The feature extraction model is then sent to the deformation prediction module, and the global feature set of historical segments is sent to the pattern-specific model construction module.

[0070] The pattern-specific model construction module is used to cluster the global features of segments and construct multiple pattern-specific models based on the clustering results for segment deformation prediction. An adaptive clustering model and an improved gradient boosting tree model are constructed as segment clustering models and pattern-specific models, respectively, and the segment clustering models and the pattern-specific models are sent to the deformation prediction module.

[0071] The deformation prediction module is used for segmental deformation prediction during the construction period of a continuous rigid frame bridge. Specifically, it combines the feature extraction model, the segment clustering model, and the pattern-specific model to perform segmental deformation prediction and obtain segmental deformation prediction results.

[0072] Example 2, see Figure 1 This embodiment is based on the above embodiment. In the construction data acquisition module, the deformation prediction raw dataset specifically includes the past deformation prediction raw dataset and the real-time deformation prediction raw dataset. The past deformation prediction raw dataset specifically includes the overall construction data of historically completed bridges and the corresponding data related to each completed bridge segment, the data related to each completed bridge segment of the currently under construction bridge, the data related to the short-term construction segments of the currently under construction bridge segment, and deformation pattern labels. The real-time deformation prediction raw dataset specifically includes the overall construction data of the currently under construction bridge and the real-time segment-related data of the currently under construction bridge segment.

[0073] The overall construction data specifically includes bridge design parameter data, construction progress data, and material information data. The bridge design parameter data specifically includes total bridge length data, span arrangement data, cross-sectional shape data, and prestressing arrangement data. The construction progress data specifically includes bridge start time data, planned construction time data for each segment, and actual construction time data for each segment. The material information data specifically includes concrete strength grade data, concrete elastic modulus data, and prestressed steel strand specification data.

[0074] The segment-related data specifically includes segment identification data, segment geometric data, segment load data, segment prestressing data, segment material shrinkage and creep parameter data, segment environmental data, and segment deformation data. The segment geometric data specifically includes segment length data, segment cross-sectional dimensions data, segment elevation data, and segment pre-camber data. The segment load data specifically includes segment concrete pouring volume data, segment self-weight data, and segment construction load data. The segment prestressing data specifically includes prestressing tensioning time data, steel strand tensioning sequence data, and steel strand tensioning force data. The segment environmental data specifically includes temperature and humidity data and wind speed and direction data during segment construction. The segment deformation data specifically includes deflection data at segment deformation monitoring points, strain value data at segment deformation monitoring points, and segment rotation angle monitoring data.

[0075] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In the preliminary data processing module, the data standardization is used to ensure the quality and consistency of construction data. Specifically, it involves processing missing values ​​through linear interpolation. The rule-based detection and correction of outliers and the Z-score standardization method eliminate the influence of dimensions, resulting in a clean and scale-uniform data sequence.

[0076] The weighted time alignment is used to solve the problem of time sequence misalignment caused by differences in construction speed among different bridge segments. Specifically, it introduces the importance weight of the process through a weighted dynamic time warping method to obtain a data sequence with time axis alignment.

[0077] The effect decoupling is used to separate the pure load response component from the total data. Specifically, it decomposes the temperature effect and the time-dependent creep effect through a regression model to obtain the load response time series data mainly caused by the construction load.

[0078] The dataset segmentation is used to obtain training data and test data, specifically by segmenting the original dataset of past deformation predictions.

[0079] The real-time deformation prediction raw dataset is pre-processed through data standardization, weighted temporal alignment, and effect decoupling to obtain a dataset to be predicted. The past deformation prediction raw dataset is pre-processed through data standardization, weighted temporal alignment, effect decoupling, and dataset segmentation to obtain a feature extraction training set and a feature extraction test set.

[0080] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In the feature extraction model construction module, a model is used to construct the global feature representation of the segment to provide a reliable data foundation for subsequent segment deformation prediction. Specifically, a multi-branch deep learning model is constructed as the feature extraction model. The multi-branch deep learning model outputs the global features of the segment through a multi-branch feature extraction and fusion mechanism.

[0081] The feature extraction model construction module specifically includes data preparation, multi-branch feature extraction, feature fusion, feature extraction and output, and model construction and training.

