Temperature-response intelligent mapping method for large-span stiff skeleton arch bridge

Abstract

The application discloses a large-span stiff skeleton arch bridge temperature-response intelligent mapping method and belongs to the technical field of bridge structure health monitoring. S100: data acquisition; S200: data preprocessing, generating a data set, and dividing into a training set and a validation set; S300: constructing a Phy-SETCN model based on the training set / validation set, using MAE, RMSE, R 2 The strain data obtained by mapping the Phy-SETCN model is evaluated; S400: real-time detection of bridge temperature-induced strain and abnormal response early warning through the Phy-SETCN model. The application can be widely applied to a large-span bridge structure health monitoring system, effectively identifies temperature-induced responses through a high-precision temperature-strain mapping model, provides a reliable basis for bridge safety evaluation, and reduces operation and maintenance costs.

Description

Technical Field

[0001] This invention relates to the field of bridge structural health monitoring technology, and more specifically to a temperature-response intelligent mapping method for long-span stiffened arch bridges. Background Technology

[0002] Bridge structures are highly temperature sensitive, and temperature-induced responses often mask crucial information such as vehicle load effects and structural damage, leading to misjudgments of safety status. This is particularly true in long-span stiffened concrete arch bridges, where the large spatial scale and numerous statically indeterminate degrees of the structure result in non-uniform temperature field distribution, significant time-delay effects, and secondary temperature effects, severely hindering the accurate acquisition of temperature-induced responses.

[0003] In existing methods, the finite element model relies on accurate material parameters and is computationally complex, while data-driven methods (such as LSTM and CNN) lack physical interpretability. Furthermore, the non-uniform temperature field distribution and significant time-delay effects of long-span arch bridges further increase the difficulty of modeling temperature-induced responses.

[0004] Therefore, how to provide a smart mapping method for temperature-response of long-span stiffness frame arch bridges is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a temperature-response intelligent mapping method for long-span stiffened arch bridges to solve the technical problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A temperature-response intelligent mapping method for long-span stiffened arch bridges includes:

[0008] S100: Acquire multi-channel temperature data and corresponding strain data of a specified section in the bridge health monitoring system;

[0009] S200: Preprocess the multi-channel temperature data and the corresponding strain data at the measuring points to generate a dataset, and divide it into a training set and a validation set;

[0010] S300: Construct and train the Phy-SETCN model based on the training set, and verify the generalization ability of the Phy-SETCN model using the validation set, using MAE, RMSE, and R... 2 Evaluation of strain data obtained by mapping the Phy-SETCN model;

[0011] S400: Real-time detection and early warning of abnormal response of bridge temperature-induced strain using the Phy-SETCN model.

[0012] Preferably, S100 includes:

[0013] Based on the initial monitoring data of the bridge, the temperature increment and strain increment are calculated. The input is multi-channel temperature increment data, and the output is the residual between theoretical strain and measured strain.

[0014] Preferably, the output is the residual between the theoretical strain and the measured strain, including:

[0015] Based on the formula for calculating the additional internal force of an arch bridge under temperature action, the theoretical strain of the arch under temperature action in a scaled-down model is derived, and the residual between the theoretical strain and the measured strain is used as the output data of the model.

[0016] Preferably, S200 includes:

[0017] The input and output datasets are constructed using a sliding window.

[0018] The input and output data are subjected to min-max normalization to eliminate differences in dimensions.

[0019] The dataset is split into a training set and a validation set.

[0020] Preferably, the Phy-SETCN model integrates the SENet module and the TCN module, and introduces physical information to construct residuals as a driving force. The specific operation process is as follows:

[0021] S310: Use the permute() function to adjust the input data;

[0022] S320: After passing through the adaptive average pooling layer, the spatial dimension of the feature map is compressed and the shape is adjusted to match the input of the fully connected layer;

[0023] S330: Define two fully connected layers as activation operations, use the Sigmoid activation function, output the importance coefficient of each channel, and adjust the shape to match the shape of the original feature map;

[0024] S340: The output dimension of the SENet module is adjusted using the transpose(·) function to facilitate its input to the TCN module;

[0025] S350: The TCN module performs convolution operations on the input time series and uses residual connections to propagate feature information across layers, capturing nonlinear data features;

[0026] S360: Use the transpose() function to reshape the output shape of the TCN module for forwarding to the linear layer;

[0027] S370: Select the output of the last time step as the final prediction result.

