A model training method, a steel bridge temperature field prediction method and device

Abstract

The application provides a model training method, a steel bridge temperature field prediction method and device. The steel bridge temperature field prediction method comprises the following steps: acquiring temperature, stress, environmental parameters and traffic load data of each target monitoring point of a steel bridge to be predicted; inputting the temperature, stress, environmental parameters and traffic load parameters into a target service steel bridge structure temperature field prediction model; and outputting temperature field prediction values of different monitoring points and different times of the steel bridge to be predicted. The service steel bridge structure temperature field prediction model is constructed based on a dynamic graph convolution and a time convolution network, and comprises a dynamic graph convolution module, a feature fusion module LSTM and a time sequence feature extraction module. Through double-driven dynamic graphs, mechanism constraints and bidirectional verification, the application strengthens the adaptability of complex working conditions, can improve the steel bridge temperature field prediction accuracy under complex working conditions, provides a reliable basis for safety evaluation, and guarantees the safe operation of the steel bridge.

Description

Technical Field

[0001] This application relates to the field of bridge structural health monitoring and intelligent prediction technology, and in particular to a model training method, a method and device for predicting the temperature field of steel bridges. Background Technology

[0002] As transportation infrastructure develops towards heavier loads and larger sizes, in-service steel bridges endure complex stress-temperature coupling effects during long-term operation. The alternating structural stresses caused by heavy traffic and the sudden temperature changes due to extreme weather conditions result in a highly nonlinear and spatiotemporally dynamic temperature field, directly impacting the accuracy of secondary temperature stress calculations and the reliability of structural safety assessments. Traditional methods for predicting the temperature field of steel bridges (such as empirical formulas and static finite element simulations) struggle to characterize the dynamic impact of stress and deformation on the heat transfer path, leading to temperature gradient calculation errors exceeding 15%, thus posing a hidden danger for structural fatigue assessments.

[0003] Existing technologies have proposed intelligent prediction methods based on GCN-LSTM, which improve upon traditional steel bridge temperature field prediction methods, but still have significant limitations: First, the Graph Convolutional Network (GCN) uses a static topology, which cannot adjust node associations in real time according to the deformation of stress concentration areas (such as welds and supports), making it difficult to capture the "stress-heat conduction" coupling law; Second, LSTM suffers from gradient vanishing problems in multi-scale time series modeling (such as the superposition of daily periodic fluctuations and seasonal trends), and prediction deviations exceed 8°C under extreme conditions (such as cold waves + braking loads); Third, it lacks constraints on heat conduction mechanisms, and purely data-driven models are prone to deviating from physical laws (such as predicting temperature fields that violate heat conservation); Fourth, the verification mechanism relies only on temperature errors and does not undergo reverse verification through measured stress values, leading to the contradiction of "temperature prediction being qualified but stress assessment failing".

[0004] Furthermore, the superposition of heavy traffic and extreme weather (such as high summer temperatures combined with dense traffic of overloaded trucks) exacerbates the stress-temperature coupling effect, causing existing methods to drop in prediction accuracy by more than 40% under such conditions. Therefore, designing a temperature field prediction method that integrates stress-temperature dual-driven dynamic graph convolution, heat conduction mechanism constraints, and bidirectional coupling verification is of significant engineering value for overcoming the bottleneck of temperature field modeling under complex conditions and improving the reliability of steel bridge safety assessment. Summary of the Invention

[0005] In view of the above problems, this application provides a model training method, a steel bridge temperature field prediction method and device. This method achieves accurate modeling and prediction of steel bridges by deeply mining and learning massive amounts of measured temperature data and combining environmental factor analysis, so as to effectively ensure the safety and stability of steel bridge structures and overcome or at least partially solve the above problems.

[0006] In a first aspect, embodiments of this application provide a model training method, including:

[0007] The hyperparameter combination of the temperature field prediction model for the in-service steel bridge structure was adjusted and optimized using the Bayesian algorithm to obtain the temperature field prediction model for the in-service steel bridge structure.

[0008] Evaluate whether the temperature field prediction model of the in-service steel bridge structure meets the engineering performance indicators. If not, return to the previous step to readjust the hyperparameters of the temperature field prediction model of the in-service steel bridge structure until the engineering performance indicators are met, and output the trained temperature field prediction model of the in-service steel bridge structure.

[0009] Based on the improved coefficient of determination, the trained temperature field prediction model for in-service steel bridge structures is tested. The tested and passed temperature field prediction model for in-service steel bridge structures is used as the target temperature field prediction model for in-service steel bridge structures for temperature field prediction.

[0010] Optionally, the temperature field prediction model for the in-service steel bridge structure is constructed based on a dynamic graph convolutional and temporal convolutional network, including a dynamic graph convolutional module, a feature fusion module, and a temporal feature extraction module;

[0011] The dynamic graph convolution module is used to extract the dynamic spatial correlation features between various monitoring points in the temperature field of the in-service steel bridge structure;

[0012] The feature fusion module is used to fuse the dynamic spatial correlation features output by the dynamic graph convolution module with the acquired multi-source data using an LSTM network; the multi-source data includes temperature, stress, environmental parameters, and traffic load data.

[0013] The temporal feature extraction module is used to extract the temporal features of the fused feature vector using a temporal convolutional network to obtain the predicted temperature value of the steel bridge.

[0014] Optionally, the step of using a Bayesian algorithm to adjust and optimize the hyperparameter combination of the in-service steel bridge structure temperature field prediction model to obtain a trained in-service steel bridge structure temperature field prediction model includes:

[0015] The preprocessed temperature, stress, environmental parameters, and traffic load data are divided into training, testing, and validation sets according to a set ratio.