[0082] The data preparation is used to convert the preprocessed data into a format that the model can process, and includes:

[0083] Time window partitioning is used to divide continuous time series data into segments of fixed length. Specifically, it involves using a sliding window method to extract data segments of fixed time steps, resulting in multiple time window samples.

[0084] Batch normalization is used to stabilize the training process and accelerate convergence. Specifically, it combines samples from multiple time windows into batch tensors, and performs mean and variance normalization on the data of each batch to obtain a stable input tensor.

[0085] The multi-branch feature extraction is used to extract feature representations of bridge deformation from different perspectives, and includes:

[0086] The spatiotemporal convolution branch is used to capture local spatiotemporal patterns in deformed data. Specifically, it extracts features by sliding along the time dimension through a one-dimensional convolution operation to obtain spatiotemporal convolution features. The formula used is as follows:

[0087] ;

[0088] In the formula, Represents spatiotemporal convolutional features. Represents the ReLU activation function. This represents the batch normalization function. Represents a one-dimensional convolution function. Indicates the input tensor;

[0089] The graph neural network branch is used to model the spatial topological relationships between bridge monitoring points. Specifically, it models the layout of bridge segment monitoring points as a graph structure, processes the connections between monitoring points through a graph convolutional network, and obtains the graph structure features. The formula used is as follows:

[0090] ;

[0091] In the formula, Representing the structural features of the graph. This represents the function that runs a graph convolutional network. Represents the adjacency matrix of a graph structure;

[0092] The physical prior branch is used to embed prior knowledge of bridge mechanics into the model. Specifically, it obtains physical guidance features by combining physical basis functions and learnable parameters, as shown in the following formula:

[0093] ;

[0094] In the formula, Indicates physical guidance characteristics. This represents the weight matrix of learnable physical parameters. This indicates the physical guidance bias term. Represents physical basis functions;

[0095] The temporal attention branch focuses on deformation changes at key time points during construction. Specifically, it calculates the importance weights of different time points using a self-attention mechanism to obtain temporal weighted features, as shown in the following formula:

[0096] ;

[0097] In the formula, Represents time-weighted features, This represents the softmax activation function. Represents the self-attention query mapping matrix. Represents the self-attention key mapping matrix. Represents the self-attention value mapping matrix. The dimension of the self-attention key is represented by T, which represents the transpose operation.

[0098] The feature fusion is used to integrate feature information extracted from multiple branches, and includes:

[0099] Feature dimension alignment is used to unify the dimensions of features from different branches for fusion. Specifically, it maps spatiotemporal convolutional features, graph structure features, physical guidance features, and temporal weighted features to a space of the same dimension through linear transformation, resulting in aligned spatiotemporal convolutional features, aligned graph structure features, aligned physical guidance features, and aligned temporal weighted features.

[0100] Cross-attention calculation is used to capture the correlation between different branch features after alignment. Specifically, from the alignment spatiotemporal convolution features, alignment graph structure features, alignment physical guidance features, and alignment temporal weighted features, one feature is selected as the query and the other feature is selected as the key and value. Cross-attention is calculated to perform pairwise fusion between features to obtain multiple cross-fused features.

[0101] Weighted feature fusion is used to determine the fusion ratio of each branch feature and generate the final comprehensive feature representation. Specifically, it involves concatenating all cross-fused features, performing a linear mapping, generating weights for each cross-fused feature using the softmax activation function, and then weighting and fusing all cross-fused features to obtain the fused comprehensive feature. The formula used is as follows:

[0102] ;

[0103] In the formula, This represents the cross-fusion feature weight of the a-th type of cross-fusion. This represents the fusion weight mapping matrix. This represents the bias term in the fusion weight mapping. This indicates the splicing and cross-fusion characteristics. Indicates integrated and comprehensive characteristics, This represents the cross-fusion feature of the a-th type of cross-fusion, where a represents the cross-fusion index. ST_GS represents the cross-fusion of spatiotemporal convolutional features and graph structure features, ST_PH represents the cross-fusion of spatiotemporal convolutional features and physical guidance features, ST_Att represents the cross-fusion of spatiotemporal convolutional features and temporal weighted features, GS_PH represents the cross-fusion of graph structure features and physical guidance features, GS_Att represents the cross-fusion of graph structure features and temporal weighted features, and PH_Att represents the cross-fusion of physical guidance features and temporal weighted features.

