Mass concrete construction temperature crack safety early warning method

Abstract

The application provides a mass concrete construction temperature crack safety early warning method, and belongs to the technical field of mass concrete construction temperature crack safety early warning. The application collects multi-dimensional data by arranging temperature sensors and strain sensors in a mass concrete pouring body, identifies hydration heat release parameters on line by using an extended Kalman filter, drives Kelvin chain finite element parallel solution of temperature stress, synchronously predicts full-field temperature by using a physical constraint embedded space-time graph attention artificial intelligence model, inputs an internal heat source space abnormal area into a heat conduction inverse problem gradient tracking algorithm, combines a temperature equivalent age dynamic tensile strength model and a Bayesian dynamic alarm threshold to calculate a stress intensity ratio, triggers a graded early warning when the ratio exceeds the threshold, and pushes the heat source space distribution abnormal area coordinates to a field management platform, thereby solving the technical problem that temperature crack risks in the mass concrete construction process cannot be precisely spatially positioned and dynamically warned in the initiation stage.

Description

Technical Field

[0001] This invention belongs to the field of safety early warning technology for temperature cracks in large-volume concrete construction, and specifically relates to a safety early warning method for temperature cracks in large-volume concrete construction. Background Technology

[0002] Large-volume concrete structures are widely used in water conservancy, transportation, and industrial construction. The temperature gradient caused by the heat released during cement hydration during construction is the root cause of temperature cracks. Current technologies typically use embedded thermocouples or fiber optic sensors to collect temperature data, combined with empirical formulas or simplified finite element models to assess temperature stress, and trigger alarms based on fixed thresholds. These methods have a certain application basis in practical engineering and can achieve basic temperature monitoring functions.

[0003] However, traditional methods have significant limitations. The parameters of cement hydration heat release are difficult to obtain accurately due to batch variations, and there is a systematic deviation between the finite element model with fixed parameters and the actual heat release process; the blind area outside the sensor deployment area lacks physical constraints for temperature field prediction, resulting in the inability to identify potential abnormal areas; the dynamic tensile strength evolution law with age is not fully integrated into the alarm mechanism, and the static setting of alarm thresholds leads to both false alarms and missed alarms.

[0004] In current monitoring of large-volume concrete construction, insufficient online identification of hydration parameters, lack of physical constraints for temperature field prediction in sensor blind zones, and inability to trace the source of abnormal spatial distribution of internal heat sources lead to the inability to accurately locate the initiation location of cracks before they penetrate the surface. This results in delayed alarm timing and unclear spatial orientation. In other words, existing technologies suffer from the technical problem of failing to achieve accurate spatial location and dynamic early warning of temperature crack risks during the initiation stage in large-volume concrete construction. Summary of the Invention

[0005] In view of this, the present invention provides a safety early warning method for temperature cracks in the construction of large-volume concrete, which can solve the technical problem in the prior art that the risk of temperature cracks in the construction of large-volume concrete cannot be accurately located and dynamically warned in the incipient stage.

[0006] This invention is implemented as follows: This invention provides a safety early warning method for temperature cracks in large-volume concrete construction, comprising the following steps:

[0007] Temperature and strain sensors are arranged in a three-dimensional grid on the body and surface of the large-volume concrete casting to collect temperature data, strain data, age data and three-dimensional coordinate data of each node;

[0008] An online parameter identification framework for extended Kalman filtering is constructed. The ultimate heat release and time parameters are used as state vector components and jointly estimated with the temperature state. Temperature data is used to drive the dynamic convergence of parameters, and the converged ultimate heat release and time parameters are output.

[0009] Substitute the converged ultimate heat release and time parameters into the Kelvin chain model ordinary differential equations, and use the parallel finite element solver of the graphics processor to output the temperature stress data of each node.

[0010] Temperature data, strain data, age data, and three-dimensional coordinate data are input into a physical constraint-embedded spatiotemporal graph attention model to output predicted temperature field data. Simultaneously, the predicted temperature field data is input into a gradient tracing crack source tracing algorithm for the inverse problem of heat conduction to output coordinates of anomaly regions in the spatial distribution of heat sources.

[0011] A dynamic tensile strength evolution model is established using temperature equivalent age as an index, and the dynamic tensile strength data of each node is output; the stress intensity ratio is obtained by calculating the ratio of temperature stress data to dynamic tensile strength data, and the alarm threshold is dynamically adjusted in combination with Bayesian confidence interval;

[0012] When the stress intensity ratio exceeds the alarm threshold, a graded early warning is triggered, the coordinates of the abnormal heat source spatial distribution area are pushed to the on-site management platform, and maintenance control instructions are issued.

[0013] Specifically, the temperature sensor and strain sensor are arranged in a three-dimensional grid, and the spacing between them is determined through statistical analysis of historical data from similar projects and iterative on-site pre-embedded tests.

[0014] Specifically, the extended Kalman filter online parameter identification framework involves performing a first-order Taylor expansion of the nonlinear system equations to approximate the nonlinear problem as linear, thereby recursively estimating the mean and covariance of the state vector. The initial prior values ​​of the ultimate heat release and time parameters are obtained by analyzing the mineral composition test report of the same batch of cement combined with laboratory adiabatic temperature rise test data.

[0015] Specifically, the Kelvin chain model ordinary differential equation system is obtained by expressing the creeping flexibility function as the sum of several exponential functions through Dirichlet series expansion. Each unit only needs to maintain an ordinary differential equation with one internal variable, transforming the historical integral constitutive relation into a recursive format that depends only on the current internal variable.

[0016] Specifically, the parallel finite element solver of the graphics processor distributes the assembly and solution of the overall stiffness matrix to multiple computing units of the graphics processor for synchronous execution. The element stiffness matrix is ​​stored in the graphics processor's video memory in blocks according to the node number. The matrix assembly stream and the equation solving stream are distributed to two independent CUDA streams for alternating execution.

[0017] Specifically, the physical constraint embedded spatiotemporal graph attention model constructs all sensor nodes into a dynamic graph. The attention weight of the graph attention layer is jointly determined by the node feature similarity and the physical correlation coefficient derived from the heat conduction equation. The physical correlation coefficient is superimposed on the standard attention score as an attention prior correction term.

[0018] Specifically, the temporal modeling of the physical constraint embedded in the spatiotemporal graph attention model adopts causal dilated convolution. Each graph attention layer is followed by a physical residual layer based on the finite difference operator. The physical residual layer calculates the prediction increment based on the current predicted temperature field data and the heat conduction equation and superimposes it onto the output of the graph attention layer in a residual manner.

[0019] Specifically, the training of the physical constraint embedded spatiotemporal graph attention model involves a training loss function composed of three weighted components: data fitting loss, physical residual loss of the heat conduction equation, and temperature gradient continuity regularization term. The learning rate is adjusted according to the comprehensive adjustment function of the thermal field state.

