A method for determining the installation elevation of main beam
Through the Informer model, the informer model integrates information such as ambient temperature, construction load and foundation settlement, and builds time series characteristics and deep neural networks, solving the accuracy and efficiency problems of main beam installation elevation calculation in the existing technology, achieving more accurate prediction and higher construction control accuracy.
Patent Information
- Application Number
- CN202411132957.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-19
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-08-19
AI Technical Summary
The existing main beam installation elevation calculation method depends on theoretical calculation and finite element simulation, and it is difficult to accurately consider complex external environmental factors and the characteristics of the structure itself. It is inefficient in computing, difficult to reflect structural state changes in real time, and difficult to effectively integrate multiple influencing factors. This leads to a deviation from the actual situation.
The main beam installation elevation calculation model MF-GEI based on Informer is adopted. By constructing time series features, training deep neural network models, integrating information such as ambient temperature, construction load, foundation settlement, etc., using feature attention mechanism and adaptive Encoder-Decoder interaction Mechanism, multi-scale spatiotemporal and factor decomposition output to achieve accurate prediction of the main beam elevation.
Achieve more accurate and comprehensive prediction of the main beam elevation, able to adaptively learn the correlation and importance of different features, comprehensively consider the complex impact of various factors on the elevation, and improve construction control accuracy and automation intelligence. level.
Smart Images

Figure QLYQS_5 
Figure QLYQS_16 
Figure QLYQS_21
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of bridge construction, and in particular to a method for determining the installation elevation of a main beam. Background Art
[0002] As a key infrastructure in the transportation network, bridges play an indispensable role in modern society. As the span and scale of bridges continue to increase, the requirements for the safety and durability of bridge structures are also increasing. In the process of bridge construction, the installation of the main beam is one of the most critical links, and its construction quality directly affects the overall performance and service life of the bridge. The precise control of the main beam installation elevation is an important prerequisite for ensuring the quality of the main beam installation.
[0003] The traditional method of calculating the installation elevation of the main beam mainly relies on theoretical calculations and finite element simulations. For example, the patent publication number CN116306087, a temperature filtering calculation method and system for the installation elevation of the main beam, discloses a method for calculating the temperature influence quantity, combined with finite element model analysis, and finally obtains the actual installation control elevation of the beam section to be erected. Although the methods in the prior art can predict the deformation behavior of the main beam to a certain extent, there are still some limitations. First, the theoretical calculation method is usually based on simplified assumptions and boundary conditions, and it is difficult to accurately consider the influence of complex external environmental factors and the characteristics of the structure itself. Secondly, the model established by finite element simulation has low calculation efficiency and it is difficult to reflect the changes in the structural state in real time. In addition, whether it is theoretical calculation or finite element simulation, it is difficult to effectively integrate a variety of different influencing factors, resulting in a certain deviation between the calculation results and the actual situation.
[0004] In recent years, the rapid development of information technology has provided new ideas for solving complex construction problems. For example, Liu Renming proposed a method of applying BP neural network in the construction control of composite beam cable-stayed bridges in his paper "Application of BP neural network in the construction control of composite beam cable-stayed bridges". By establishing a neural network model, the parameters that affect the bridge line shape and internal force are identified, thereby improving the accuracy of construction control and overcoming the deviation between theoretical values and actual values.
[0005] In the process of studying related technologies, we found that Informer showed unique advantages in the calculation of main beam installation elevation. First, its excellent sequence modeling ability enables it to effectively capture the dynamic characteristics of long-term influencing factors such as ambient temperature, construction load, and foundation settlement. Through the ProbSparse self-attention mechanism, Informer significantly reduces the computational complexity and can process longer input sequences, thereby more comprehensively considering the impact of historical data on future elevations. Secondly, Informer's self-attention mechanism naturally has the ability to fuse multiple data. In the process of main beam installation, multiple data sources such as temperature, load, and settlement are involved. Informer can adaptively learn the correlation and importance between different features and achieve effective integration of these multi-source data. This fusion capability enables the model to comprehensively consider the complex impact of various factors on the elevation, thereby providing more accurate and comprehensive prediction results. In addition, Informer's self-attention mechanism can also capture the dependencies in time and space dimensions at the same time, which is crucial for accurately characterizing the deformation behavior of the main beam. In summary, we believe that Informer's advantages in sequence modeling and multiple data fusion make it an ideal choice for dealing with the complex engineering problem of main beam installation elevation prediction.