[0104] The feature extraction and output are used to extract and output segmental global feature representations from the fused comprehensive features, and include the following:

[0105] Global average pooling is used to reduce feature dimensionality while preserving global information. Specifically, it involves performing global average pooling on the fused composite features along the time dimension to obtain the pooled composite features. The formula used is as follows:

[0106] ;

[0107] In the formula, This indicates the pooling synthesis characteristics. This represents the global average pooling function;

[0108] The activation function is designed to introduce nonlinearity and adapt to the bounded output characteristics of deformation prediction. Specifically, the deformation activation function is designed by combining the Swish function and the tanh function, and the formula used is as follows:

[0109] ;

[0110] In the formula, Let x represent the deformation activation function, and let x represent the input independent variable. This represents the sigmoid activation function. Represents the hyperbolic tangent function. , and Represents distinct learnable parameters;

[0111] The global feature output is used to generate the final global feature representation. Specifically, the pooled comprehensive features are mapped to a low-dimensional space through a fully connected layer with a deformable activation function to obtain the segment global features.

[0112] The model construction and training specifically involves constructing a multi-branch deep learning model by integrating the data preparation, multi-branch feature extraction, feature fusion, and feature extraction and output. The model is then trained and its performance is verified based on the feature extraction training set and the feature extraction test set, resulting in the trained multi-branch deep learning model as the feature extraction model.

[0113] Using the dataset to be predicted and the feature extraction training set as input to the feature extraction model, the global feature set of historical segments is extracted based on the overall construction data of the bridge under construction in the dataset to be predicted, the overall construction data of the historical completed bridges in the feature extraction training set, the relevant data of each completed bridge segment, and the relevant data of each completed bridge segment of the bridge under construction.

[0114] By performing the above operations, this solution addresses the technical problems of traditional continuous rigid frame bridge deformation prediction during construction, which lacks a dedicated feature extraction mechanism, fails to fully explore the spatiotemporal correlation patterns and physical laws contained in the construction data, resulting in insufficient capture of key features of bridge structural behavior evolution and susceptibility of prediction models to noise interference. This solution creatively adopts a multi-branch deep learning model as the feature extraction model. Through multi-perspective feature fusion, it integrates the temporal dynamics, spatial topological relationships, and physical and mechanical priors of the bridge structural response, significantly improving the completeness and discriminative power of feature expression, and laying a reliable data foundation for subsequent accurate deformation prediction.

[0115] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In the pattern-specific model construction module, it is used to cluster the global features of the segments and construct multiple pattern-specific models based on the clustering results to perform segment deformation prediction. Specifically, it constructs an adaptive clustering model and a gradient boosting tree model as the segment clustering model and the pattern-specific model, respectively.

[0116] The pattern-specific model construction module specifically includes adaptive clustering analysis, pattern-specific model construction, input segment identification and matching, and model fine-tuning and prediction.

[0117] The adaptive clustering analysis is used to cluster the global feature set of historical segments to discover bridge segments with similar construction conditions. The content includes:

[0118] Feature space similarity calculation is used to quantify the degree of similarity between different segments in the feature space. Specifically, it is to obtain the standard feature distance by calculating the normalized Euclidean distance between global features of historical segments.

[0119] The temporal shape similarity metric is used to evaluate the shape similarity of deformation development curves caused by construction loads in different segments. Specifically, it is calculated by introducing time-weighted Fréchet distance, emphasizing the shape matching of key construction nodes, to obtain the temporal shape distance. The formula used is as follows:

[0120] ;

[0121] In the formula, This represents the temporal shape distance between segment m and segment n. This represents the load response time series data for segment m. This represents the load response time series data for segment n. and Let t represent distinct mapping functions, and t represent time steps. Euclidean norm operations Indicates key construction milestones. Indicates the time adjustment parameter;

[0122] The composite distance metric is used to generate a composite distance that comprehensively considers feature space similarity and temporal shape similarity. Specifically, it adjusts the contribution ratio of feature space similarity and temporal shape similarity by using weighting coefficients that change over time to obtain the composite distance. The formula used is as follows:

[0123] ;