[0020] Specifically, the comprehensive adjustment function for the thermal field state is calculated by weighting three data points: the stress intensity ratio, the normalized value of the hydration heat release rate, and the normalized value of the physical residual loss of the validation set, to obtain the comprehensive thermal field state index. ,in accordance with The range of values ​​is adjusted by multiplying the learning rate by the corresponding coefficient.

[0021] Specifically, the gradient tracing crack source tracing algorithm for the inverse heat conduction problem is based on the Tikhonov regularization framework. It uses the three-dimensional unsteady heat conduction equation as the forward problem operator, sets the spatial distribution of internal hydration heat sources as the variable to be inverted, and uses the conjugate gradient method to solve iteratively. The regularization parameter is automatically determined by the L-curve method.

[0022] Specifically, the iterative solution of the gradient tracing crack source tracing algorithm for the inverse heat conduction problem involves calculating the predicted temperature field data using the positive problem operator at each step, calculating the residuals, solving the adjoint equation to obtain the adjoint variables, calculating the objective functional gradient, and updating the spatial distribution of the heat source along the negative gradient direction. The step size is determined by line search, and the computational complexity is accelerated to a certain level through multigrid processing. Magnitude.

[0023] Specifically, the dynamic tensile strength evolution model uses the temperature equivalent age as the time index. The dynamic tensile strength data increases with the temperature equivalent age in accordance with the time-varying function form recommended by the FIB specification. The temperature equivalent age is calculated based on maturity theory.

[0024] Specifically, the Bayesian confidence interval dynamically adjusts the alarm threshold by using the false alarm rate and missed alarm rate in historical early warning events as prior information, and employing the Bayesian update method to continuously correct the upper and lower confidence boundaries of the alarm threshold during the construction process.

[0025] The graded early warning system is divided into three levels: the first level is the observation level, the second level is the warning level, and the third level is the emergency level. The threshold boundaries of the three levels are determined by statistical analysis of the historical engineering crack event database.

[0026] The spacing between the temperature sensor and the strain sensor ranges from 0.5 to 2.0 mm. The initial prior value range for the ultimate heat release is 200–420. The initial prior value range for the time parameter is 10–80. The number of Kelvin chain units ranges from 4 to 8; the alarm threshold ranges from 0.7 to 0.9; the stress intensity ratio corresponds to the observation level being in the range of 0.7 to 0.8, the stress intensity ratio corresponds to the warning level being in the range of 0.8 to 0.9, and the stress intensity ratio corresponds to the emergency level being greater than 0.9; the convergence threshold range of the gradient tracing crack source tracing algorithm for the inverse heat conduction problem is... ~ The maximum number of iterations is 50 to 200.

[0027] This invention organically integrates extended Kalman filter online parameter identification, physical constraint embedded spatiotemporal graph attention model, and gradient tracking crack source tracing algorithm for the inverse problem of heat conduction, constructing a complete closed loop from dynamic parameter convergence to spatial inversion of heat source anomalies. This solves the technical problem that temperature crack risk cannot be accurately located and dynamically warned of in its incipient stage during the construction of large-volume concrete.

[0028] The extended Kalman filter framework continuously corrects the hydration exothermic parameters using measured temperature data, eliminating systematic biases in fixed-parameter models and making the heat source terms relied upon by subsequent finite element calculations closer to the actual hydration process. The physical constraint-embedded spatiotemporal graph attention model can still output a physically reasonable temperature field that satisfies the heat conduction equation even in sensor blind zones, providing a reliable full-field input for inverse problem tracing and avoiding non-physical prediction jumps in purely data-driven methods under sparse observation conditions. The gradient tracing crack tracing algorithm for the heat conduction inverse problem uses the Tikhonov regularization framework to invert abnormal regions of internal heat source spatial distribution, achieving spatial advance warning of crack initiation locations before they penetrate the surface, overcoming the deficiency of traditional methods that rely solely on surface temperature gradients for warning lag.

[0029] In summary, this invention solves the technical problem mentioned in the background art of the inability to accurately locate and dynamically warn of the risk of temperature cracks in the early stage during the construction of large-volume concrete. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention.

[0031] Figure 2 This is a comparison chart of the temperature evolution of representative nodes in the core area and surface layer from 0h to 168h after pouring, and the predicted temperature field data with the measured temperature data.

[0032] Figure 3 This is a three-dimensional spatial distribution of the stress intensity ratio of each node inside the cast body and a map showing the abnormal heat source area at 60 hours after pouring.

[0033] Figure 4 This is a timeline diagram showing the process from the start of pouring to the 168th hour early warning. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0035] like Figure 1 The diagram shown is a flowchart of a safety early warning method for temperature cracks in large-volume concrete construction provided by the present invention. This method includes the following steps:

[0036] S01. Temperature and strain sensors are arranged in a three-dimensional grid on the body and surface of the large-volume concrete casting to collect temperature data, strain data, age data and three-dimensional coordinate data of each node.

[0037] S02. Construct an online parameter identification framework for extended Kalman filtering, take the ultimate heat release and time parameters as state vector components, estimate them together with the temperature state, use temperature data to drive the dynamic convergence of parameters, and output the converged ultimate heat release and time parameters.

[0038] S03. Substitute the converged ultimate heat release and time parameters into the Kelvin chain model ordinary differential equations, and use the parallel finite element solver of the graphics processor to output the temperature stress data of each node.

[0039] S04. Input the temperature data, strain data, age data and three-dimensional coordinate data into the physical constraint embedding spatiotemporal graph attention model, and output the predicted temperature field data; simultaneously input the predicted temperature field data into the gradient tracing crack source tracing algorithm of the inverse heat conduction problem, and output the coordinates of the abnormal area of ​​the heat source spatial distribution.

[0040] S05. Establish a dynamic tensile strength evolution model using temperature equivalent age as an index, and output the dynamic tensile strength data of each node; calculate the ratio of temperature stress data to dynamic tensile strength data to obtain the stress intensity ratio, and dynamically adjust the alarm threshold in combination with Bayesian confidence interval;

[0041] S06. When the stress intensity ratio exceeds the alarm threshold, a graded early warning is triggered, the coordinates of the abnormal heat source spatial distribution area are pushed to the on-site management platform, and maintenance control instructions are issued.

[0042] The spacing between the temperature sensor and the strain sensor was determined through statistical analysis of historical data from similar projects and iterative on-site pre-embedded tests, with a spacing range of 0.5 to 2.0 mm. The specific spacing is selected based on the geometric dimensions of the concrete casting and the graded classification of cement hydration heat release intensity.