[0006] In view of this, we proposed an Informer-based main girder installation elevation calculation model MF-GEI (Informer Based Multi Factor Main Girder Elevation Model) to determine the main girder installation elevation. This model can fully consider a variety of information including ambient temperature, load action, foundation settlement, etc., and train a deep neural network model by constructing time series features to achieve accurate prediction of the main girder elevation. Summary of the invention
[0007] In view of the existing problems in the calculation of the main girder installation elevation, the main girder installation elevation calculation model MF-GEI (Informer Based Multi Factor Main Girder Elevation Model) is used to complete the calculation of the main girder installation elevation. Specifically, a method for determining the main girder installation elevation is used. This method comprehensively considers key influencing factors such as ambient temperature, construction load, foundation settlement, etc., and determines the main girder installation elevation through MF-GEI:
[0008] Integrate the multi-dimensional feature representation of key influencing factors. Construct the monitoring data of factors such as ambient temperature, construction load, foundation settlement, etc. into a time series feature matrix ,in is the sequence length, is the total feature dimension. The feature attention mechanism is used to adaptively weight different types of features to highlight key information. For each type of feature subsequence, the attention vector To adjust the importance of its different dimensions:
[0009] ;
[0010] in represents broadcast multiplication. The attention vector is generated by a two-layer feedforward neural network. They represent the ambient temperature, construction load, key factors affecting foundation settlement and other auxiliary factors. For the The dimension of the class factor.
[0011] Adopt an adaptive Encoder-Decoder interaction mechanism. Introduce an adaptive interaction matrix at each layer of the Decoder To adjust its attention to the Encoder output and capture the dynamic changes of influencing factors, among which, is the sequence length of the Decoder; It is the dimension of each layer of Decoder. The calculation formula is:
[0012] ;
[0013] in is the output of the previous layer of Decoder, is the final output of the Encoder. Modify the attention calculation formula to:
[0014] ;
[0015] Multi-scale spatiotemporal feature fusion: One-dimensional convolution and pooling operations with different kernel sizes are used in the Encoder and Decoder to extract multi-scale features.
[0016] Factor decomposition output: The final output of the Decoder Through three independent fully connected layers, it is mapped to the three latent spaces of temperature, load, and settlement to obtain the corresponding latent space representation , is the number of layers of the Decoder. Then, adaptive fusion is performed through the factor attention mechanism to obtain the final predicted value of the main beam elevation:
[0017] ;
[0018] in is the attention weight of each factor.
[0019] Using multi-task loss function: considering the overall prediction loss at the same time and the independent losses predicted by each factor , guiding the model to take into account both overall accuracy and accurate characterization of influencing factors:
[0020] .
[0021] The Early Stopping strategy is used during model training to control the number of iterations and update parameters.
[0022] During inference, the prediction results are post-processed, including moving average smoothing, outlier removal, and dynamic error compensation. The moving average smoothing uses a length of The sliding window of the forecast sequence is smoothed, and the outlier removal will exceed the range of historical data The predicted value is truncated, and the dynamic error compensation uses the measured value Correct the predicted value:
[0023] ;
[0024] in is the compensation coefficient.