[0016] The training set data is input into the temperature field prediction model of the in-service steel bridge structure to learn the coupling law of multiple factors and temperature field, fit the model hyperparameters, and train the temperature field prediction model of the in-service steel bridge structure.

[0017] The validation set data is input into the trained temperature field prediction model of the in-service steel bridge structure, and the predicted temperature value is output.

[0018] Based on the obtained temperature prediction values ​​and the real-time temperatures of the corresponding monitoring points in the validation set, the basic mean square error loss and heat conduction loss terms are calculated to determine the total model loss.

[0019] The confidence of the surrogate model in Bayesian optimization is updated based on the total model loss.

[0020] By using the expectation enhancement criterion, select the next set of hyperparameter combinations in the hyperparameter space to be evaluated until the preset parameter convergence condition is met, and determine the optimal hyperparameter combination.

[0021] The optimal hyperparameter combination after the last optimization is used as the parameters of the temperature field prediction model for the in-service steel bridge structure, and the trained temperature field prediction model for the in-service steel bridge structure is output.

[0022] Optionally, the step of inputting the validation set data into the trained temperature field prediction model for the in-service steel bridge structure and outputting the predicted temperature value includes:

[0023] Each monitoring point of the steel bridge is used as a spatial node of a dynamic graph convolutional network, and the dynamic spatial correlation features of the temperature field are extracted through the dynamic graph convolutional network.

[0024] Temperature, stress, meteorological, and traffic load data, along with dynamic spatial correlation features output from a dynamic graph convolutional network, are used as input layer data for the LSTM model, which outputs a fused feature vector.

[0025] The fused feature vector is used as input to a temporal convolutional network. The temporal convolutional network is then used to extract the temporal features of the fused feature vector to obtain the predicted temperature of the steel bridge for the current time period.

[0026] Optionally, the step of calculating the basic mean square error loss and heat conduction loss based on the obtained temperature prediction value and the real-time temperature of the corresponding monitoring points in the validation set, and determining the total model loss, includes:

[0027] Based on the simulation results of finite element modeling of the steel bridge, the main prediction region and the secondary prediction region are divided to determine the temperature prediction region of the steel bridge.

[0028] Collect the real-time temperature of the corresponding monitoring point in the temperature prediction zone and calculate the basic mean square error loss;

[0029] By combining the heat source term of traffic load with the theory of heat conduction, the heat conduction loss is calculated;

[0030] The total model loss is obtained based on the basic mean square error loss and the heat conduction loss; the formula for calculating the total model loss is as follows:

[0031] ;

[0032] in, This represents the total loss of the model; As the weight for heat conduction loss, Based on the mean square error loss, This is due to heat conduction loss.

[0033] Optionally, evaluate whether the temperature field prediction model for the in-service steel bridge structure meets the engineering performance indicators. If not, return to the previous step to readjust the hyperparameters of the temperature field prediction model for the in-service steel bridge structure until the engineering performance indicators are met, and output the trained temperature field prediction model for the in-service steel bridge structure, including:

[0034] The difference between the predicted temperature value and the real-time temperature value is taken as the temperature error. It is determined whether the temperature error is less than or equal to 2℃. If so, proceed to the next step; otherwise, readjust the hyperparameters of the temperature field prediction model of the in-service steel bridge structure.

[0035] Calculate the stress deviation between the predicted temperature secondary stress and the actual temperature secondary stress. If the stress deviation exceeds 3%, readjust the hyperparameters of the temperature field prediction model for the in-service steel bridge structure. Otherwise, output the temperature field prediction model for the in-service steel bridge structure that meets the engineering performance indicators.

[0036] Optionally, based on the improved coefficient of determination, the trained temperature field prediction model for in-service steel bridge structures is tested. The tested and successful temperature field prediction model for in-service steel bridge structures is used as the target temperature field prediction model for in-service steel bridge structures, for predicting the temperature field of in-service steel bridge structures, including:

[0037] The test set data is input into the trained temperature field prediction model of the in-service steel bridge structure to obtain the corresponding temperature prediction value.

[0038] The improved determination coefficient R² is calculated based on the predicted temperature value and the real-time temperature value; the formula for calculating the improved determination coefficient R² is as follows:

[0039] ;

[0040] in, Assuming the weighting of traffic load impact, Let be the vehicle density at time t, x(t) be the real-time temperature value, and y(t) be the predicted temperature value. The average of the real-time temperature values ​​is given, and n represents the total time.

[0041] If R² ≥ 0.9 and stress deviation ≤ 3% At that time, among them, If the yield strength of the steel is given, the temperature field prediction model for the in-service steel bridge structure passes the test, and the model at this time is used as the target temperature field prediction model for the in-service steel bridge structure.

[0042] Secondly, embodiments of this application provide a method for predicting the temperature field of a steel bridge, including:

[0043] Obtain temperature, stress, environmental parameters, and traffic load data at various monitoring points of the steel bridge to be predicted;

[0044] Temperature, stress, environmental parameters, and traffic load data are input into the temperature field prediction model of the target in-service steel bridge structure; the temperature field prediction model of the target in-service steel bridge structure is obtained by one of the model training methods described above;

[0045] Output the predicted temperature field values ​​of the steel bridge at different monitoring points and at different times.

[0046] Thirdly, embodiments of this application provide a steel bridge temperature field prediction device, the device comprising:

[0047] The data acquisition module is used to acquire temperature, stress, environmental parameters, and traffic load data at various monitoring points of the steel bridge to be predicted.

[0048] The prediction module is used to input temperature, stress, environmental parameters, and traffic load data into the temperature field prediction model of the target service steel bridge structure; the temperature field prediction model of the target service steel bridge structure is obtained as described above.

[0049] The output module is used to output the predicted temperature field values ​​of the steel bridge at different monitoring points and at different times.