[0043] The extended Kalman filter (EPF) online parameter identification framework is a nonlinear state estimation framework based on EPF. EPF performs a first-order Taylor expansion on the nonlinear system equations, approximating the nonlinear problem as linear, thereby recursively estimating the mean and covariance of the state vector. The initial prior value range for the ultimate heat release is 200–420. The initial prior value range of the time parameter is 10 to 80. The initial prior values ​​are obtained by combining the mineral composition test report of the same batch of cement with the data analysis of the laboratory adiabatic temperature rise test. The laboratory test needs to cover no less than 5 batches of cement samples to establish the prior distribution range. At each time step, the estimated values ​​of the ultimate heat release and time parameters are corrected by the difference between the measured temperature data and the model predicted temperature, so that the parameters continue to converge as the construction progresses.

[0044] The Kelvin chain model's ordinary differential equations are obtained by expressing the creep compliance function as a sum of several exponential functions through Dirichlet series expansion, corresponding to a Kelvin chain model in mechanics composed of parallel units of springs and dampers connected in series. Each unit in the Kelvin chain model only needs to maintain an ordinary differential equation with one internal variable, thus transforming the historical integral constitutive relation into a recursive format that depends only on the current internal variable, reducing the computational complexity from... Down to ,in This represents the total number of time steps. The number of units in a Kelvin chain ranges from 4 to 8, and the unit relaxation time ranges from 0.01... Up to 1000 The elements are selected according to a logarithmic uniform distribution, and the specific element parameters are determined by least-squares fitting iteration on early-age concrete creep test data of no less than three sets of mix proportions. The parallel finite element solver of the graphics processor distributes the assembly and solution of the global stiffness matrix to multiple computing units of the graphics processor for synchronous execution. The video memory allocation strategy is to store the element stiffness matrix in blocks according to node numbers in the graphics processor's video memory; the CUDA stream allocation strategy is to allocate the matrix assembly stream and the equation solving stream to two independent CUDA streams for alternating execution to mask data transmission delays.

[0045] The physical constraint-embedded spatiotemporal graph attention model is based on deep learning, and its structure is described as follows: All sensor nodes are constructed as a dynamic graph. Each node's feature vector contains six components: temperature data, strain data, age data, and three-dimensional coordinate data. The edge weights between nodes are calculated by weighting the physical distance between nodes with the length of the heat diffusion path. The graph attention layer consists of three layers. The attention weight of each layer is determined by the node feature similarity and the physical correlation coefficient derived from the heat conduction equation. The physical correlation coefficient is used as a priori correction term for attention and is superimposed on the standard attention score, forcing the heat transfer delay effect between distant nodes to be reflected in the graph message propagation mechanism. Temporal modeling uses causal dilated convolution, with the dilation factor adjusted according to... Incremental, covering from 1 Up to 7 The model simulates multi-scale temperature evolution while maintaining strict causality. Each graph attention layer is followed by a physical residual layer based on a finite difference operator. This physical residual layer calculates the prediction increment based on the current predicted temperature field data and the heat conduction equation, and superimposes it onto the graph attention layer output as a residual, forcing the output to satisfy energy conservation constraints. The training loss function consists of a weighted average of three parts: data fitting loss, heat conduction equation physical residual loss, and temperature gradient continuity regularization term. The weight coefficients for these three parts range from 0.5 to 0.8, 0.1 to 0.3, and 0.05 to 0.15, respectively. The specific weights are determined through multiple rounds of ablation experiments on the validation set. During the inference phase, if the stress intensity ratio exceeds 0.85, an adaptive time step shortening mechanism is triggered within the model, re-inferring the state at the next time step with a smaller time step, forming a closed-loop adaptive warning cycle. The connection weights between neurons are grouped and stored into three categories: graph attention layer, causal dilated convolutional layer, and physical residual layer. Inter-layer activation tensors are pipelined in causal temporal order. Data looping between the graph attention layer and the physical residual layer employs a double-buffering strategy. GPU memory is dynamically allocated based on the product of node batch size and the receptive field length of the causal dilated convolution. Each CUDA stream independently maintains its input and output buffers. The graph attention computation stream and the physical residual computation stream are allocated to two parallel CUDA streams for alternating execution. Forward propagation is allocated to the main GPU computing cluster, while loss gradient backpropagation is allocated to an independent gradient computation stream. The number of spatial difference nodes in the finite difference operator of the physical residual layer maintains a linear correspondence with the number of Kelvin chain units in step S03. As the number of Kelvin chain units increases, the number of spatial difference nodes expands synchronously.

[0046] The steps for establishing the training dataset for the physical constraint-embedded spatiotemporal graph attention model specifically include: collecting full-process measured temperature data, strain data, age data, and three-dimensional coordinate data of no less than 10 similar large-volume concrete projects; recording the cement mineral composition, mix proportion, pouring temperature, ambient temperature, and curing measures for each project; imputing missing values ​​and removing outliers from the original data; and using a sliding window method to segment the time-series samples, with a window length ranging from 24 to 168. The step size ranges from 1 to 6. The training set, validation set, and test set are divided into three parts in a 7:2:1 ratio. The input features are normalized to zero mean and unit variance for each node, and the normalization parameters are determined by the statistics of the training set.

[0047] The specific steps for training the physical constraint-embedded spatiotemporal graph attention model include: using the Adam optimizer, with an initial learning rate range of [value missing]. ~ The batch size ranges from 16 to 64; the maximum number of training rounds is 500 rounds, and the training is terminated early when the validation set loss does not decrease for 20 consecutive rounds; after each round of training, the data fitting loss and the physical residual loss of the heat conduction equation are evaluated on the validation set, and the learning rate is adjusted according to the comprehensive adjustment function of the thermal field state; after training is completed, the error of the predicted temperature field data and the accuracy of crack early warning are verified on the test set. If the requirements are not met, the training dataset is expanded and retraining is performed.

[0048] The technical effects of embedding physical constraints into the spatiotemporal graph attention model are reflected in the following aspects: By embedding the physical heat conduction equation into the graph attention network in two forms—attention prior correction terms and physical residual layers—the model can still maintain physically reasonable predicted temperature field data in sensor blind spots, avoiding non-physical prediction jumps in purely data-driven models; causal dilated convolution enables the model to simultaneously perceive short-term abrupt changes and long-term trends without introducing future information; the adaptive time step shortening mechanism enables the early warning response to be automatically encrypted during high-risk periods, balancing computational efficiency and early warning timeliness; the overall framework organically integrates mechanistic knowledge and data learning, improving the generalization ability of predicted temperature field data under different engineering conditions.