[0025] The main beam installation elevation prediction method proposed in the present invention can fully explore and integrate the complex spatiotemporal characteristics of key influencing factors such as ambient temperature, construction load, foundation settlement, and establish an end-to-end mapping from heterogeneous monitoring data to main beam elevation prediction values. Through a carefully designed deep learning model architecture, including feature attention, adaptive Encoder-Decoder interaction, multi-scale feature fusion, factor decomposition output and other mechanisms, this method can accurately capture the dynamic change law of the main beam elevation, and analyze and explain the contribution of each influencing factor. In the model training and reasoning links, parameter optimization, Early Stopping, post-processing and other strategies are adopted to further ensure the accuracy and robustness of the prediction results. The present invention can provide a reliable elevation reference for the main beam installation construction, improve the level of construction automation and intelligence, and has engineering application value. DETAILED DESCRIPTION
[0026] In order to explain the present invention more clearly, the present invention is further described below in conjunction with an embodiment. A method for determining the main beam installation elevation is used to calculate the main beam installation elevation by using the main beam installation elevation calculation model MF-GEI based on Informer. The specific scheme is as follows:
[0027] The calculation of the main beam installation elevation is defined as a multivariate time series regression problem. Given a historical observation sequence, the MF-GEI model predicts the main beam elevation values at several future moments. Assume that the input features are ,in is the length of the historical sequence, is the feature dimension. It can be expressed as:
[0028] ;
[0029] Input features include ambient temperature, construction load, and foundation settlement information. The output of the model is the future The main beam elevation prediction sequence at each moment .
[0030] The optimization goal of the model is to minimize the mean absolute error (MAE) between the predicted value and the true value:
[0031] ;
[0032] in, are model parameters, and Respectively The predicted value and actual value of the main beam elevation at a certain moment.
[0033] Temperature change is an important factor affecting the main beam elevation. This paper collects real-time temperature monitoring data from multiple points on the surface and inside of the beam to characterize the temporal and spatial distribution characteristics of the temperature field. temperature sensor, the temperature characteristic can be expressed as :
[0034] ;
[0035] in, Indicates Moment In order to extract the spatial characteristics of the temperature field, this paper performs weighted averaging of the temperature values at different measuring points:
[0036] ;
[0037] in, For the The weight coefficient of each measuring point reflects the influence of the temperature at that point on the main beam elevation. The weight can be set according to the location of the measuring point and experience. By weighted averaging, a scalar characteristic reflecting the overall temperature effect can be obtained. .
[0038] In order to characterize the time evolution characteristics of temperature, this paper further calculates the differential characteristics of historical temperature values:
[0039] ;
[0040] It reflects the rate of change of temperature and helps capture the dynamic effect of temperature. Finally, the temperature characteristic can be expressed as .
[0041] Construction load is the direct factor affecting the deformation of the main beam. If there are loads of this type (including but not limited to static loads such as crane load, deadweight of construction equipment, load of construction personnel, and dynamic loads such as wind load, rain and snow load), the load characteristics can be expressed as :
[0042] ;
[0043] in, Indicates Moment Similar to the temperature characteristics, this paper performs weighted summation of different types of loads to obtain the comprehensive load effect:
[0044] ;
[0045] in, For the The weight coefficient of the class load is calculated. At the same time, the differential characteristics of the comprehensive load are calculated. , reflecting the dynamic changes of the load. Finally, the load characteristics can be expressed as .
[0046] Uneven settlement of the pier foundation will cause changes in the elevation of the superstructure. This paper uses a high-precision level and GPS to regularly observe the piers and obtain settlement data. Assume that the number of piers is , then the settlement characteristics can be expressed as :
[0047] ;
[0048] in, Indicates Moment In order to quantify the overall effect of the settlement, this paper calculates the weighted average of the settlement of each pier:
[0049] ;
[0050] in, For the The weight coefficient of each pier reflects the influence of the pier settlement on the main beam elevation. The weight can be set according to the location and force characteristics of the pier. In addition, the settlement rate characteristics It can reflect the dynamic trend of settlement. Finally, the settlement characteristics can be expressed as .
[0051] In terms of feature normalization, this paper adopts the maximum and minimum normalization method to map the temperature, load, settlement and other dimensional features into the range of [0, 1] respectively, eliminating the differences in dimension and numerical range, and improving the training efficiency and generalization performance of the model.
[0052] After normalization, the final model input features can be expressed as:
[0053] ;
[0054] in, Represents other auxiliary feature vectors, such as construction stage, segment type, etc.