[0050] The specific beneficial effects are as follows:

[0051] First, this invention constructs a stress-temperature dual-driven dynamic graph convolution (DG), which can adapt to changes in heat conduction paths under structural stress deformation in real time, effectively enhancing the accuracy of spatial feature extraction and overcoming the problem of insufficient adaptability of static graph convolutional networks (GCN) when dealing with structural deformation.

[0052] Second, this invention employs a temporal convolutional network (TCN). By leveraging the design of dilated convolution and residual connections, it significantly improves computational efficiency in long-term (annual scale) temperature field prediction, while avoiding the gradient vanishing problem that is prone to occur in traditional long short-term memory networks (LSTM), thus ensuring the stability of long-term prediction.

[0053] Third, the heat conduction mechanism constraint introduced in this invention can effectively constrain the deviation between the predicted temperature field and the actual physical laws, make up for the "physical deviation" defect of the pure data-driven model, and make the model prediction more in line with the objective laws of heat conduction.

[0054] In summary, the method proposed in this invention solves the problems of poor adaptability of static models, insufficient accuracy under extreme conditions, and weak physical interpretability in traditional steel bridge temperature field prediction by constructing an integrated framework of "dynamic spatial modeling - multi-scale temporal capture - mechanism constraint verification". It is applicable to scenarios with significant stress-temperature coupling, such as heavy-load railway bridges and long-span steel bridges in extreme climate zones. It can directly support structural fatigue assessment and preventative maintenance decisions, reduce safety risks caused by temperature prediction deviations, and promote the upgrade of bridge health monitoring from "data fitting" to "mechanism-data dual-driven" approaches. It has significant engineering application value and promising prospects for technology promotion. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of the present invention;

[0057] Figure 2 This is a schematic diagram of the typical temperature measurement points and vehicle load arrangement of the steel bridge according to the present invention;

[0058] Figure 3 This is an architecture diagram of the DG-TCN model of the present invention;

[0059] Figure 4 This is a schematic diagram of the comparison arrangement for predicting the temperature field of in-service steel bridge structures based on DG-TCN according to the present invention.

[0060] Figure 5 This is the overall logic block diagram of the method proposed in this invention;

[0061] Figure 6 These are the prediction results of the three models compared in the experimental case; where (a) is the prediction result of the DG-TCN model; and (b) is the prediction result of the LSTM and CNN models.

[0062] Among them, 1-bridge deck pavement, 2-top surface of bridge deck, 3-corner weld, 4-U-shaped longitudinal rib, 5-diaphragm, 6-temperature sensor, 7-bridge load detector. Detailed Implementation

[0063] Exemplary embodiments of this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0064] Combination Figure 1 This application provides a model training method, which includes the following steps:

[0065] Step 1: Obtain temperature, stress, environmental parameters, and traffic load data at various monitoring points of the steel bridge structure. Based on the preprocessed data, perform a refined finite element modeling of the target steel bridge to simulate its temperature-stress coupling distribution under typical working conditions (such as "heavy load + high temperature" and "alternating daily average temperature"). Based on the simulation results, identify stress-temperature dual-sensitive areas on the steel bridge (such as stress concentration areas near welds and temperature difference areas at the junction of sunlight and shadow).

[0066] Specifically, in combination Figure 2 The steel bridge structure is the top surface 2 of the bridge deck pavement 1. The steel plate below the bridge deck pavement 1 protects the steel components below the bridge deck and can regulate the local heat conduction environment. The bridge deck pavement 1 is connected to the main body of the steel bridge through fillet welds 3. This application selects key parts of the steel bridge as monitoring points, including: the heat-affected zone of the fillet welds below the bridge deck pavement (this area is more sensitive to temperature and stress response due to welding process and stress concentration characteristics), the stress concentration zone of the web of the steel box girder (this is the key area of ​​stress on the steel bridge structure, where the coupling effect of stress and temperature is significant), and the connection area of ​​the U-shaped longitudinal ribs 4 (the longitudinal rib connection is related to the overall stress state and heat distribution uniformity of the steel bridge deck). High-precision temperature sensors 6 (accuracy ±0.2℃, sampling frequency 5min) and strain sensors (accuracy ±1με, sampling frequency 10Hz) are embedded at each monitoring point. A meteorological station is installed simultaneously, and traffic monitoring equipment is deployed. The sensor spacing is ≤3m to cover the stress-temperature coupling sensitive area. Temperature and stress data at each monitoring point are collected using temperature sensor 6 and strain sensor, and environmental parameters (ambient temperature Ta, solar radiation intensity I, wind speed v, relative humidity RH) are collected using meteorological station, and traffic monitoring equipment (i.e. bridge load detector 7) is used to collect bridge deck traffic load data (vehicle density, average load, and traffic speed).

[0067] Based on the steel bridge, a refined finite element model was built to calculate the temperature-stress coupling distribution under typical working conditions (heavy load + high temperature, cold wave + braking load, stress alternation period), thereby identifying stress-temperature dual-sensitive areas (such as stress concentration areas near welds, temperature difference areas at the junction of sunlight and shadow). According to the stress change rate (>5MPa / h) and temperature gradient (>8℃ / m), the main prediction area (such as the connection between the web and the transverse diaphragm 5) and the secondary prediction area (such as the non-stressed base material area) were delineated in the stress-temperature dual-sensitive areas. The node density of the main prediction area was 3 times that of the secondary prediction area.

[0068] After obtaining the above multi-dimensional raw data, preprocessing operations such as outlier removal, stress moving average smoothing, and data normalization are performed. Then, based on monitoring data such as temperature and stress, a matrix form is obtained, ultimately forming a node feature matrix containing temperature and stress information. .