[0049] The thermal field state comprehensive adjustment function is used to adjust the learning rate of the physical constraint embedded spatiotemporal graph attention model. This function is calculated by weighting three data points: the stress intensity ratio, the normalized value of the hydration heat release rate, and the normalized value of the validation set physical residual loss, to obtain the thermal field state comprehensive index. . The calculation formula is expressed as follows:

[0050] ;

[0051] in The stress-intensity ratio, The unit is , The unit is The ratio of the two is a dimensionless value. The ratio of the current hydration heat release rate to the reference hydration heat release rate, in units of 1 / 2. The ratio of the two is a dimensionless value. This is the ratio of the physical residual loss of the current validation set to the physical residual loss of the initial validation set. Both have the same dimension, and the ratio is a dimensionless value. These are weighting coefficients; all three are dimensionless and their sum is 1. Their specific values ​​are determined through a grid search on the validation set with a search step size of 0.05. When the learning rate is multiplied by a coefficient of 1.2; when When the learning rate remains unchanged, the learning rate remains at its current value; when When the learning rate is multiplied by a factor of 0.5, when... At that time, the learning rate is multiplied by a coefficient of 0.1 and an adaptive time step shortening mechanism is triggered simultaneously; the above interval boundary values ​​of 0.3, 0.6, and 0.85 are determined iteratively by performing piecewise statistical analysis on the training curves on no less than 5 engineering validation sets.

[0052] The reference hydration heat release rate is the peak value of the hydration heat release rate of the same batch of cement under standard curing conditions, measured in the laboratory, and the value ranges from 5 to 25. The data were obtained through analysis of laboratory isothermal calorimetry test data.

[0053] The heat conduction inverse problem gradient tracing crack source tracing algorithm is based on the Tikhonov regularization framework. It uses the three-dimensional unsteady heat conduction equation as the positive problem operator, setting the spatial distribution of internal hydration heat sources as the variable to be inverted. The known variables are the observations from a finite number of temperature measurement points on the surface and inside, i.e., temperature data. The first term of the regularization objective functional is the data fitting residual, and the second term is the spatial smoothness regularization constraint. The regularization parameter is automatically determined by the L-curve method. Iterative solution is achieved using the conjugate gradient method: each step calculates the predicted temperature field data using the positive problem operator, calculates the residual, solves the adjoint equation to obtain the adjoint variable, then calculates the gradient of the objective functional and updates the spatial distribution of heat sources along the negative gradient direction. The step size is determined by line search. The algorithm stops when the gradient norm is less than the convergence threshold or the number of iterations exceeds the upper limit. Finally, it locates the potential crack initiation location based on the anomalous clustering areas of the inverted heat source spatial distribution and outputs the coordinates of the anomalous areas. The computational complexity is accelerated to a certain level through multigrid processing. Quantity, of which This represents the number of spatially discrete nodes. The convergence threshold range is... ~ The upper limit of the number of iterations is 50 to 200 times, and the specific value is determined by iterative convergence analysis experiments on no less than 3 sets of typical engineering simulation cases.

[0054] The technical advantages of the gradient tracing crack source tracing algorithm for the inverse heat conduction problem are as follows: by inverting the spatial distribution of internal heat sources rather than relying solely on surface temperature gradients, it can detect abnormal hydration reaction zones inside the crack before the crack penetrates to the surface, achieving spatial early warning of crack initiation; the Tikhonov regularization constraint ensures that the inversion results maintain spatial smoothness even when the temperature data contains noise, suppressing false anomalies caused by measurement errors; and multi-grid acceleration enables the algorithm to complete iterations within a practical engineering timescale, meeting near real-time early warning requirements.

[0055] The temperature equivalent age is based on maturity theory, and a dynamic tensile strength evolution model is established using the temperature equivalent age as a time index. The dynamic tensile strength data increases with the temperature equivalent age following the time-varying function form recommended by the FIB standard. The alarm threshold range is 0.7 to 0.9, and the specific value is determined iteratively by conducting full-process temperature stress monitoring and destructive tests on no less than 5 groups of large-volume concrete test bodies with different cement mix ratios, and statistically analyzing the correspondence between the stress intensity ratio and the crack initiation time. The method of dynamically adjusting the alarm threshold using the Bayesian confidence interval is as follows: using the false alarm rate and false negative rate in historical early warning events as prior information, the upper and lower confidence boundaries of the alarm threshold are continuously corrected during construction using the Bayesian update method, so that the alarm threshold automatically approaches the optimum as temperature data accumulates.

[0056] The graded early warning is divided into three levels: Level 1 is the observation level, corresponding to a stress intensity ratio in the range of 0.7 to 0.8; Level 2 is the warning level, corresponding to a stress intensity ratio in the range of 0.8 to 0.9; Level 3 is the emergency level, corresponding to a stress intensity ratio exceeding 0.9. The threshold boundaries of the three levels are determined by statistical analysis of a database of historical engineering crack events, with a database sample size of no less than 30 engineering cases.

[0057] The stress intensity ratio is the ratio of the temperature stress data output in step S03 to the dynamic tensile strength data output in step S05. It is a dimensionless value, and the unit of the temperature stress data is... The unit of dynamic tensile strength data is .

[0058] Optionally, the present invention also provides a method for implementing a large-volume concrete construction temperature crack safety early warning system by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions can execute the above-mentioned method when running in the computer.

[0059] The specific implementation of step S01 is as follows: Temperature sensors and strain sensors are uniformly arranged in a three-dimensional grid on the body and surface of the large-volume concrete casting. The three-dimensional coordinates of the sensor nodes are accurately recorded and numbered during the arrangement phase. Temperature data, strain data, age data, and three-dimensional coordinate data are collected synchronously at each node. The arrangement spacing is determined through statistical analysis of historical data from similar projects and iterative on-site pre-embedded tests, with a spacing range of 0.5 to 2.0. The parameters are selected based on the geometric dimensions of the concrete casting and the intensity of cement hydration heat release. The collected multidimensional data forms the basis for subsequent parameter identification and model-driven processing, and the spatial density of the sensor arrangement directly determines the blind zone range and the accuracy of subsequent inversion.

[0060] The specific implementation of step S02 is as follows: A nonlinear state estimation framework based on extended Kalman filtering is constructed. The ultimate heat release and time parameters are used as state vector components, which are then combined with the temperature states of each node to form an augmented state vector. Extended Kalman filtering performs a first-order Taylor expansion on the nonlinear system equations, approximating the nonlinear problem as linear, and then recursively estimates the mean and covariance of the state vector. The initial prior value range for the ultimate heat release is 200–420. The initial prior value range for the time parameter is 10–80. The initial prior values ​​were obtained by combining the mineral composition test report of the same batch of cement with laboratory adiabatic temperature rise test data. At each time step, the difference between the measured temperature data and the model predicted temperature was used as the residual signal to correct the estimated values ​​of the ultimate heat release and time parameters, so that the parameters would continue to converge as the construction progressed. Finally, the converged ultimate heat release and time parameters were output, providing accurate heat source terms for subsequent finite element calculations.