[0055] Constructing the MF-GEI model integrating key influencing factors:
[0056] The ME-GEI model based on Informer has significant advantages in processing sequential time series data. In order to better predict the main beam installation elevation, this paper makes targeted improvements to the original Informer model to fully integrate key influencing factors such as ambient temperature, construction load and foundation settlement. The ME-GEI model includes:
[0057] Input encoding layer:
[0058] The input of the model includes three key features: ambient temperature, construction load, foundation settlement, and other auxiliary features, forming a multi-channel time series data representation. Suppose the input sequence is ,in is the sequence length, is the total feature dimension. It can be further expressed as:
[0059] ;
[0060] in, are the temperature, load and settlement feature subsequences, respectively. are other auxiliary feature subsequences.
[0061] For each type of feature subsequence, an attention vector is used to weight the importance of its different dimensions:
[0062] ;
[0063] in, is the weighted feature subsequence, is the corresponding attention vector, represents broadcast multiplication. The attention vector is generated through a two-layer feedforward neural network:
[0064] ;
[0065] in, are the parameters of the attention network, is the middle-level dimension, Indicates averaging over the time dimension.
[0066] By introducing feature-level attention, the model can adaptively adjust the weights of different feature dimensions, highlighting the role of key influencing factors such as ambient temperature, construction load, and foundation settlement, while suppressing the interference of unimportant features. The weighted feature subsequences are concatenated in the channel dimension and mapped to a high-dimensional space through position encoding to form the final input embedding. :
[0067] ;
[0068] in, is the embedding matrix, is the position encoding matrix.
[0069] Adaptive Encoder-Decoder Interaction:
[0070] In the original Informer model, the Encoder and Decoder interact with each other through the attention mechanism, but the intensity of the interaction remains unchanged throughout the calculation process. Considering that the effects of ambient temperature, construction load and foundation settlement on the main beam elevation may vary in different periods, this paper proposes an adaptive Encoder-Decoder interaction mechanism.
[0071] Specifically, in the In the calculation of the layer decoder, an adaptive interaction matrix is introduced To adjust the Decoder's attention to the Encoder output. The calculation formula is:
[0072] ;
[0073] in, For the The output of layer Decoder, is the final output of Encoder, are the parameters of the interaction matrix.
[0074] When calculating the Encoder-Decoder attention, the adaptive interaction matrix is multiplied by the value matrix to obtain the adjusted attention output:
[0075] ;
[0076] in, is the query matrix, is a key-value matrix. Through the adaptive interaction matrix, the Decoder can dynamically adjust the degree of attention to the Encoder output according to its own state, highlight the influencing factors of the critical period, and improve the timeliness of the prediction.
[0077] Multi-scale spatiotemporal feature fusion:
[0078] Since the influence of factors such as ambient temperature, construction load, and foundation settlement on the main beam elevation has multi-scale spatiotemporal characteristics, this paper introduces a multi-scale feature fusion module in the Encoder and Decoder to better capture the influence patterns at different spatiotemporal scales.
[0079] In the Encoder Layer, the calculation process of the multi-scale feature fusion module is as follows:
[0080] ;
[0081] ;
[0082] in, The kernel size is The one-dimensional convolution operation, Indicates the step length is The one-dimensional maximum pooling operation, is the number of scales, is the feature fusion matrix, is the dimension of the convolution output.
[0083] Accordingly, in the Decoder The multi-scale feature fusion module is introduced at the layer:
[0084] ;
[0085] ;
[0086] in, Indicates the step length is The one-dimensional upsampling operation is is the feature fusion matrix.
[0087] Through multi-scale feature fusion, the model can simultaneously extract and combine influence patterns at different temporal and spatial scales, enhance the ability to characterize complex nonlinear relationships, and thus improve the accuracy of main beam elevation prediction.
[0088] Factorization output layer:
[0089] In the main beam installation elevation prediction task, the contributions of ambient temperature, construction load and foundation settlement to the prediction results may be different, and this difference may change dynamically over time. In order to more accurately evaluate the impact of each factor, this paper designs a factor decomposition output layer and the corresponding loss function.
[0090] For the final output of the Decoder , firstly maps it to the latent space corresponding to the three factors of temperature, load and settlement through three independent fully connected layers:
[0091] ;
[0092] in, are the latent space representations corresponding to temperature, load and settlement factors, respectively. is the corresponding fully connected layer.