[0124] In the formula, Represents the similarity weight coefficient. Indicates the time-mapped weights. This represents the composite distance between segment m and segment n. This represents the standard characteristic distance between segment m and segment n;

[0125] Local density calculation is used to identify density peaks in the feature space. Specifically, it obtains the local density by combining a kernel density estimation method based on K-nearest neighbors and local scale adjustment. The formula used is as follows:

[0126] ;

[0127] In the formula, This represents the local density of segment m. Indicates the cutoff distance. The segment m is represented by a local scale based on the K nearest neighbors;

[0128] Relative distance calculation is used to determine the distance between each segment and segments with higher local density. Specifically, it involves adjusting the original distance calculation by introducing time consistency constraints to obtain the relative distance under the constraints. The formula used is as follows:

[0129] ;

[0130] In the formula, This represents the relative distance between segment m and segments with higher local density under minimum constraints. This represents the local density of segment n. and These represent distinct time consistency parameters. This represents the actual construction time of segment m. This represents the actual construction time of segment n;

[0131] The decision value calculation and center selection are used to automatically identify suitable cluster centers. Specifically, the decision value is calculated by combining local density and relative distance under constraints, and segments with decision values ​​greater than a preset decision value threshold are selected as cluster centers. The formula used for decision value calculation is as follows:

[0132] ;

[0133] In the formula, This represents the decision value for segment m;

[0134] A clustering assignment strategy is designed to assign segments to appropriate clusters. Specifically, it uses a combination of composite distance metrics and time interval penalties to assign segments to corresponding clusters, resulting in the segment clustering results. The formula used is as follows:

[0135] ;

[0136] In the formula, Let C represent the cluster label of segment m, and let C represent the set of cluster centers. This represents the composite distance between segment m and cluster center c. This represents the time consistency weighting coefficient. This represents the average actual construction time of the segments in cluster c. Let represent the variance of the actual construction time of the segments in cluster c. This indicates the time interval adjustment parameter;

[0137] The pattern-specific model construction is used to train a lightweight prediction model for each cluster. Specifically, based on the global features of historical segments in each cluster, a corresponding gradient boosting tree model is constructed and trained as a pattern-specific model. The pattern-specific models corresponding to each cluster are independent of each other and are used to output the deformation prediction results of continuous rigid frame bridge segments during the construction period. The deformation prediction results are specifically the predicted deformation pattern labels of bridge segments.

[0138] The input segment identification and matching is used to match the current bridge segment under construction to an existing cluster category, and includes:

[0139] Global feature extraction is used to obtain the feature representation of the current bridge segment under construction for pattern recognition. Specifically, the feature extraction model is used to process the data of the current bridge segment under construction to obtain the global features of the current segment.

[0140] Fuzzy membership calculation is used for soft classification of the current bridge segment under construction to each cluster category. Specifically, the softmax activation function is used to calculate the fuzzy membership of the current segment's global features to each cluster category based on the composite distance between the current segment's global features and the mean of each cluster, and the current bridge segment under construction is matched to the cluster category with the largest fuzzy membership.

[0141] The model fine-tuning and prediction system is used to quickly adjust and predict the model based on the current bridge segment data under construction. It includes the following:

[0142] Model parameter initialization is used to set the starting point of the fine-tuning process. Specifically, it loads the corresponding pattern-specific model based on the matched clustering category, and uses its model parameters as initial parameters to obtain the initial parameter set.

[0143] The fine-tuning intensity is determined to control the degree of fine-tuning. Specifically, the fine-tuning intensity is adjusted based on the fuzzy membership degree and the data uncertainty value to obtain the actual fine-tuning intensity. This intensity is then used as the learning rate to update the initial parameter set in gradient descent, resulting in the updated parameter set. The formula used to determine the fine-tuning intensity is as follows:

[0144] ;

[0145] In the formula, Indicates the actual fine-tuning intensity. Indicates the basic fine-tuning strength. and These represent distinct fine-tuning strength adjustment parameters, where Fd represents the fuzzy membership degree of the clustering category matched by the current bridge segment under construction. This represents the composite distance between the global features of the current segment and the c-mean of the cluster.

[0146] The output prediction results are used for real-time deformation prediction. Specifically, the updated parameter set is used to build a fine-tuned pattern-specific model, and the deformation prediction results are output based on the global features of the current segment.