[0028] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a temperature-response intelligent mapping method for long-span stiffened arch bridges, which has good generalization ability, is suitable for long-term analysis of actual bridge monitoring data, and has the following beneficial effects:

[0029] (1) SENet is used to dynamically adjust the channel weights of TCN, which alleviates the imbalance of TCN focusing too much on early information and ignoring recent information. The influence of non-uniform temperature distribution is considered through multi-channel input, and the need for preprocessing of the time lag effect of the original data is eliminated. The original data can be directly input into the model, which significantly improves the model efficiency and the ability to capture key time series features.

[0030] (2) By calculating the theoretical temperature-induced response of the rigid frame arch bridge, the theoretical strain and measured strain residuals were constructed during the neural network training process. The temperature-strain mapping model driven by the residuals was built, which significantly improved the mapping accuracy and interpretability of the model. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0032] Figure 1 This is a schematic diagram of the structure of the present invention;

[0033] Figure 2 Calculation diagram of additional internal forces in the main arch caused by temperature changes;

[0034] Figure 3 The diagram shows the location of each test section and the distribution of measuring points across the cross section.

[0035] Figure 4(a) is the time history curve corresponding to 3-T3;

[0036] Figure 4(b) shows the spectrum corresponding to 3-T3;

[0037] Figure 5(a) is the time history curve corresponding to 3-S2;

[0038] Figure 5(b) shows the spectrum corresponding to 3-S2;

[0039] Figure 6 The correlation analysis diagram of 3-T3 and 3-S2 at section 3 is shown;

[0040] Figure 7(a) is a scatter plot of the full-time correlation between the temperatures 15-T3 and 3-S2 of section 15, which is symmetrical about the arch of section 3;

[0041] Figure 7(b) is a scatter plot of the correlation between the temperatures 15-T3 and 3-S2 of section 15, which are symmetrical about the arch, on a certain day after downsampling.

[0042] Figure 8(a) shows the temperature probability density distribution.

[0043] Figure 8(b) shows the strain probability density distribution.

[0044] Figure 9(a) is a schematic diagram of the input temperature data for model training;

[0045] Figure 9(b) is a schematic diagram of the strain data output by the model training;

[0046] Figure 10 A detailed diagram illustrating the sliding window operation;

[0047] Figure 11 A schematic diagram illustrating the specific training process for each model;

[0048] Figure 12 A comparison diagram of the mapped strain and measured strain of the validation set;

[0049] Figure 13 Radial bar charts for the evaluation indicators of each model;

[0050] Figure 14 This is a schematic diagram of the validation set results for Phy-SETCN. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Example 1:

[0053] This invention discloses a temperature-response intelligent mapping method for long-span stiffened arch bridges. The overall process of this method is as follows: Figure 1 As shown, the specific steps are as follows:

[0054] (1) Data preparation: Using the measured value at the start time of the period of interest as the benchmark, the strain increment and temperature increment data within that period are calculated, and the temperature increment is used as the input. Based on the calculation formula of the additional internal force under temperature action of the arch bridge, the theoretical strain under temperature action of the scaled model arch in this embodiment is derived, and the residual between the theoretical strain and the measured strain is used as the model output data.

[0055] (2) Dataset Partitioning: An appropriate input-output correspondence is established using a sliding window. Based on model complexity and dataset size, the dataset is reasonably partitioned to ensure effective model training. The training set is X1 = {x1, x2, ..., x...}. c-1 ,x c The validation set is X2 = {x} c+1 ,x c+2 ,...,x c+m-1 ,x c+m}, where c and m are the lengths of the training set data and the validation set data, respectively.