[0093] Then, the three latent space representations are adaptively fused through the factor attention mechanism to obtain the main beam elevation value:
[0094] ;
[0095] ;
[0096] in, is the attention weight of temperature, load and settlement factors, is the attention parameter.
[0097] Loss function:
[0098] In order to better guide model learning, this paper designs a multi-task loss function that takes into account both the overall error of the predicted value and the independent error of each factor prediction:
[0099] ;
[0100] in,
[0101] ;
[0102] ;
[0103] ;
[0104] ;
[0105] in, is the overall prediction loss, Independent losses predicted for temperature, load and settlement factors, is the balance coefficient. This loss function can guide the model to improve the overall prediction accuracy while also focusing on the accurate characterization of the impact of each key factor.
[0106] Through the factor decomposition output layer and multi-task loss function, the model can have a deeper understanding of the impact mechanism of ambient temperature, construction load and foundation settlement on the main beam elevation, and provide more interpretable and operational prediction results.
[0107] Based on the above model architecture and loss function, the training process of the ME-GEI model is Algorithm 1.
[0108] Algorithm 1 Training algorithm of the main beam installation elevation prediction model:
[0109] Input: training set , validation set , the maximum number of training rounds , number of early stop rounds ;
[0110] Output: trained model parameters .
[0111] 1: Randomly initialize model parameters ;
[0112] 2: ;
[0113] 3: while and do;
[0114] 4: ;
[0115] 5: for do;
[0116] 6: ;
[0117] 7: ;
[0118] 8: ;
[0119] 9: end for;
[0120] 10: ;
[0121] 11: if then;
[0122] 12: ;
[0123] 13: ;
[0124] 14: ;
[0125] 15: else;
[0126] 16: ;
[0127] 17: end if;
[0128] 18: end while;
[0129] 19: return ;
[0130] In each training round, for each sample in the training set First, the overall prediction value is obtained by forward calculation of the ME-GEI model And independent predictions for temperature, load, and settlement factors Then, the total loss is calculated based on the multi-task loss function , and update the model parameters .in, is the learning rate, is the exponential decay rate of the first and second moment estimates.
[0131] At the end of each round of training, Evaluate the model performance on the validation loss .if Lower than the current best validation loss , then update and the optimal model parameters , and reset the early stop counter Otherwise, increase When the number of training rounds exceeds Or the early stop counter exceeds the threshold When , the training process terminates and returns the optimal model parameters .
[0132] Using the trained ME-GEI model, new monitoring data can be inferred in real time to predict the main beam elevation at several moments in the future. At the same time, the model can also output independent prediction values of three key factors: ambient temperature, construction load, and foundation settlement, to evaluate their contribution to the prediction results. The reasoning process is shown in Algorithm 2.
[0133] Algorithm 2 Reasoning algorithm of the main beam installation elevation prediction model:
[0134] Input: trained model parameters , new monitoring data ;
[0135] Output: Predicted value of main beam elevation , Temperature factor prediction value , load factor prediction value , predicted value of settlement factor .
[0136] 1: ;
[0137] 2: ;
[0138] 3: ;
[0139] 4: ;
[0140] 5: ;
[0141] 6: return .
[0142] First, use the trained model parameters New monitoring data Perform forward calculation to obtain the overall predicted value of the main beam elevation And the independent predictive value of each factor .
[0143] Then, the prediction results are post-processed. The function includes the following steps:
[0144] Moving average smoothing: Use a moving average of length The sliding window of is used to smooth the forecast sequence and eliminate short-term fluctuations:
[0145] ;
[0146] Outlier elimination: Eliminate the predicted values that are beyond the range of historical data. Outliers:
[0147] ;
[0148] Dynamic error compensation: using measured values Dynamically correct the predicted value to reduce the cumulative error:
[0149] ;
[0150] in, is the compensation coefficient.
[0151] Finally, the overall predicted value of the main beam elevation after post-processing is returned And the predicted value of each factor These prediction results can be directly used to guide the main beam installation construction and evaluate the impact of ambient temperature, construction load, and foundation settlement on the elevation.