[0147] By performing the above operations, this solution addresses the technical problem of traditional continuous rigid frame bridge deformation prediction during construction. It uses a unified global model to process the deformation prediction of all segments, neglecting the heterogeneity of working conditions and the diversity of patterns during construction. When there are significant differences in construction conditions and structural responses between segments, the unified model parameters cannot simultaneously meet the prediction needs of all situations, leading to increased prediction bias. This solution creatively adopts a progressive strategy of first extracting and clustering features, and then grouping and independently modeling. By intelligently grouping segments with similar features and designing a dedicated predictor for each group, it not only fully considers the specificity of different construction stages but also improves prediction accuracy through targeted model optimization, thus solving the problem of pattern diversity in complex construction systems.

[0148] Example 6, see Figure 1 This embodiment is based on the above embodiment. In the deformation prediction module, it is used for segmental deformation prediction during the construction period of a continuous rigid frame bridge. Specifically, the dataset to be predicted is processed as the input of the feature extraction model to obtain the global features of the current segment. These features are then input into the segment clustering model. The pattern-specific model corresponding to the clustering category is loaded and fine-tuned. The fine-tuned pattern-specific model takes the global features of the current segment as input to obtain the deformation prediction result of the current bridge segment under construction.

[0149] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0150] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0151] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent deformation prediction system for continuous rigid frame bridges during construction, characterized in that: The system includes a construction data acquisition module, a preliminary data processing module, a feature extraction model construction module, a pattern-specific model construction module, and a deformation prediction module; The construction data acquisition module obtains a deformation prediction raw dataset by collecting data. The deformation prediction raw dataset specifically includes a past deformation prediction raw dataset and a real-time deformation prediction raw dataset. The past deformation prediction raw dataset specifically includes the overall construction data of historically completed bridges and the corresponding data of each completed bridge segment, the data of each completed bridge segment of the currently under construction bridge, the data of the short-term construction segments of the currently under construction bridge segment, and deformation pattern labels. The real-time deformation prediction raw dataset specifically includes the overall construction data of the currently under construction bridge and the real-time segment-related data of the currently under construction bridge segment. The overall construction data specifically includes bridge design parameter data, construction progress data, and material information data. The segment-related data specifically includes segment identification data, segment geometric data, segment load data, segment prestress data, segment material shrinkage and creep parameter data, segment environmental data, and segment deformation data. The preliminary data processing module employs data standardization, weighted time-series alignment, effect decoupling, and dataset segmentation to obtain the dataset to be predicted, the feature extraction training set, and the feature extraction test set. The feature extraction model construction module is used to construct the model required for extracting global feature representations of segments, providing a reliable data foundation for subsequent segment deformation prediction. Specifically, it constructs a multi-branch deep learning model as the feature extraction model. The multi-branch deep learning model outputs global segment features through a multi-branch feature extraction and fusion mechanism. The pattern-specific model construction module is used to cluster the global features of segments and construct multiple pattern-specific models based on the clustering results to predict segment deformation. Specifically, it constructs an adaptive clustering model and a gradient boosting tree model as the segment clustering model and the pattern-specific model, respectively. The deformation prediction module is used for segmental deformation prediction during the construction period of a continuous rigid frame bridge. Specifically, the dataset to be predicted is processed as the input of the feature extraction model to obtain the global features of the current segment. These features are then input into the segment clustering model. The corresponding pattern-specific model for the clustering category is loaded and fine-tuned. The fine-tuned pattern-specific model uses the global features of the current segment as input to obtain the deformation prediction result of the current bridge segment under construction.

2. The intelligent deformation prediction system for continuous rigid frame bridges during construction as described in claim 1, characterized in that: The feature extraction model construction module specifically includes data preparation, multi-branch feature extraction, feature fusion, feature extraction and output, and model construction and training.