[0056] (3) Data normalization: In order to improve the training speed and stability of the model and enhance the performance of the final model, this embodiment uses max-min normalization to eliminate the negative impact of data magnitude differences on subsequent model training and prediction.

[0057] (4) Model training: The training set is used for feature learning, the validation set is used to adjust the hyperparameters of the model, and the trained model is used to map the strain of unknown time periods to evaluate the generalization ability of the model.

[0058] (5) Denormalization: In order to ensure that all model output values ​​are consistent with the range and dimension of the original data, the model output values ​​are denormalized.

[0059] (6) Data Reconstruction: The inversely normalized output, i.e., the mapped residual, is added to the theoretical values ​​corresponding to the time sequence to obtain the mapped strain increment. MAE, RMSE, and R... 2 The strain data obtained from model mapping are evaluated.

[0060] This embodiment simplifies the temperature load calculation by performing the temperature effect calculation in two steps: first, the self-stress and axial free deformation of a micro-segment of the stiffened concrete frame are calculated; then, the axial deformation of the micro-segment is substituted into the basic arch structure to calculate the additional internal forces. The calculation diagram of the additional internal forces of the main arch caused by temperature changes is shown below. Figure 2 As shown.

[0061] The deformation compatibility condition between steel pipe and concrete under uniform temperature change is:

[0062]

[0063] Where: ε z σ represents the axial free strain of the micro-segment; s σ c1 and σ c2 E represents the axial self-stress of the steel, the concrete inside the pipe, and the concrete casing, respectively. s E c1 and E c2 These are the elastic moduli of steel, concrete inside the pipe, and concrete outside the pipe, respectively, α.s α c1 and α c2 The coefficients of linear expansion for steel, concrete inside the pipe, and concrete outside the pipe are ΔT, respectively. s ΔT c1 and ΔT c2 These represent the temperature changes of the steel, the concrete inside the pipe, and the outer concrete, respectively.

[0064] The internal force equilibrium condition of the composite section is:

[0065] σ s A s +σ c1 A c1 +σ c2 A c2 =0 (2)

[0066] In the formula: A s A c1 and A c2 These are the cross-sectional areas of the steel, the concrete inside the pipe, and the concrete covering the outside, respectively.

[0067] Combining equations (1) and (2), we get:

[0068]

[0069] ε z Substitution Figure 2 The additional internal forces of the arch in the basic structure shown are calculated, and the redundant constraint forces of the basic system are:

[0070]

[0071] Δl t =ε z l (8)

[0072] In the formula, H t The horizontal force generated at the elastic center by temperature change; δ′ 42 Δl is the horizontal displacement at the elastic center caused by a unit horizontal force. t The displacement of the arch axis in the horizontal direction caused by temperature changes; EA and EI are the axial stiffness and bending stiffness of the composite section, respectively; l is the calculated span; The horizontal angle is defined at any position on the arch axis.

[0073] Therefore, the additional internal forces at the composite section are:

[0074] M t =-H t y = -H t (y s -y1) (9)

[0075]

[0076] In the formula, M t and N t y represents the additional bending moment and axial force at any cross section of the arch caused by temperature changes. s Let y1 be the location of the elastic center and y1 be the distance from any cross section to the top of the arch.

[0077] The secondary internal forces in the steel pipe, core concrete, and outer concrete sections are:

[0078]

[0079] In one specific embodiment, model training specifically includes:

[0080] This embodiment uses a temperature-strain mapping model that integrates SENet and TCN, and introduces physically-based residuals as the model driver. This combination aims to fully leverage the strengths of each model to improve mapping accuracy. In this composite model, the task allocation for the different models is as follows:

[0081] (1) SENet explicitly models the dependencies between feature channels and recalibrates the channel weights to emphasize key features and suppress useless features, thereby enhancing the representational ability of convolutional neural networks and solving the problem of TCN focusing on early information while ignoring recent information in the receptive field.