Claims
1. A method for determining the main beam installation elevation, characterized in that: The method uses key influencing factors including ambient temperature, construction load, foundation settlement and other auxiliary factors to determine the main beam installation elevation through MF-GEI. The MF-GEI is an informer-based main beam installation elevation calculation model, which specifically includes: multi-dimensional feature representation of key influencing factors; multi-scale spatiotemporal feature fusion; factor decomposition output; multi-task loss function; Early Stopping strategy is used to control the number of iterations and update parameters to train MF-GEI; the trained MF-GEI is used to obtain the overall prediction value of the main beam elevation and the independent prediction value of each key influencing factor; the overall prediction value is post-processed to determine the final main beam installation elevation. The multi-dimensional feature representation of the key influencing factors is specifically: the monitoring data of ambient temperature, construction load and foundation settlement factors are constructed as a time series feature matrix ,in is the sequence length, is the total feature dimension; different types of features are adaptively weighted through the feature attention mechanism; for each type of feature subsequence, the attention vector is used To adjust the importance of its different dimensions: ; in represents broadcast multiplication, and the attention vector is generated through a two-layer feedforward neural network; They represent ambient temperature, construction load, key factors affecting foundation settlement and other auxiliary factors, among which, are the temperature, load and settlement feature subsequences, respectively. are other auxiliary feature subsequences, is the weighted feature subsequence, is the corresponding attention vector, For the Dimensions of class factors; is the weighted feature subsequence, For the corresponding attention vector, an adaptive Encoder-Decoder interaction mechanism is adopted, and an adaptive interaction matrix is introduced at each layer of the Decoder To adjust its attention to the Encoder output and capture the dynamic changes of influencing factors; is the sequence length of the Decoder, is the dimension of each layer of Decoder; the factor decomposition output includes: the final output of Decoder Through three independent fully connected layers, it is mapped to the three latent spaces of ambient temperature, construction load, and foundation settlement to obtain the corresponding latent space representation , where adaptive fusion is performed through the factor attention mechanism to obtain the main beam elevation value: ; in are the attention weights of key influencing factors of ambient temperature, construction load, and foundation settlement, is the number of layers of the Decoder; The multi-task loss function is: ,in For the overall loss, are the independent losses of ambient temperature, construction load and foundation settlement factors, is the balance coefficient, is the model parameter.
2. The method for determining the main beam installation elevation according to claim 1, characterized in that: The multi-scale spatiotemporal feature fusion is specifically: using one-dimensional convolution and pooling operations with different kernel sizes in the Encoder and the Decoder to extract multi-scale features.
3. The method for determining the main beam installation elevation according to claim 2, characterized in that: The EarlyStopping strategy is used to control the number of iterations and update the parameters of MF-GEI training as follows: In each training round, for each sample in the training set First, the overall prediction value is obtained by forward calculation of the ME-GEI model And independent prediction values of ambient temperature, construction load, and foundation settlement factors ; Then, the total loss is calculated based on the multi-task loss function , and update the model parameters ; At the end of each round of training, Evaluate the model performance on the validation loss ;if Lower than the current best validation loss , then update and the optimal model parameters , and reset the early stop counter Otherwise, increase When the number of training rounds exceeds Or the early stop counter exceeds the threshold When , the training process terminates and returns the optimal model parameters .
4. The method for determining the main beam installation elevation according to claim 3, characterized in that: The method of using the trained MF-GEI to obtain the overall prediction value of the main beam elevation and the independent prediction value of each key influencing factor specifically includes: using the trained MF-GEI parameters New monitoring data Perform forward calculation to obtain the overall predicted value of the main beam elevation And the independent prediction values of three key influencing factors: ambient temperature, construction load and foundation settlement .
Citation Information
Patent Citations
Method for fusing monitoring data of small and medium-sized bridges
CN113255035A
Bridge structure parameter correction model establishment and prediction method based on time domain data
CN115510759A
Temperature filtering calculation method and system for girder installation elevation
CN116306087A
Ship traffic flow long sequence space-time prediction method based on ST-Informer
CN116821784A
Correction method for upward shift amount, and correction method and program for upward shift amount
JP2016003475A