[0061] The specific implementation of step S03 is as follows: Substitute the converged final heat release and time parameters into the Kelvin chain model's ordinary differential equations. The Kelvin chain model is obtained by expressing the creep compliance function as the sum of several exponential functions through Dirichlet series expansion. Mechanically, it corresponds to a Kelvin chain composed of parallel series units of springs and dampers. Each unit only needs to maintain an ordinary differential equation with one internal variable, transforming the historical integral constitutive relation into a recursive format that depends only on the current internal variable. The computational complexity is reduced from... Down to ,in This represents the total number of time steps. The parallel finite element solver of the graphics processor distributes the assembly and solution of the global stiffness matrix to multiple computing units of the graphics processor for synchronous execution. The element stiffness matrix is ​​stored in the graphics processor's video memory in blocks according to the node number. The matrix assembly stream and the equation solving stream are distributed to two independent CUDA streams and executed alternately to mask the data transmission delay. Finally, the temperature stress data of each node is output.

[0062] The specific implementation of step S04 is as follows: Temperature data, strain data, age data, and three-dimensional coordinate data are input into a spatiotemporal graph attention model with physical constraints. The model constructs all sensor nodes as a dynamic graph, and the edge weights between nodes are calculated by weighting the physical distance between nodes with the length of the heat diffusion path. There are three graph attention layers, and the attention weight of each layer is jointly determined by the node feature similarity and the physical correlation coefficient derived from the heat conduction equation. The physical correlation coefficient is superimposed as an attention prior correction term, forcing the heat transfer delay effect between distant nodes to be reflected in graph message propagation. Temporal modeling uses causal dilated convolution, covering from 1 Up to 7 The multi-scale temperature evolution is maintained with strict causality. Each graph attention layer is followed by a physical residual layer based on a finite difference operator, forcing the output to satisfy energy conservation constraints, ultimately outputting the predicted temperature field data. The predicted temperature field data is simultaneously input into a gradient tracing crack source tracing algorithm based on the inverse heat conduction problem. This algorithm, based on the Tikhonov regularization framework, uses the three-dimensional unsteady heat conduction equation as the forward problem operator and employs the conjugate gradient method to iteratively invert the spatial distribution of internal hydration heat sources. The regularization parameters are automatically determined by the L-curve method, and the computational complexity is accelerated to a minimum through multigrid processing. The magnitude is determined, and the final output is the coordinates of the spatially abnormal area of ​​the heat source.

[0063] The specific implementation of step S05 is as follows: using the temperature equivalent age as the time index, the temperature equivalent age of each node is calculated based on maturity theory. The dynamic tensile strength data increases with the temperature equivalent age following the time-varying function form recommended by the FIB standard, and the dynamic tensile strength data of each node is output. The stress intensity ratio is calculated by comparing the temperature stress data output in step S03 with the dynamic tensile strength data, and the alarm threshold range is 0.7 to 0.9. The method of dynamically adjusting the alarm threshold using the Bayesian confidence interval is as follows: using the false alarm rate and false negative rate in historical early warning events as prior information, the upper and lower confidence boundaries of the alarm threshold are continuously corrected during construction using the Bayesian update method, so that the alarm threshold automatically approaches the optimal criterion as temperature data accumulates.

[0064] The specific implementation of step S06 is as follows: When the stress intensity ratio exceeds the alarm threshold, a graded early warning is triggered. The graded early warning is divided into three levels: Level 1 (Observation Level) corresponds to a stress intensity ratio in the range of 0.7 to 0.8; Level 2 (Warning Level) corresponds to a stress intensity ratio in the range of 0.8 to 0.9; and Level 3 (Emergency Level) corresponds to a stress intensity ratio exceeding 0.9. The threshold boundaries for the three levels are determined through statistical analysis of a historical engineering crack event database, with a database sample size of no less than 30 engineering cases. After triggering the early warning, the coordinates of the abnormal spatial distribution area of ​​the heat source output in step S04 are pushed to the on-site management platform, and corresponding maintenance and control instructions are issued according to the early warning level to achieve proactive and directional intervention in the spatial location of crack initiation.

[0065] It should be noted that the key technologies of this invention include: extended Kalman filter online parameter identification technology, which incorporates hydration exothermic parameters into the state vector and jointly estimates the temperature state, enabling the hydration parameters to dynamically converge with the construction process, avoiding the systematic bias of fixed parameter models, and providing continuously updated accurate heat source terms for temperature stress calculation; physical constraint embedding spatiotemporal graph attention technology, which embeds the heat conduction equation into the deep learning model in two forms: attention prior correction term and physical residual layer, ensuring that the temperature field prediction in the sensor blind zone remains physically reasonable, and avoiding non-physical jumps in pure data-driven methods under sparse observation conditions; and heat conduction inverse problem gradient tracing crack source tracing technology, which uses the Tikhonov regularization framework to invert the spatial distribution anomaly zone of internal heat sources, realizing spatial early warning of crack initiation before it penetrates the surface. The three technologies form a causal and progressive relationship: online identification improves the driving accuracy of temperature field prediction, physical constraint prediction provides a globally consistent full-field input for the inverse problem, and inverse problem tracing transforms temperature field information into spatial coordinate output for crack initiation. The synergistic effect of the three technologies enables the overall early warning system to maintain effective operation under the triple constraints of parameter uncertainty, sensor sparsity, and the invisibility of internal anomalies.

[0066] It should be noted that during the construction of large-volume concrete, a significant temperature gradient exists between the core area and the surface of the poured body. The hydration heat release rate exhibits a nonlinear time-varying characteristic of rapid increase followed by slow decay in the early stages. Due to the differences in mineral composition between different batches of cement, the hydration heat release parameters cannot be accurately determined before pouring. Furthermore, the ambient temperature at the construction site, curing measures, and the actual pouring temperature further exacerbate the uncertainty of these parameters. Under the above conditions, if the hydration heat release parameters relied upon for the finite element method calculation of temperature stress deviate systematically from the actual process, the calculated temperature stress field will be shifted overall, leading to a distortion in the basis for judging the stress intensity ratio. The warning timing may be either too early or too late, failing to accurately reflect the actual crack risk.

[0067] The aforementioned technical problems arise because hydration exothermic reaction is a complex nonlinear chemical reaction process influenced by cement mineral composition, water-cement ratio, and temperature history. The exothermic parameters are theoretically time-dependent variables dependent on construction conditions, rather than fixed constants. Traditional methods rely on pre-calibrated, fixed parameters input into the finite element model, neglecting the systematic differences between on-site construction conditions and laboratory conditions. Furthermore, they fail to capture the dynamic changes in the exothermic rate of the same batch of cement under different temperature histories, resulting in a persistent systematic deviation between the finite element calculation results and the measured temperature field.