3. The intelligent deformation prediction system for continuous rigid frame bridges during construction, as described in claim 2, is characterized in that: The data preparation is used to convert the preprocessed data into a format that the model can process, and includes: Time window partitioning is used to divide continuous time series data into segments of fixed length. Specifically, it involves using a sliding window method to extract data segments of fixed time steps, resulting in multiple time window samples. Batch normalization is used to stabilize the training process and accelerate convergence. Specifically, it combines samples from multiple time windows into batch tensors, and performs mean and variance normalization on the data of each batch to obtain a stable input tensor. The multi-branch feature extraction is used to extract feature representations of bridge deformation from different perspectives, and includes: The spatiotemporal convolution branch is used to capture the local spatiotemporal patterns of deformed data. Specifically, it extracts features by sliding along the time dimension through a one-dimensional convolution operation to obtain spatiotemporal convolution features. The graph neural network branch is used to model the spatial topological relationship between bridge monitoring points. Specifically, the layout of bridge segment monitoring points is modeled as a graph structure, and the connection relationship between monitoring points is processed by a graph convolutional network to obtain graph structure features. The physical prior branch is used to embed prior knowledge of bridge mechanics into the model. Specifically, it obtains physical guidance features by combining physical basis functions and learnable parameters. The temporal attention branch is used to focus on deformation changes at key time points during construction. Specifically, it calculates the importance weights of different time points through a self-attention mechanism to obtain temporal weighted features. The feature fusion is used to integrate feature information extracted from multiple branches, and includes: Feature dimension alignment is used to unify the dimensions of features from different branches for fusion. Specifically, it maps spatiotemporal convolutional features, graph structure features, physical guidance features, and temporal weighted features to a space of the same dimension through linear transformation, resulting in aligned spatiotemporal convolutional features, aligned graph structure features, aligned physical guidance features, and aligned temporal weighted features. Cross-attention calculation is used to capture the correlation between different branch features after alignment. Specifically, from the alignment spatiotemporal convolution features, alignment graph structure features, alignment physical guidance features, and alignment temporal weighted features, one feature is selected as the query and the other feature is selected as the key and value. Cross-attention is calculated to perform pairwise fusion between features to obtain multiple cross-fused features. Weighted feature fusion is used to determine the fusion ratio of each branch feature and generate the final comprehensive feature representation. Specifically, it involves concatenating all cross-fusion features, performing linear mapping, generating the weights of each cross-fusion feature using the softmax activation function, and then weighted fusion of all cross-fusion features to obtain the fused comprehensive feature. The feature extraction and output are used to extract and output segmental global feature representations from the fused comprehensive features, and include the following: Global average pooling is used to reduce the feature dimensionality while retaining global information. Specifically, it involves performing global average pooling on the fused composite features along the time dimension to obtain pooled composite features. The activation function is designed to introduce nonlinearity and adapt to the bounded output characteristics of deformation prediction. Specifically, the deformation activation function is designed by combining the Swish function and the tanh function. The global feature output is used to generate the final global feature representation. Specifically, the pooled comprehensive features are mapped to a low-dimensional space through a fully connected layer with a deformable activation function to obtain the segment global features. The model construction and training specifically involves constructing a multi-branch deep learning model by integrating the data preparation, multi-branch feature extraction, feature fusion, and feature extraction and output. The model is then trained and its performance is verified based on the feature extraction training set and the feature extraction test set, resulting in the trained multi-branch deep learning model as the feature extraction model. Using the dataset to be predicted and the feature extraction training set as input to the feature extraction model, the global feature set of historical segments is extracted based on the overall construction data of the bridge under construction in the dataset to be predicted, the overall construction data of the historical completed bridges in the feature extraction training set, the relevant data of each completed bridge segment, and the relevant data of each completed bridge segment of the bridge under construction.

4. The intelligent deformation prediction system for continuous rigid frame bridges during construction as described in claim 1, characterized in that: The pattern-specific model building module specifically includes adaptive clustering analysis, pattern-specific model building, input segment identification and matching, and model fine-tuning and prediction.