[0069] Furthermore, all the collected raw data are preprocessed (outlier removal, stress data smoothing, and normalization).

[0070] The formula for smoothing stress data is as follows: ;

[0071] In the formula, The shading efficiency is determined experimentally, with a value ranging from 0.1 to 0.3, and I' is the corrected solar radiation intensity (W / m²). The solar radiation intensity (W / m²) before correction. The stress value at time t is the smoothed value. N is the window length for the moving average (here N = 5, meaning that the stress values ​​at 5 consecutive moments are averaged). For the summation index (values ​​from 0 to (N-1), The time sampling interval for stress data. for The raw stress monitoring value at that moment (unsmoothed stress data). Vehicle density.

[0072] By using moving average smoothing, interference from solar radiation fluctuations and equipment acquisition errors in stress monitoring can be filtered out, while retaining the stress variation trend driven by "corrected solar radiation + traffic load + other factors." This provides a cleaner and more regular stress characteristic input for subsequent temperature field prediction models. By correcting for solar radiation, the contribution of heat input to force can be accurately quantified. Stress smoothing can filter out high-frequency noise from sensors and retain effective trends, which is beneficial for extracting stress characteristics strongly correlated with solar radiation and traffic load, supporting the accurate prediction of the temperature field by subsequent models.

[0073] The thermal effect of the input traffic load data is preprocessed by quantifying the bridge deck equivalent temperature increment. The formula for calculating the bridge deck equivalent temperature increment (caused by traffic load) is as follows:

[0074] ;

[0075] In the formula, The energy conversion factor is 0.05. This is the total load energy term of the bridge deck related to the average vehicle speed (integrating the combined effect of the total bridge deck load G and the average vehicle speed v on the thermal effect). A is the bridge deck area (m²). Where c is the density of steel (kg / m³), and c is the specific heat capacity. .

[0076] The above formula can convert the "mechanical parameters of traffic load (total load G, average vehicle speed v)" into "thermal parameters (equivalent temperature increment)" through quantitative preprocessing of the thermal effect of traffic load. This approach maps traffic load characteristics from the "mechanical domain" to temperature characteristics in the "thermal domain," enabling subsequent temperature field prediction models (DG-TCN-LSTM models) to directly incorporate the contribution of traffic load to temperature, thus improving the model's prediction accuracy for complex coupled scenarios. The equivalent temperature increment of the bridge deck caused by traffic load alone was quantified.

[0077] Step 2: Based on the preprocessed data, establish a temperature field prediction model for in-service steel bridge structures based on DG-TCN;

[0078] Combination Figure 3 The temperature field prediction model for the in-service steel bridge structure is constructed based on Dynamic Graph Convolution (DG) and Temporal Convolutional Network (TCN), and includes a Dynamic Graph Convolution module, a Feature Fusion Module (LSTM), and a Temporal Feature Extraction module; wherein:

[0079] The dynamic graph convolution module is used to extract the dynamic spatial correlation features between various monitoring points in the temperature field of the in-service steel bridge structure;

[0080] The feature fusion module is used to fuse the spatial features output by the dynamic graph convolution module with the acquired multi-source data using an LSTM network; the multi-source data includes temperature, stress, meteorological, and traffic load data.

[0081] The temporal feature extraction module is used to extract the temporal features of the fused feature vector using a temporal convolutional network to obtain the predicted temperature value of the steel bridge.

[0082] Step 3: Use the Bayesian algorithm to adjust and optimize the hyperparameter combination of the temperature field prediction model of the in-service steel bridge structure to obtain the trained temperature field prediction model of the in-service steel bridge structure.

[0083] Optionally, step 3 may include the following sub-steps:

[0084] Step 3.1: Divide the preprocessed temperature, stress, environmental parameters, and traffic load data into training, testing, and validation sets according to the specified proportions;

[0085] Step 3.2: Input the training set data into the temperature field prediction model of the in-service steel bridge structure, learn the coupling law of multi-factor-temperature field, fit the model hyperparameters, and train the temperature field prediction model of the in-service steel bridge structure.

[0086] The training set data includes multi-source data such as corrected solar radiation intensity, traffic load, stress, and environmental parameters, along with corresponding temperature labels; the hyperparameters include the weight matrix of dynamic graph convolution. Examples of parameters include the kernel size and dilation coefficient of TCN, the hidden layer dimension of LSTM, and the gating parameters of LSTM.

[0087] Step 3.3: Input the validation set data into the trained temperature field prediction model of the in-service steel bridge structure and output the predicted temperature value; the validation set data includes temperature, stress, environmental parameters, traffic load data, etc.

[0088] Furthermore, the construction steps in step 3.3 are as follows:

[0089] Step 3.3.1: Use the monitoring points of the steel bridge as spatial nodes of the dynamic graph convolutional network, and extract the dynamic spatial correlation features of the temperature field through the dynamic graph convolutional network;

[0090] Step 3.3.2: Use temperature, stress, environmental parameters, traffic load data and spatial features output by the dynamic graph convolutional network as input layer data for the LSTM model, determine the structure of the hidden and output layers of the LSTM model, and output the fused feature vector.

[0091] Step 3.3.3: Use the fused feature vector as input to the temporal convolutional network, and use the temporal convolutional network to extract the temporal features of the fused feature vector to obtain the temperature prediction value of the steel bridge in the current time period.