[0082] (2) Utilize designs such as extended causal convolution and residual connections in TCN to help the model effectively capture the linear and nonlinear characteristics of signal data sequences.

[0083] (3) The theoretical response of the structure under temperature is derived from the mechanical principle. The model is driven by the residual between the theoretical response and the measured response, and interpretable deep learning with physical mechanism embedded is realized.

[0084] The specific operation process of this combined model (Phy-SETCN model) is as follows:

[0085] Step 1: Use the permute() function to adjust the input data to match the input size of the subsequent model;

[0086] Step 2: After passing through the adaptive average pooling layer, the spatial dimension of the feature map is compressed to 1x1, and the shape is adjusted to match the input of the fully connected layer;

[0087] Step 3: Define two fully connected layers as activation operations, use the Sigmoid activation function, output the importance coefficient of each channel, and adjust the shape to match the shape of the original feature map;

[0088] Step 4: The transpose(·) function is used to adjust the dimensions of the SENet output to facilitate its input to the TCN;

[0089] Step 5: TCN performs convolution operations on the input time series and uses residual connections to propagate feature information across layers, capturing nonlinear data features;

[0090] Step 6: Use the transpose(·) function to reshape the output shape of the TCN to facilitate its forwarding to the linear layer;

[0091] Step 7: Finally, select the output of the last time step as the final prediction result.

[0092] In one specific embodiment, to evaluate the model's performance, mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are selected. 2 As evaluation metrics, these metrics have been widely used in related research. RMSE amplifies large errors caused by squaring, making it sensitive to large errors between measured and predicted values. In contrast, MAE treats all errors equally and assigns equal weight to each error. Therefore, MAE and RMSE can effectively evaluate overall error and extreme errors, respectively. The closer the values ​​of MAE and RMSE are to 0, the smaller the model error. 2 The goodness of fit between predicted and observed values ​​is measured; the closer the value is to 1, the better the regression model's performance. The three evaluation metrics are calculated using the following formulas:

[0093] In the formula, m is the sample size, and P i Represents the i-th predicted value, O i Represents the i-th measured value. This represents the average value of the measured data.

[0094] Example 2:

[0095] This embodiment presents a measurement scheme for acquiring temperature and strain data, and performs data processing and analysis on the measured temperature and strain. The time-frequency characteristics of temperature and strain within the time period of interest, as well as the correlation between them, are clarified, and the non-Gaussian properties of the data are analyzed.

[0096] This embodiment is based on a 1:10 scale model (60 meters) of the world's largest span arch bridge. This bridge is a rigid-framed concrete arch bridge with a catenary axis for the main arch. The calculated span is 600m, the rise is 125m, the rise-to-span ratio is f = 1 / 4.8, and the arch axis coefficient is m = 1.9. The rigid frame of this scaled model is entirely constructed using Q420D material. The internal concrete is C80 fine-grained concrete with good fluidity, while the external concrete is C50 fine-grained concrete.

[0097] Specifically, the model has a total of 17 strain test sections, and the location of each test section and the distribution of measuring points on the cross section are detailed. Figure 3 As shown, fiber optic sensors were used in all tests. At each test section (2-16), 14 strain gauges and 3 temperature gauges were installed. At the arch foot sections (1 and 17), an additional 12 strain gauges were installed at the steel pipe. The model had a total of 262 strain gauges and 51 temperature gauges.

[0098] This embodiment selects temperature data at point T3 and strain data at point S2 at three cross-sections from July 4th to July 12th, 2023 for data analysis. To facilitate the analysis of the correlation between temperature and strain, and to conform to the data characteristics of actual bridge health monitoring (health monitoring system data is generally incremental), the data at the initial moment within this time period is used as the initial value. The incremental value within the time period is obtained by subtracting the initial value from the measured data. All data mentioned thereafter are incremental values. The time history curves and corresponding spectrum diagrams for 3-T3 and 3-S2 are shown below. Figures 4(a)-4(b) as well as Figures 5(a)-5(b) As shown.