5. The intelligent deformation prediction system for continuous rigid frame bridges during construction as described in claim 4, characterized in that: The adaptive clustering analysis is used to cluster the global feature set of historical segments to discover bridge segments with similar construction conditions. The content includes: Feature space similarity calculation is used to quantify the degree of similarity between different segments in the feature space. Specifically, it is to obtain the standard feature distance by calculating the normalized Euclidean distance between global features of historical segments. The temporal shape similarity metric is used to evaluate the shape similarity of the deformation development curves of different segments caused by construction loads. Specifically, it is calculated by introducing time weights into the Fréchet distance calculation, emphasizing the shape matching of key construction nodes, and obtaining the temporal shape distance. The composite distance metric is used to generate a composite distance that comprehensively considers feature space similarity and temporal shape similarity. Specifically, it adjusts the contribution ratio of feature space similarity and temporal shape similarity by weighting coefficients that change over time to obtain the composite distance. Local density calculation is used to identify density peak points in the feature space. Specifically, it obtains the local density by combining a kernel density estimation method based on local scale adjustment of K nearest neighbors. Relative distance calculation is used to determine the distance between each segment and higher local density segments. Specifically, it involves adjusting the original distance calculation by introducing time consistency constraints to obtain the relative distance under the constraints. The decision value calculation and center selection are used to automatically identify suitable cluster centers. Specifically, the decision value is calculated by combining local density and relative distance under constraints, and the segment with the decision value greater than the preset decision value threshold is selected as the cluster center. Clustering assignment strategy design is used to assign segments to appropriate clusters. Specifically, by combining a composite distance metric and a time interval penalty assignment method, segments are assigned to the corresponding clusters to obtain segment clustering results. The pattern-specific model construction is used to train a lightweight prediction model for each cluster. Specifically, based on the global features of historical segments in each cluster, a corresponding gradient boosting tree model is constructed and trained as a pattern-specific model. The pattern-specific models corresponding to each cluster are independent of each other and are used to output the deformation prediction results of continuous rigid frame bridge segments during the construction period. The deformation prediction results are specifically the predicted deformation pattern labels of bridge segments. The input segment identification and matching is used to match the current bridge segment under construction to an existing cluster category, and includes: Global feature extraction is used to obtain the feature representation of the current bridge segment under construction for pattern recognition. Specifically, the feature extraction model is used to process the data of the current bridge segment under construction to obtain the global features of the current segment. Fuzzy membership calculation is used for soft classification of the current bridge segment under construction to each cluster category. Specifically, the softmax activation function is used to calculate the fuzzy membership of the current segment's global features to each cluster category based on the composite distance between the current segment's global features and the mean of each cluster, and the current bridge segment under construction is matched to the cluster category with the largest fuzzy membership. The model fine-tuning and prediction system is used to quickly adjust and predict the model based on the current bridge segment data under construction. It includes the following: Model parameter initialization is used to set the starting point of the fine-tuning process. Specifically, it loads the corresponding pattern-specific model based on the matched clustering category, and uses its model parameters as initial parameters to obtain the initial parameter set. The fine-tuning intensity is determined to control the degree of fine-tuning. Specifically, the fine-tuning intensity is adjusted based on the fuzzy membership degree and the data uncertainty value to obtain the actual fine-tuning intensity, which is used as the learning rate to update the initial parameter set in gradient descent to obtain the updated parameter set. The output prediction results are used for real-time deformation prediction. Specifically, the updated parameter set is used to build a fine-tuned pattern-specific model, and the deformation prediction results are output based on the global features of the current segment.

6. The intelligent deformation prediction system for continuous rigid frame bridges during construction according to claim 1, characterized in that: In the preliminary data processing module, data standardization is used to ensure the quality and consistency of construction data. Specifically, it involves processing missing values ​​through linear interpolation. The rule-based detection and correction of outliers and the Z-score standardization method eliminate the influence of dimensions, resulting in a clean and scale-uniform data sequence. The weighted time alignment is used to solve the problem of time sequence misalignment caused by differences in construction speed among different bridge segments. Specifically, it introduces the importance weight of the process through a weighted dynamic time warping method to obtain a data sequence with time axis alignment. The effect decoupling is used to separate the pure load response component from the total data. Specifically, it decomposes the temperature effect and the time-dependent creep effect through a regression model to obtain the load response time series data mainly caused by the construction load. The dataset segmentation is used to obtain training data and test data, specifically by segmenting the original dataset of past deformation predictions. The real-time deformation prediction raw dataset is pre-processed through data standardization, weighted temporal alignment, and effect decoupling to obtain a dataset to be predicted. The past deformation prediction raw dataset is pre-processed through data standardization, weighted temporal alignment, effect decoupling, and dataset segmentation to obtain a feature extraction training set and a feature extraction test set.

Images