[0092] Furthermore, the specific steps of step 3.3.1 include:

[0093] Step 3.3.1.1: Determine the stress difference between nodes based on the stress and temperature between two different monitoring points. and temperature gradient ;

[0094] Specifically, the dynamic spatial node set composed of monitoring points is defined as ,in, Let represent the nth node input at time t; the stress difference between nodes i and j can be obtained from the absolute value of the difference between the input stress and temperature of any two nodes i and j. and temperature gradient ;

[0095] Step 3.3.1.2: Based on the stress difference between nodes Temperature gradient Construct a time-varying adjacency matrix using distance and heat conduction weights;

[0096] First, determine the heat conduction weights between nodes i and j. for:

[0097] ;

[0098] in, The thermal conductivity coefficient between two nodes reflects the steel's ability to transfer heat. Let be the straight-line distance between node i and node j; Let be the stress at node i at time t. Let be the stress at node j at time t; Let be the temperature of node i at time t. Let J be the temperature of node j at time t. As the stress weight ratio, =0.6 (higher than temperature weight) (to highlight the effect of stress on heat conduction). The stress influence coefficient is... =5; The yield strength of the steel; This is the temperature influence coefficient. =3; The maximum temperature difference threshold, =50°; This is a traffic load correction factor. =0.03; Initial vehicle density;

[0099] Secondly, a time-varying adjacency matrix is ​​constructed based on heat conduction weights: ;

[0100] Based on the time-varying adjacency matrix, the dynamic spatial correlation characteristics of the temperature field output by the dynamic graph convolution module are as follows:

[0101] ;

[0102] in, For dynamic spatial correlation features, It is a time-varying adjacency matrix. This is a node feature matrix containing temperature and stress information. This is the DG weight matrix. For bias terms, This is an activation function used to enhance nonlinear expressions.

[0103] Furthermore, in the LSTM module of step 3.3.2, the input gate Forgotten Gate Output gate and cell state The update formula is as follows:

[0104] Input Gate: ;

[0105] Forgotten Gate: ;

[0106] Cell state candidate values: ;

[0107] Cell status update: ;

[0108] Output gate: ;

[0109] Hidden state: ;

[0110] In the formula, [ ] represents the fusion vector, x t For temperature data, , , , as well as , , , Both are weight matrices. , , , All are bias terms, σ is the sigmoid function, and ⊙ represents element-wise multiplication. The hidden state at time t-1 Let be the hidden state at time t. Let be the candidate values ​​for the cell state at time t. Let t represent the cell state at time t. The cell state at time t-1. is the output of the output gate, and tanh is the hyperbolic tangent function.

[0111] Furthermore, in step 3.3.3, the Temporal Feature Extraction (TCN) module uses a 3-layer dilated convolution to extract temporal features, the third of which... The output features of the layer are:

[0112] ;

[0113] in, The number of layers in the dilated convolution. The fused vector output by the LSTM module (containing temperature data, Dynamic Graph (DG) output features, and vehicle density, i.e., the fused vector) =[ ], The kernel size for dilated convolution. =7; The coefficient of thermal expansion is 1 / 3. =2 l-1 (Doubling the layer by layer to capture multi-scale time series patterns on a daily / weekly / monthly scale); These are residual connections used to prevent gradient vanishing in deep networks.

[0114] Step 3.4: Based on the obtained temperature prediction values ​​and the real-time temperatures of the corresponding monitoring points in the validation set, calculate the basic mean square error loss and heat conduction loss, and determine the total model loss;

[0115] Optionally, step 3.4 may include the following sub-steps:

[0116] Step 3.4.1: Based on the finite element model of the steel bridge in Step 1, refine the main prediction region and the secondary prediction region according to the simulation results, and determine the temperature prediction region of the steel bridge;

[0117] The primary and secondary prediction regions are further refined using numerical simulations (finite element numerical simulation results). Based on the primary and secondary prediction regions defined in the previous step, numerical simulations are used to refine the region based on stress-temperature coupling and load response, narrowing the scope to obtain a more precise primary prediction region (a key local area with significant stress-temperature coupling and extremely high requirements for temperature field prediction accuracy) and a refined secondary prediction region (areas with relatively stable coupling but still requiring monitoring). These refined primary and secondary prediction regions are then considered as the temperature prediction region.

[0118] In finite element modeling, a heat conduction mechanism constraint term is introduced to ensure that the prediction conforms to the physical law of "heat diffusion + external heat source". The heat conduction mechanism constraint term is as follows:

[0119] ;

[0120] in, Let c be the density of the steel, c be the specific heat capacity of the steel, and k be the thermal conductivity of the steel. This is a heat generation term induced by the stress field. This refers to the heat input corresponding to environmental loads (solar radiation, wind speed, etc.). This refers to the heat input item corresponding to traffic loads (vehicle friction, braking, etc.). Expressing the request gradient, This represents the average temperature. This represents the partial derivative of the average temperature with respect to time t;

[0121] By discretizing and solving the above equations using the finite difference method or the finite element method, the temperature changes at different locations can be obtained.

[0122] Step 3.4.2: Collect the real-time temperature of the corresponding monitoring point in the temperature prediction zone and calculate the basic mean square error loss;

[0123] The formula for calculating the basic mean square error loss is as follows:

[0124] ;

[0125] in, The total number of samples in the validation set. Mean squared error loss, Let i be the predicted temperature value for the i-th sample. The true temperature value of the i-th sample;

[0126] Step 3.4.3: Calculate heat transfer loss by combining the traffic load heat source term with heat conduction theory;

[0127] The formula for calculating the heat conduction loss is:

[0128] ;

[0129] in, For heat conduction loss, The stress-heat conduction correlation weights, For the predicted temperature secondary stress at the j-th monitoring point, Let M be the actual temperature secondary stress at the j-th monitoring point, and M represent the total number of steel bridge monitoring points involved in the calculation of heat conduction loss.