[0068] The usual solution to the above technical problems is to conduct isothermal calorimetry tests on cement samples before pouring to obtain exothermic parameters and then input them into the model. However, isothermal calorimetry tests are conducted at standard temperatures and cannot reflect the impact of complex temperature histories on the hydration process. The samples used in the tests are from the same batch of cement, but there are inherent batch variations within the same batch. More importantly, the actual temperature inside the poured body changes continuously during the hydration process, and the feedback effect of temperature history on the hydration rate is non-linear. Fixed-parameter models cannot fundamentally track this dynamic process, and even more refined laboratory calibration cannot eliminate deviations from field conditions.

[0069] This invention effectively solves this technical problem by extending the Kalman filter online parameter identification framework. It incorporates the ultimate heat release and time parameters into the augmented state vector, jointly estimating them with the temperature states at each node. The difference between measured temperature data and model-predicted temperature is used as a residual signal to continuously drive parameter convergence. The optimal recursive estimation theory of Kalman filtering ensures that at each time step, the parameter estimate approximates the true value in the mean square sense. As construction progresses and temperature data accumulates, the covariance of the parameter estimate continuously shrinks, eventually converging to parameter values ​​consistent with the current construction conditions. This mechanism ensures that the heat source term in the finite element model consistently tracks the actual hydration heat release dynamics throughout the construction process. The systematic deviation of the temperature stress calculation results continuously decreases as the parameters converge, thus providing a reliable computational basis for accurately judging the stress intensity ratio and guaranteeing the accuracy of early warning timing and spatial positioning.

[0070] Specifically, the principle of this invention is:

[0071] The fundamental reason why this invention can solve the above-mentioned technical problems is that its technical solution forms a logically consistent causal chain from three links: dynamic identification of hydration parameters, physical constraint prediction of the whole field temperature, and reverse problem tracing of internal heat sources. The output of each link provides the necessary high-quality input for the next link.

[0072] First, the uncertainty of the hydration heat release parameters is the root cause limiting the accuracy of temperature stress calculation. The extended Kalman filter framework incorporates the ultimate heat release and time parameters into the state vector, jointly estimating them with the temperature state. The difference between the measured temperature and the model-predicted temperature is used as a residual signal to continuously correct the parameter estimates. The rationale for this approach lies in the fact that the optimal recursive estimation theory of Kalman filtering guarantees the minimum estimation error of the online identification results in the mean square sense, rather than relying on fixed values ​​from offline calibration. This allows the heat source terms used in the Kelvin chain finite element model to dynamically approximate the actual hydration process as construction progresses, and the systematic bias of the temperature stress calculation results continuously decreases as the parameters converge.

[0073] Secondly, the sparse sensor deployment results in numerous blind zones that cannot be directly observed. Purely data-driven interpolation methods are prone to non-physical prediction jumps in these blind zones, failing to provide reliable global temperature input for the inverse problem. The physical constraint-embedded spatiotemporal graph attention model enforces the output to satisfy the heat conduction equation through two mechanisms: first, it uses the physical correlation coefficient derived from the heat conduction equation as a priori correction term for attention, ensuring that the heat transfer delay effect between distant nodes is reflected in graph message propagation; second, it connects a physical residual layer in parallel after each graph attention layer, superimposing the prediction increments calculated from the heat conduction equation as residuals to enforce energy conservation constraints. These two constraints work together to maintain a physically reasonable temperature field under sparse observation conditions, providing a globally consistent input for subsequent inversion.

[0074] Third, traditional methods rely on surface temperature gradients to assess crack risk. However, crack initiation typically begins in areas of abnormal internal hydration reactions, where surface features are not yet significant, leading to delayed early warning. The heat conduction inverse problem gradient tracing crack source tracing algorithm uses the three-dimensional unsteady heat conduction equation as the forward problem operator. It iteratively inverts the spatial distribution of internal heat sources using the conjugate gradient method. Tikhonov regularization constraints ensure the spatial smoothness of the inversion results under noisy temperature data conditions, suppressing the generation of false anomalies. The abnormal clusters of heat source spatial distribution obtained from the inversion directly correspond to the locations of abnormal internal hydration reactions, i.e., potential crack initiation zones, achieving spatial early warning.

[0075] Finally, the dynamic tensile strength evolution model uses temperature equivalent age as an index and dynamically adjusts the alarm threshold using Bayesian confidence intervals. This allows the alarm criterion for the stress-intensity ratio to continuously adapt as the concrete strength develops, fundamentally eliminating false alarms and missed alarms caused by static thresholds. It also ensures that the timing of the graded early warning commands is synchronized with the actual mechanical state of the concrete. The logical progression of these four steps determines the technical self-consistency and effectiveness of this invention.

[0076] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0077] The specific implementation of step S01 involves arranging temperature and strain sensors in a three-dimensional grid on the body and surface of the large-volume concrete casting to collect temperature, strain, age, and three-dimensional coordinate data at each node. The spacing ranges from 0.5 to 2.0 meters, and is selected based on the geometric dimensions of the concrete casting and the graded intensity of cement hydration heat release. The specific spacing is determined through statistical analysis of historical data from similar projects and iterative on-site pre-embedded tests.

[0078] The specific implementation of step S02 is to construct a nonlinear state estimation framework based on extended Kalman filtering, which will determine the ultimate heat release. With time parameters As a component of the state vector, it is combined with the temperature state to form an augmented state vector. The formula is expressed as follows:

[0079] ;

[0080] In the formula, For the first The node temperature at each time step, in °C. For the first The ultimate heat release estimated by the time step, in units of The initial prior value range is 200 to 420. For the first The time parameter for time step estimation is in hours (h). The initial prior value ranges from 10 to 80 hours. The initial prior value is obtained by analyzing the mineral composition test report of the same batch of cement combined with laboratory adiabatic temperature rise test data. The laboratory test must cover no fewer than 5 batches of cement samples. State transition function. The hydration heat release rate equation is discretized and expressed as follows:

[0081] ;

[0082] In the formula, The time step is in hours. This refers to the density of concrete, in units of... Specific heat capacity of concrete, unit: The thermal conductivity of concrete is given by [value missing]. The Laplace operator is implemented using the finite difference operator, with units of 1. The degree of hydration is dimensionless, ranging from 0 to 1, and its dependence on dimensionless time... partial derivatives Calculated from the hydration kinetic equation, it is usually approximated as follows: The discrete form is dimensionless. The state prediction equation and the covariance prediction equation are as follows:

[0083] ;

[0084] ;

[0085] In the formula, For the first Step-by-step prediction of state vector Nonlinear state transition function To predict the covariance matrix for exist The first-order Taylor expansion Jacobian matrix at the location The process noise covariance matrix For the first The posterior state vector is given by the measurement update equation:

[0086] ;

[0087] ;

[0088] ;

[0089] In the formula, Kalman gain matrix For observation function exist Jacobian matrix at the location The observation function maps the state vector to the observation space. To observe the noise covariance matrix For the first Step-by-step measured temperature vector, unit: °C identity matrix For the first Posterior covariance matrix For the first The posterior state vector is calculated at each time step. The difference between the measured temperature data and the model-predicted temperature is used to correct the error at each time step. and The estimated values ​​are used to achieve continuous convergence of parameters as construction progresses, and the final output is the converged value. and .