[0099] observe Figures 4(a)-4(b) as well as Figures 5(a)-5(b) It can be observed that strain changes in the opposite direction to temperature; strain decreases as temperature increases. Furthermore, from... Figures 4(a)-4(b) as well as Figures 5(a)-5(b) It can be observed that the spectral characteristics of temperature and strain are extremely similar, exhibiting a clear diurnal periodic variation. (The period corresponding to a frequency of 1.158E-5 is approximately 86356s, roughly equivalent to 86400s of a day). After analyzing the time-frequency characteristics of the data, the correlation between temperature and strain was further investigated. The scatter plot of the correlation between temperature and strain is shown below. Figure 6 and Figures 7(a)-7(b) As shown.

[0100] Figure 6 The correlation analysis of 3-T3 and 3-S2 at section 3 is presented, i.e., the temperature-strain correlation analysis of the same section. Figures 7(a)-(b) show the correlation analysis of temperature 15-T3 and 3-S2 at section 3, which is symmetrical about the arch crown at section 15, i.e., the temperature and strain data come from different sections with a large longitudinal span. Figure 6 and Figures 7(a)-7(b) It can be observed that temperature and strain generally exhibit a significant negative correlation. Figure 6The scatter points in Figure 7(a)-(b) are more concentrated and the correlation is more significant. However, the scatter points in Figures 7(a)-(b) exhibit obvious loop closure characteristics. To study this characteristic more thoroughly, the data was downsampled, and data from a specific day was selected and presented in Figure 7(b). Figure 7(b) shows that strain decreases as temperature increases and increases as temperature decreases, but the same temperature corresponds to different strains, with a maximum difference of 29.34 με. The above analysis indicates that the temperature distribution is non-uniform, the temperature-strain correlation is more significant at the same cross-section, and there is a relatively obvious time lag effect between temperature and strain at different cross-sections. In this case, the correlation between temperature and strain cannot be described by a simple linear relationship.

[0101] Higher-order statistical properties are primarily determined through skewness and kurtosis. Data is considered non-Gaussian when skewness is not equal to 0 or kurtosis is not equal to 3. This embodiment uses excess kurtosis to evaluate kurtosis; data is considered non-Gaussian when kurtosis is not equal to 0. For probability density functions, comparing them with a standard Gaussian distribution provides a more intuitive assessment. The probability density function is as follows: Figures 8(a)-8(b) As shown.

[0102] The calculated kurtosis values ​​for temperature and strain data are -0.43 and -0.47, respectively, both less than 0, indicating that the probability density distributions of temperature and strain have light tails. This characteristic is consistent with the features shown in Figures 8(a)-(b), indicating fewer extreme values ​​compared to a standard normal distribution. Regarding skewness, the temperature skewness is 0.065, showing positive skewness with a slightly rightward bias. The strain skewness is -0.081, showing negative skewness with a slightly leftward bias, consistent with Figures 8(a)-(b). Furthermore, three methods were used to test the normality of the data, and the results are shown in Table 1. All three methods showed that at the 0.05 level, the temperature and strain data do not significantly originate from a normally distributed population. In conclusion, the temperature and strain data in this scheme are non-Gaussian.

[0103] Table 1 Results of the normality test

[0104]

[0105] Specifically, regarding the construction of the dataset, data from July 4th to July 20th, 2023, a total of 17 days, were selected for the study. Sampling was performed once per minute, resulting in a data length of 24480. The 3-T3 temperature range and its difference from 15-T3 were used as the model input data, and 3-S2 was used as the model output data, as shown below. Figures 9(a)-9(b) As shown.