[0130] Step 3.4.4: Based on the basic mean square error loss and heat conduction loss, obtain the total model loss;

[0131] By introducing a heat conduction mechanism constraint, we can ensure that the prediction conforms to the physical laws of "heat diffusion + external heat source". The hybrid loss function incorporating the heat conduction mechanism constraint is:

[0132] ;

[0133] in, This represents the total loss of the model; As the weight for heat conduction loss, Based on the mean square error loss, This is due to heat conduction loss;

[0134] Step 3.5: Update the surrogate model confidence in Bayesian optimization based on the total model loss;

[0135] Bayes' theorem uses the prior distribution of unknown parameters combined with the likelihood of observed data to form the posterior distribution of the parameters. Its core lies in calculating the posterior probability by updating the prior probability. The Expected Utility Improvement criterion is a Bayesian decision theory criterion used for optimizing decisions. This criterion selects the optimal decision by maximizing the expected utility improvement and is often used to handle uncertain decision-making problems.

[0136] The mathematical expression of Bayes' theorem is:

[0137] ;

[0138] In the formula: This represents a dataset containing vehicle load observations. Represents the posterior probability. This represents the probability of the likelihood distribution. Prior probability represents an assumption about the distribution of an unknown objective function. This represents the marginal likelihood distribution.

[0139] The surrogate model is an auxiliary model used in Bayesian optimization to approximate the complex mapping from hyperparameter combinations to model loss, such as a Gaussian process; based on the validation set loss of the new hyperparameter combination, it is calculated using Bayes' theorem. ( For loss, For hyperparameters, To obtain the posterior distribution (for validation set loss), the prior distribution and the likelihood of new samples are fused to obtain the posterior distribution to correct the confidence. The validation set loss is used as feedback on the true performance to guide the optimization of the surrogate model.

[0140] Step 3.6: Using the expectation enhancement criterion, select the next set of hyperparameter combinations in the hyperparameter space to be evaluated until the preset parameter convergence condition is met, and determine the optimal hyperparameter combination;

[0141] Specifically, the expected improvement criterion refers to the "selection of the most promising hyperparameter" criterion derived from Bayes' theorem, and its formula is as follows:

[0142] ;

[0143] in, For hyperparameters The corresponding expected improvement value, Expressing expectations, This represents the current minimum validation set loss. Hyperparameters predicted for surrogate models Corresponding loss;

[0144] The parameter convergence condition refers to the stopping condition of the Bayesian optimization iteration, such as the decrease in validation set loss being less than 10 after k consecutive iterations. -4 ;

[0145] In this embodiment, the reason for using the validation set for hyperparameter optimization is that using the training set to optimize hyperparameters is prone to overfitting, while the validation set is an independent "intermediate evaluation set" that can reflect the model's adaptability to unseen data and ensure generalization ability. Therefore, the training set is used to fit the model parameters (such as DG weights, LSTM gating parameters, and TCN convolution parameters).

[0146] Step 3.7: Using the optimal hyperparameter combination after the last optimization as the model parameters, output the trained temperature field prediction model of the service steel bridge structure;

[0147] Step 4: Evaluate whether the temperature field prediction model of the in-service steel bridge structure meets the engineering performance indicators. If not, return to step 3 to readjust the hyperparameters of the model until the engineering performance indicators are met, and output the trained temperature field prediction model of the in-service steel bridge structure.

[0148] Optionally, step 4 may include the following sub-steps:

[0149] Step 4.1: Take the difference between the predicted temperature and the real-time temperature as the temperature error, and determine whether the temperature error is within the specified error range; if yes, proceed to step 4.2; otherwise, return to step 3 to readjust the model's hyperparameters.

[0150] Step 4.2: Calculate the stress deviation between the predicted temperature secondary stress and the measured temperature secondary stress. If the stress deviation exceeds 3%, return to step 3 to readjust the hyperparameters of the model. Otherwise, output the temperature field prediction model of the service steel bridge structure that meets the engineering performance indicators.

[0151] Specifically, when the temperature error is within the specified range (≤2℃), the process proceeds to the next step (stress error verification stage). Based on the thermo-mechanical coupling relationship between temperature and stress, the process continues to determine whether the stress prediction error in the corresponding area meets the standard. Otherwise, when the temperature error exceeds the specified range (>2℃), the model parameter correction mechanism is triggered, and the process returns to step 3 to readjust the model hyperparameters. The model is then retrained based on the adjusted hyperparameters until the temperature prediction accuracy meets the requirements. Finally, the temperature field prediction model of the service steel bridge structure after the last parameter adjustment is output.

[0152] The two thresholds for error verification are:

[0153] ;

[0154] in, To predict secondary temperature stress, For the measured temperature secondary stress, To predict temperature, This is the measured temperature.

[0155] By setting dual thresholds, consistency between temperature prediction and stress response can be ensured, thereby improving reliability.

[0156] Temperature secondary stress refers to the additional internal forces that arise when a structure is subjected to forced flexural or axial expansion and contraction under the influence of ambient temperature. These forces, initially constrained by redundant constraints, generate constraint forces that cause additional internal forces in the structure. These additional internal forces are called temperature secondary internal forces, and the resulting stresses are called temperature secondary stresses.

[0157] The formula for calculating secondary temperature stress (thermoelastic theory) is as follows:

[0158] ;

[0159] Where E=210GPa is the elastic modulus of steel (GPa). =12e-6 is the linear expansion coefficient (1 / ℃). To predict temperature Compared with reference temperature The difference (°C). The value is Poisson's ratio (taken as 0.3).

[0160] Step 5: Based on the improved coefficient of determination, test the trained temperature field prediction model of the in-service steel bridge structure, and use the model that passes the test as the target temperature field prediction model of the in-service steel bridge structure for temperature field prediction of the in-service steel bridge structure.

[0161] After the hyperparameters are optimized, the model's final generalization performance is evaluated on the test set to ensure that the model ultimately meets the engineering performance indicators.