[0090] The specific implementation of step S03 is to convert the converged... and Substituting the creep constitutive equations based on the Kelvin chain model, and using a parallel finite element solver with a graphics processor, the temperature and stress data at each node are output. The Kelvin chain model incorporates the creep compliance function. It can be expressed as the sum of several exponential functions, as shown in the following formula:

[0091] ;

[0092] In the formula, The current time is expressed in days. The loading age is expressed in days (d). The elastic modulus at the loading age is expressed in units of 1000 m / s. The number of Kelvin chain units, ranging from 4 to 8. For the first The elastic modulus of a unit, in units of For the first The relaxation time for each unit, in days, is selected based on a logarithmic uniform distribution within the range of 0.01 days to 1000 days. and The determination was made through least-squares fitting iteration using early-age concrete creep test data from at least three mix proportions. Internal variables were introduced for each element. Its ordinary differential equation is:

[0093] ;

[0094] In the formula, For the first The internal variables of each Kelvin chain unit have a dimension of 1, consistent with the dimension of the dependent variable. Stress rate, unit: Using this recursive format, the computational complexity is reduced from... Down to , This represents the total number of time steps. The parallel finite element solver of the graphics processor stores the element stiffness matrix in video memory in blocks according to node numbers. The matrix assembly stream and the equation solving stream are assigned to two independent CUDA streams and executed alternately to mask data transmission latency.

[0095] The specific implementation of step S04 involves inputting temperature data, strain data, age data, and three-dimensional coordinate data into a physical constraint-embedded spatiotemporal graph attention model, and outputting predicted temperature field data. Feature vectors for each node... For a 6-dimensional vector, the formula is as follows:

[0096] ;

[0097] In the formula, For nodes Temperature, in °C For nodes The strain, dimensionless For nodes The concrete age is measured in hours (h). For nodes The three-dimensional coordinates, in meters. Edge weights between nodes. Physical distance between nodes With the length of the heat diffusion path Weighted calculation, the formula is expressed as follows:

[0098] ;

[0099] In the formula, For nodes With nodes Euclidean distance, in meters For nodes To node The length of the heat diffusion path, in meters. and The corresponding reference distance is in meters (m), and the empirical value is taken as the average spacing between sensors. and The weight coefficients calculated for the edge weights are all dimensionless values ​​and The default value is 0.5 for each. The value is dimensionless. The attention layer consists of 3 layers, the first... Layer nodes With nodes Attention score between Introducing physical correlation coefficient As a priori correction term, the formula is expressed as follows:

[0100] ;

[0101] In the formula, For the first Layer nodes Hidden state vector For the first Layer learnable weight matrix For the first Layer attention coefficient vector Represents vector concatenation operation The physical correlation coefficient, derived from the heat conduction equation, is dimensionless and is expressed by the following formula:

[0102] ;

[0103] In the formula, This refers to the thermal diffusivity of concrete, expressed in units of... ,Depend on Calculated The reference time step is in seconds (s), and the empirical value is the sampling interval. For reference thermal diffusion time, the unit is seconds, and the empirical value is taken as... Reflecting nodes With nodes The degree of information attenuation during propagation is determined by the physical mechanism of heat conduction. Temporal modeling employs causal dilated convolution, with the dilation factor adjusted according to... Incremental, covering multi-scale temperature evolution from 1 hour to 7 days. The physical residual layer is based on the currently predicted temperature field. Calculate the predicted increment using the heat conduction equation The formula is expressed as follows:

[0104] ;

[0105] In the formula, The model predicts the temperature field, in °C. This is the heat source for hydration, in units of... The temperature prediction increment calculated by the heat conduction equation per unit time, in units of The residuals are superimposed onto the output of the graph attention layer, forcibly satisfying the energy conservation constraint. Training loss function. It consists of three parts, and the formula is expressed as follows:

[0106] ;

[0107] In the formula, For data fitting loss, The physical residual loss of the heat conduction equation, This is the temperature gradient continuity regularization term. These are the corresponding initial loss values, used for normalization. The range is 0.5 to 0.8. The range is 0.1 to 0.3. The range is 0.05 to 0.15, and all three are dimensionless values. The specific values ​​were determined through multiple rounds of ablation experiments on the validation set. These are dimensionless values. The formulas for the three loss functions are expressed as follows:

[0108] ;

[0109] ;

[0110] ;

[0111] In the formula, Total number of sample nodes For nodes The model predicts the temperature, in °C. For nodes The measured temperature, in °C This is a reference temperature, in °C. The empirical value is taken from the initial temperature of the cast-in-place body. For nodes The temperature prediction increment calculated by the heat conduction equation is given in units of . For nodes Spatial gradient of predicted temperature, in units of For nodes Spatial gradient of measured temperature, in units of All three equations are dimensionless values. During the inference phase, when the stress intensity ratio exceeds 0.85, the model's internal adaptive time step shortening mechanism is triggered, re-inferring the state at the next moment with a smaller time step. The number of finite difference operator spatial difference nodes and the number of Kelvin chain units in the physical residual layer. Maintaining a linear correspondence, when As the number of nodes increases, the number of spatial difference nodes expands synchronously. The predicted temperature field data is simultaneously input into the gradient tracing crack source tracing algorithm for the inverse heat conduction problem. This algorithm is based on the Zikhonov regularization framework, using the three-dimensional unsteady heat conduction equation as the forward problem operator, and the variable to be inverted is the spatial distribution of internal hydration heat sources. Regularization objective functional The formula is expressed as follows:

[0112] ;

[0113] In the formula, This represents the temperature vector observed at a finite number of temperature measurement points, with units of °C. The predicted temperature vector given by the positive problem operator, in °C. The regularization parameter is dimensionless and is automatically determined by the L-curve method. For spatial location Spatial distribution of hydration heat sources at location, in units of For reference heat source intensity, the unit is... The empirical value is taken as the peak value of hydration exothermic reaction. It is a spatial position vector, with units of m. These are dimensionless values. The conjugate gradient method is used for iterative solution, with the adjoint variable obtained by solving the adjoint equation at each step. The formula for the adjoint equation is as follows:

[0114] ;

[0115] In the formula, Let be the adjoint variable, in °C, obtained by inverse integration of the above adjoint equation from the termination time to the initial time. Objective functional gradient. The formula is expressed as follows:

[0116] ;

[0117] Considering right The dimensions of the derivative are given by the above formula, which can be rearranged as follows:

[0118] ;

[0119] In the formula, The unit is Let be the gradient of the objective functional with respect to the heat source distribution; update the spatial distribution of the heat sources along the negative gradient direction, with the step size determined by the line search, when the gradient norm... Less than the convergence threshold (range: ~ The unit is The computation stops when the number of iterations exceeds the upper limit (ranging from 50 to 200), and finally outputs the coordinates of the spatial distribution anomaly region of the heat source. The computational complexity is accelerated to a certain level through multigrid processing. Magnitude.