[0106] As shown in Figure 9(a), the temperature difference trend is roughly similar to that of 3-T3. Comparing the two subplots of Figure 9(a) and (b), it can be seen that the strain mainly fluctuates within the range of [-140, 40], which is much larger than the temperature data. During the training of a neural network model, features with large numerical values ​​have too much influence on the model, while features with small numerical values ​​may be ignored. If the distribution of the input data is very different, it will also lead to very drastic changes in the gradient, making the gradient descent algorithm unstable during training, or even unable to converge. Moreover, many operations during training involve numerical calculations, and an excessively large range of input data will make the related operations complex and time-consuming. Therefore, before further constructing the dataset for training, the data needs to be standardized or normalized. Through standardization or normalization, all features can be scaled to the same numerical range, thereby eliminating dimensional differences, making the model easier to converge, and making the data distribution more stable. This helps the gradient descent algorithm find the optimal solution faster, improving the model's running efficiency while having more stable performance. This embodiment selects the min-max normalization method to scale the data to [0,1]. This method is a linear transformation and is suitable for situations where the range of input data is known (maximum and minimum values ​​are known). The specific formula is as follows:

[0107]

[0108] In the formula, x min x max Let x' be the minimum and maximum values ​​of the data, and x' be the normalized value. As shown in the formula, it conforms to the form y = kx + b, which is a linear transformation. Its inverse transformation is:

[0109] x=x′·(x max -x min )+x min (twenty one)

[0110] The temperature and strain data used are typical time series data. After normalization, a sliding window method was chosen to construct the dataset. The sliding window method can extract fixed-length subsequences (windows) from continuous time series data. Each window contains information from a past period, effectively capturing the time dependence of the data. Considering the sampling frequency of the monitoring data, this embodiment sets the window size to 1440, meaning one window can cover a whole day's worth of data, enhancing the ability to capture periodic features. Furthermore, by using a sliding window, multiple overlapping samples can be created from the original data, increasing the size of the training set. More training samples help the model learn richer features, improving the model's training performance. Therefore, the sliding window step size is set to 1. The specific sliding window sampling process is as follows... Figure 10As shown, the temperature sliding window data sequence is used as input, and the strain data with a sequence length of 1 is used as output. The sampling time of the strain corresponds to the sampling time of the 1440th temperature data in the input temperature data sequence. Subtracting the window size, 23040 pairs of input and output data are finally obtained, where the input temperature data size is (23040, 1440, 2) and the output strain data size is (23040, 1).

[0111] Specifically, to accurately evaluate the fitting performance of the trained neural network, the dataset was divided into a training set (80%, 19104 pairs) and a validation set (20%, 4776 pairs). Only the training set data was used when training the neural network model parameters; the validation set data was used to verify whether the network model had overfitting or underfitting issues during network model evaluation.

[0112] In one specific embodiment, the computer hardware configuration used for building the Phy-SETCN model is as follows: NVIDIA RTX 3060 GPU, 16GB RAM, and 6GB VRAM. The deep learning framework PyTorch (version 2.4.0) is used for neural network training and validation, with CUDA 12.4 and cuDNN 12.5 for GPU acceleration. The programming language is Python 3.10.0. During training, the hyperparameters of the model are determined using a grid search method. The Adam optimizer, known for its good optimization performance, is used with a learning rate of 0.00015, a batch size of 256, and 100 training steps. SmoothL1Loss is used to calculate the error between the true and predicted values, and Dropout is set to 0. Specific parameter values ​​are shown in Table 2 below.

[0113] Table 2. Model hyperparameter values

[0114]

[0115] Specifically, to verify the practicality and superiority of the SETCN method proposed by improving TCN using SENet, the performance of baseline models such as TCN, LSTM, and CNN in establishing temperature-strain correlation models was compared. The specific training process of each model is as follows: Figure 11 As shown.

[0116] Figure 11The convergence process of the loss curves of each model is shown. The trends of the training loss curves of each model are similar, and they all decrease rapidly and tend to be stable in the first 25 epochs. The validation losses of SETCN and TCN decrease rapidly and tend to be stable in the first 25 epochs, with minimum values of 0.000113 and 0.000139 respectively. The validation losses of LSTM and CNN tend to be stable only at the 50th epoch, with minimum values of 0.000171 and 0.000654 respectively. This shows that the proposed method is superior to other models in terms of convergence speed and stability. Further, the mapped strain of the validation set is compared with the measured strain and shown in Figure 12 in.