[0162] Specifically, after training, the features (stress, solar radiation, traffic load, etc.) in the test set (a steel bridge monitoring dataset that does not overlap with the training set and has independent distribution characteristics) are input into the trained temperature field prediction model of the service steel bridge structure to obtain the temperature prediction value corresponding to the test set. Then, the prediction value is compared with the real-time temperature collected at the corresponding time and in the corresponding area of ​​the test set (the real temperature data synchronously monitored by the steel bridge temperature sensor) to evaluate the prediction accuracy of the model.

[0163] In this embodiment, after the model training is completed, an independent test set is used to verify the final generalization performance of the model using the improved coefficient of determination R².

[0164] ;

[0165] Where ω2 is the traffic load influence weight (taken as 0.01). Let be the vehicle density at time t, x(t) be the real-time temperature value (i.e., the measured temperature), and y(t) be the predicted temperature value. This is the average of the real-time temperature values.

[0166] When R²≥0.9 and stress deviation≤3%, the model is deemed valid, and the model that passes the test is output as the temperature field prediction model for the target service steel bridge structure.

[0167] Reference Figure 5 , Figure 5 A flowchart illustrating a method for predicting the temperature field of a steel bridge based on dual-driven dynamic graph convolution, provided in this application embodiment, is included in the following steps:

[0168] Step 1: Obtain temperature, stress, environmental parameters, and traffic load data for each target monitoring point of the steel bridge to be predicted;

[0169] Step 2: Input temperature, stress, environmental parameters, and traffic load data into the temperature field prediction model of the target in-service steel bridge structure;

[0170] Step 3: Output the predicted temperature field values ​​for different monitoring points and at different times on the steel bridge. The results are as follows: Figure 4 As shown in the figure, the measured temperature fits the predicted temperature.

[0171] This application also provides a steel bridge temperature field prediction device based on dual-driven dynamic graph convolution, including:

[0172] The data acquisition module is used to acquire temperature, stress, environmental parameters, and traffic load data for each target monitoring point of the steel bridge to be predicted.

[0173] The prediction module is used to input temperature, stress, environmental parameters and traffic load data into the temperature field prediction model of the target in-service steel bridge structure;

[0174] The output module is used to output the predicted temperature field values ​​of the steel bridge at different monitoring points and at different times.

[0175] Experimental Case: To further verify the performance of the proposed model, simulation experiments were conducted to compare the prediction accuracy of LSTM, CNN, and the proposed DG-TCN model. Figure 6 It can be seen that the curves of the three prediction models overlap significantly, indicating that all three models perform well in extracting information from the sample data. Especially in cases like... Figure 6In the DG-TCN model prediction results shown in (a), the predicted value curve almost completely overlaps with the actual value curve, achieving accurate fitting of temperature time series fluctuations (such as peaks, troughs, and periodic changes), demonstrating extremely strong time series matching ability. Meanwhile... Figure 6 In the prediction results of the LSTM / CNN neural network models shown in Figure (b), the prediction curves of LSTM or CNN are significantly weaker than the prediction accuracy of the DG-TCN model. Furthermore, as shown in Table 1, the DG-TCN model proposed in this invention has the smallest prediction error among the three models. Experimental results demonstrate that the DG-TCN prediction data is highly consistent with the real data, exhibiting excellent prediction performance. Its prediction accuracy is the highest among the three models, showcasing its advantages in predicting temporal and spatial sequence data. The DG-TCN model accurately fits the temporal fluctuation patterns and significantly compresses the prediction error, ultimately achieving significantly higher accuracy than LSTM and CNN in prediction tasks coupled with temporal and spatial data (such as the temperature field of a steel bridge).

[0176] Table 1 Comparison of prediction accuracy among the three models:

[0177] ;

[0178] Although preferred embodiments of the embodiments of this application have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0179] Finally, 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 terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0180] The above provides a detailed description of the model training method, steel bridge temperature field prediction method, and apparatus provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A model training method, characterized in that, include: The temperature, stress, environmental parameters and traffic load data of each monitoring point of the steel bridge are obtained, and finite element modeling is performed based on the preprocessed temperature, stress, environmental parameters and traffic load data to identify the main prediction area and the secondary prediction area. Based on preprocessed temperature, stress, environmental parameters and traffic load data, a temperature field prediction model for in-service steel bridge structures is established based on dynamic graph convolution (DG) and temporal convolutional network (TCN). The hyperparameter combination of the temperature field prediction model for the in-service steel bridge structure was adjusted and optimized using the Bayesian algorithm to obtain the temperature field prediction model for the in-service steel bridge structure. Evaluate whether the temperature field prediction model of the in-service steel bridge structure meets the engineering performance indicators. If not, return to the previous step to readjust the hyperparameters of the temperature field prediction model of the in-service steel bridge structure until the engineering performance indicators are met, and output the trained temperature field prediction model of the in-service steel bridge structure. Based on the improved coefficient of determination, the trained temperature field prediction model for in-service steel bridge structures is tested, and the tested temperature field prediction model for in-service steel bridge structures is used as the target temperature field prediction model for in-service steel bridge structures for temperature field prediction. Specifically, it includes: The preprocessed temperature, stress, environmental parameters, and traffic load data are divided into training, testing, and validation sets according to a set ratio. The test set data is input into the trained temperature field prediction model of the in-service steel bridge structure to obtain the corresponding temperature prediction value. The improved determination coefficient R² is calculated based on the predicted temperature value and the real-time temperature value; the formula for calculating the improved determination coefficient R² is as follows: ; in, Assuming the weighting of traffic load impact, for Real-time vehicle density This is the real-time temperature value. This is a predicted temperature value. This is the average of the real-time temperature values. Represents the total time; like Furthermore, the stress deviation between the predicted temperature secondary stress and the actual temperature secondary stress... At that time, among them, If the yield strength of the steel is given, the temperature field prediction model for the in-service steel bridge structure passes the test, and the model at this time is used as the target temperature field prediction model for the in-service steel bridge structure.