[0117] From Figure 12 it can be observed that the strains mapped by SETCN, TCN and LSTM roughly coincide with the changes in the measured strain, successfully capturing the strain changes of each day. Compared with the results of TCN and LSTM, the SETCN results are closer to the measured strain at the peaks and valleys, and the mapped strain is smoother with fewer burrs. The CNN results are in good agreement with the measured strain at the first 2000 time points, but perform poorly in the subsequent mapping, with an increasing deviation from the measured strain, especially obvious at the minima near time points 3000 and 4500. To sum up, the proposed method has stronger feature capture ability, faster convergence speed and higher mapping accuracy. To quantify the differences between different methods, the specific evaluation index values and radial bar charts are shown in Table 3 and Figure 13 as shown.

[0118] Table 3 Comparison table of model accuracies

[0119]

[0120]

[0121] It can be found from Table 3 that under the two evaluation indexes of MAE and RMSE, the performance ranking of each model is CNN < LSTM < TCN < SETCN. The performance of CNN is significantly behind that of other models, indicating that CNN is less suitable for processing long-time series data. As for R 2 the results of each model do not differ much. By observing the specific values in Table 3, it can be found that although the time consumption of SETCN is slightly higher than that of TCN, compared with TCN, the MAE and RMSE of SETCN are reduced by 9.40% and 8.92% respectively. R 2The proposed method achieved a 0.62% improvement in accuracy and fit, indicating a significant improvement in both. Furthermore, the increase in runtime due to the more complex model structure is within an acceptable range. While LSTM and CNN are faster, their model accuracy is lower. Compared to LSTM, SETCN reduces MAE and RMSE by 17.70% and 18.21%, respectively, and R... 2 This represents a 1.46% improvement. Compared to CNN, SETCN reduced MAE and RMSE by 53.98% and 57.33%, respectively, and R... 2 It increased by 14.47%.

[0122] A residual-driven model, Phy-SETCN, embedding a physical mechanism, was constructed by subtracting the theoretical stress values ​​calculated from physical formulas from the measured stress values. The loss curves of Phy-SETCN and SETCN are compared. Observation shows that the training and validation losses of Phy-SETCN decrease faster and more smoothly than those of SETCN. The minimum training and validation losses of Phy-SETCN are 0.0000740 and 0.0000745, respectively, representing reductions of 5.88% and 34.66% compared to SETCN. Further comparisons of the mapped strain and measured strain on the validation set for both methods are presented. Figure 14 The specific indicators are shown in Table 4.

[0123] Table 4 Model Performance Comparison

[0124] MAE RMSE <![CDATA[R 2 ]]> Time SETCN 1.590 1.981 0.973 66.288 Phy-SETCN 1.220 1.537 0.984 67.698 Improvement -23.27％ -22.41％ 1.13％ 2.13％

[0125] Analysis shows that Phy-SETCN outperforms SETCN on the validation set, especially in the area highlighted by the red dashed line. The stress mapped by Phy-SETCN is closer to the measured stress, and the curve trend matches better. The index values ​​in Table 4 also confirm the superior performance of Phy-SETCN, although R... 2 While only improving by 1.13%, its accuracy is significantly improved. Compared to SETCN, Phy-SETCN reduces MAE and RMSE by 23.27% and 22.41% respectively, with only a 2.13% increase in runtime.

[0126] To reflect the model's performance when faced with entirely new and unseen data, a test set was selected, excluding the training and validation sets, to evaluate the model's generalization ability. Using data from a week in March 2024 as an example, the mapped strain shows a good match with the measured strain trend, exhibiting a clear diurnal cycle. Furthermore, as shown in the red dashed box, the model effectively captures stress changes caused by atypical temperature variations. The MAE on the Phy-SETCN test set is 5.426, the RMSE is 6.261, and the R...2 The value is 0.926. The above analysis confirms that Phy-SETCN has good generalization ability, and the trained model can be applied to completely new and unseen data, avoiding the computational resource and time loss caused by repeated training.