2. The method according to claim 1, characterized in that, The temperature field prediction model for the in-service steel bridge structure is constructed based on dynamic graph convolution and temporal convolution networks, including a dynamic graph convolution module, a feature fusion module, and a temporal feature extraction module. The dynamic graph convolution module is used to extract the dynamic spatial correlation features between various monitoring points in the temperature field of the in-service steel bridge structure; The feature fusion module is used to fuse the dynamic spatial correlation features output by the dynamic graph convolution module with the acquired multi-source data using an LSTM network; the multi-source data includes temperature, stress, environmental parameters, and traffic load data. The temporal feature extraction module is used to extract the temporal features of the fused feature vector using a temporal convolutional network to obtain the predicted temperature value of the steel bridge.

3. The method according to claim 2, characterized in that, The process involves using a Bayesian algorithm to adjust and optimize the hyperparameter combination of the in-service steel bridge structure temperature field prediction model, resulting in a trained in-service steel bridge structure temperature field prediction model, including: The training set data is input into the temperature field prediction model of the in-service steel bridge structure to learn the coupling law of multiple factors and temperature field, fit the model hyperparameters, and train the temperature field prediction model of the in-service steel bridge structure. The validation set data is input into the trained temperature field prediction model of the in-service steel bridge structure, and the predicted temperature value is output. Based on the obtained temperature prediction values ​​and the real-time temperatures of the corresponding monitoring points in the validation set, the basic mean square error loss and heat conduction loss terms are calculated to determine the total model loss. The confidence of the surrogate model in Bayesian optimization is updated based on the total model loss. By using the expectation enhancement criterion, select the next set of hyperparameter combinations in the hyperparameter space to be evaluated until the preset parameter convergence condition is met, and determine the optimal hyperparameter combination. The optimal hyperparameter combination after the last optimization is used as the parameters of the temperature field prediction model for the in-service steel bridge structure, and the trained temperature field prediction model for the in-service steel bridge structure is output.

4. The method according to claim 3, characterized in that, The step of inputting validation set data into the trained temperature field prediction model for in-service steel bridge structures and outputting predicted temperature values ​​includes: Each monitoring point of the steel bridge is used as a spatial node of a dynamic graph convolutional network, and the dynamic spatial correlation features of the temperature field are extracted through the dynamic graph convolutional network. Temperature, stress, meteorological, and traffic load data, along with dynamic spatial correlation features output from a dynamic graph convolutional network, are used as input layer data for the LSTM model, which outputs a fused feature vector. The fused feature vector is used as input to a temporal convolutional network. The temporal convolutional network is then used to extract the temporal features of the fused feature vector to obtain the predicted temperature of the steel bridge for the current time period.

5. The method according to claim 4, characterized in that, The step involves calculating the basic mean square error loss and heat conduction loss based on the obtained temperature prediction value and the real-time temperature of the corresponding monitoring points in the validation set, and determining the total model loss, including: Based on the simulation results of finite element modeling of the steel bridge, the main prediction region and the secondary prediction region are divided to determine the temperature prediction region of the steel bridge. Collect the real-time temperature of the corresponding monitoring point in the temperature prediction zone and calculate the basic mean square error loss; By combining the heat source term of traffic load with the theory of heat conduction, the heat conduction loss is calculated; The total model loss is obtained based on the basic mean square error loss and the heat conduction loss; the formula for calculating the total model loss is as follows: ; in, This represents the total loss of the model; As the weight for heat conduction loss, Based on the mean square error loss, This is due to heat conduction loss.

6. The method according to claim 5, characterized in that, Evaluate whether the temperature field prediction model for the in-service steel bridge structure meets the engineering performance indicators. If not, return to the previous step to readjust the hyperparameters of the temperature field prediction model until it meets the engineering performance indicators. Output the trained temperature field prediction model for the in-service steel bridge structure, including: The difference between the predicted temperature value and the real-time temperature value is taken as the temperature error. It is determined whether the temperature error is less than or equal to 2℃. If so, proceed to the next step; otherwise, readjust the hyperparameters of the temperature field prediction model of the in-service steel bridge structure. Calculate the stress deviation between the predicted temperature secondary stress and the actual temperature secondary stress. If the stress deviation exceeds... If the condition is met, the hyperparameters of the temperature field prediction model for the in-service steel bridge structure are readjusted; otherwise, the temperature field prediction model for the in-service steel bridge structure that meets the engineering performance indicators is output.

7. A method for predicting the temperature field of a steel bridge, characterized in that, The method includes: Obtain temperature, stress, environmental parameters, and traffic load data at various monitoring points of the steel bridge to be predicted; Temperature, stress, environmental parameters, and traffic load data are input into the temperature field prediction model of the target service steel bridge structure; the temperature field prediction model of the target service steel bridge structure is obtained by the method described in any one of claims 1-6; Output the predicted temperature field values ​​of the steel bridge at different monitoring points and at different times.

8. A device for predicting the temperature field of a steel bridge, characterized in that, The device includes: The data acquisition module is used to acquire temperature, stress, environmental parameters, and traffic load data at various monitoring points of the steel bridge to be predicted. The prediction module is used to input temperature, stress, environmental parameters, and traffic load data into the temperature field prediction model of the target service steel bridge structure; the temperature field prediction model of the target service steel bridge structure is obtained by the method described in any one of claims 1-6; The output module is used to output the predicted temperature field values ​​of the steel bridge at different monitoring points and at different times.

Images