[0127] Specifically, a temperature-response intelligent mapping method for long-span stiffened arch bridges embeds physical mechanisms into a neural network to establish a temperature-strain mapping model based on Phy-SETCN. This method considers cross-sectional temperature, longitudinal non-uniformity of temperature distribution, and physical information, effectively improving the model's convergence stability, generalization ability, and mapping accuracy, providing a solution. The main components include:

[0128] (1) The temperature distribution exhibits obvious non-uniformity. The correlation between temperature and strain within the same cross section is more significant, while there is a significant time lag effect between temperature and strain at different cross sections. This indicates that the relationship between temperature and strain cannot be simply described by a linear model, and the measured temperature and strain data both show significant non-Gaussian characteristics.

[0129] (2) SENet effectively characterizes the dependencies between feature channels and recalibrates the channel weights to emphasize key features and suppress useless features. This method significantly enhances the feature representation capability of TCN and solves the problem that TCN's receptive field focuses on early information while ignoring recent information. Experimental results show that SENet outperforms TCN in MAE, RMSE, and R... 2 In terms of metrics, SETCN outperforms the comparison model. Compared to TCN, SETCN reduces MAE and RMSE by 9.40% and 8.92%, respectively.

[0130] (3) By using the theoretical temperature-induced response calculation formula of the stiffened frame arch bridge of the present invention, the physical mechanism is embedded into a neural network to build a residual-driven mapping model, which further improves the mapping accuracy and interpretability of the model. Compared with SETCN, Phy-SETCN reduces MAE and RMSE by 23.27% and 22.41% respectively, while increasing the running time by only 2.13%.

[0131] (4) This invention not only demonstrates excellent mapping accuracy, but also exhibits good generalization ability. The trained model can be effectively applied to completely new and unseen data, avoiding the computational resource and time losses caused by repeated training.

[0132] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0133] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A temperature-response intelligent mapping method for long-span stiffened arch bridges, characterized in that, include: S100: Acquire multi-channel temperature data and corresponding strain data at specified measuring points of a bridge health monitoring system, including: Using the measured value of the start time of the period of interest as the benchmark, the strain increment and temperature increment data within the period are calculated. The temperature increment is used as input. Based on the calculation formula of the additional internal force under the action of temperature on the arch bridge, the theoretical strain under the action of temperature on the scaled model arch is derived. The residual between the theoretical strain and the measured strain is used as the model output data. S200: Preprocess the multi-channel temperature data and the corresponding strain data at the measuring points to generate a dataset, and divide it into a training set and a validation set; S300: Construct and train the Phy-SETCN model based on the training set, and verify the generalization ability of the Phy-SETCN model using the validation set. MAE , RMSE , R 2 Evaluation of strain data obtained by mapping the Phy-SETCN model; S400: Real-time detection and early warning of abnormal response of bridge temperature-induced strain using the Phy-SETCN model; The Phy-SETCN model integrates the SENet and TCN modules and introduces physical information to construct residuals as the driving force. The specific operation process is as follows: S310: Use the permute() function to adjust the input data; S320: After passing through the adaptive average pooling layer, the spatial dimension of the feature map is compressed and the shape is adjusted to match the input of the fully connected layer; S330: Define two fully connected layers as activation operations, use the Sigmoid activation function, output the importance coefficient of each channel, and adjust the shape to match the shape of the original feature map; S340: The output dimension of the SENet module is adjusted using the transpose(·) function to facilitate its input to the TCN module; S350: The TCN module performs convolution operations on the input time series and uses residual connections to propagate feature information across layers, capturing nonlinear data features; S360: Use the transpose() function to reshape the output shape of the TCN module for forwarding to the linear layer; S370: Select the output of the last time step as the final prediction result.

2. The intelligent temperature-response mapping method for a long-span stiffened arch bridge according to claim 1, characterized in that, S200 includes: The input and output datasets are constructed using a sliding window. The input and output data are subjected to min-max normalization to eliminate differences in dimensions. The dataset is split into a training set and a validation set.

Images