[0120] The specific implementation of step S05 is to use temperature equivalent age. To establish a dynamic tensile strength evolution model for the index, Based on maturity theory, the formula is expressed as follows:

[0121] ;

[0122] In the formula, The apparent activation energy of cement hydration reaction is expressed in units of 1000 kJ / m². The experience value is 35,000 to 45,000. This is the gas constant, with a value of 8.314. For reference curing temperature, the unit is K, and the value is usually taken as 293.15K. for Concrete temperature at any given time, in °C The time variable is the integral time variable, and the unit is h. The unit is h. Dynamic tensile strength. Following the time-varying function form recommended by the FIB specification, the formula is expressed as follows:

[0123] ;

[0124] In the formula, Age Dynamic tensile strength at time, unit is Tensile strength at 28 days standard age, in units of This is the time correction parameter, in hours, determined by fitting experimental data. This is a dimensionless fitting exponent, determined by fitting experimental data. Stress intensity ratio. The formula is expressed as follows:

[0125] ;

[0126] In the formula, The temperature stress output in step S03, in units of This refers to dynamic tensile strength, in units of... The value is dimensionless. The thermal state comprehensive adjustment function is used to adjust the learning rate of the physical constraint embedding spatiotemporal graph attention model, and the thermal state comprehensive index is... The formula is expressed as follows:

[0127] ;

[0128] In the formula, That is, stress intensity ratio Dimensionless The current hydration heat release rate, in units of For reference hydration heat release rate, the unit is... The value represents the peak hydration heat release rate of the same batch of cement under standard curing conditions, measured in the laboratory, with a range of 5–25. For the current validation set physical residual loss, and Same dimensions The physical residual loss for the initial validation set is used for normalization. The weighting coefficients are dimensionless and The value was determined by a grid search on the validation set with a search step size of 0.05. It is a dimensionless value. When Multiply the learning rate by 1.2, when The learning rate remains constant when Multiply the learning rate by 0.5 when The learning rate is multiplied by 0.1, and an adaptive time step shortening mechanism is triggered simultaneously. The aforementioned boundary values ​​are determined iteratively through piecewise statistical analysis of training curves on at least five engineering validation sets. The alarm threshold ranges from 0.7 to 0.9. The upper and lower confidence boundaries of the alarm threshold are continuously corrected during construction using a Bayesian update method. Specific values ​​are determined through full-process temperature stress monitoring and destructive testing on at least five groups of large-volume concrete test specimens with different cement mix proportions. The correspondence between the crack initiation time and the crack initiation time was determined iteratively through statistical analysis.

[0129] The specific implementation of step S06 is that when the stress intensity ratio When the alarm threshold is exceeded, a tiered warning system is triggered: Level 1 is the observation level, corresponding to... The first level is in the range of 0.7 to 0.8; the second level is the warning level, corresponding to... The value is in the range of 0.8 to 0.9; Level 3 is the emergency level, corresponding to... The threshold exceeds 0.9. The Level 3 threshold boundary is determined through statistical analysis of a historical crack event database with a sample size of no less than 30 engineering cases. After the early warning is triggered, the coordinates of the abnormal spatial distribution area of ​​the heat source are pushed to the on-site management platform, and maintenance and control instructions are issued.

[0130] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0131] To verify the effectiveness of this invention, technicians set up a test environment and deployed the method described in this invention throughout a large-volume concrete slab pouring project to dynamically monitor and provide early warning of the risk of temperature cracks in the poured body. The planar dimensions of the poured body were 42. ×36 The thickness is 3.2. The concrete design strength grade is C40, and the cement dosage is 380. The water-cement ratio is 0.42, the pouring temperature is 22℃, the ambient temperature is 18℃, and a covering curing measure is adopted.

[0132] In step S01, temperature sensors and strain sensors are arranged according to a three-dimensional mesh, with a planar spacing of 1.0. Five layers are arranged in the thickness direction, with a spacing of 0.7m. A total of 2940 sensor nodes were deployed, each synchronously collecting temperature data, strain data, age data, and three-dimensional coordinate data, with a data acquisition time step of 0.5. .

[0133] In step S02, the initial prior value of the ultimate heat release is 310. The initial prior value of the time parameter is 45. The prior values ​​were obtained by combining the mineral composition test report of this batch of cement with the data from five batches of adiabatic temperature rise tests. The extended Kalman filter online identification framework was applied 12 days after pouring. At that time, the final heat release estimate converged to 324. The time parameter estimate converged to 38. The parameter covariance decreased to 8% of the initial value, indicating that the parameter has converged sufficiently.

[0134] In step S03, the number of Kelvin chain units is set to 6, and the relaxation time is uniformly distributed logarithmically from 0.01... Up to 1000 The GPU parallel finite element solver distributes the assembly and solution of the overall stiffness matrix to multiple computing units of the GPU for synchronous execution. The matrix assembly stream and the equation solving stream are distributed to two independent CUDA streams for alternating execution, and output temperature and stress data for each node.

[0135] In step S04, the training dataset for the physical constraint-embedded spatiotemporal graph attention model was collected from full-process experimental data of 12 similar projects. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio. The Adam optimizer was used, with an initial learning rate of [value missing]. The batch size is 32, and training is terminated early on the validation set after 400 rounds. (24th day after pouring) Up to the 72nd During the peak of hydration heat release, the model's adaptive time step shortening mechanism is automatically triggered when the stress intensity ratio exceeds 0.85, reducing the inference time step from 0.5. Reduced to 0.1 The gradient tracing crack source tracing algorithm for the inverse problem of heat conduction uses regularization parameters automatically determined by the L-curve method, solved iteratively by the conjugate gradient method, and sets a convergence threshold of [value missing]. The maximum number of iterations is set to 100, and multi-grid acceleration makes a single inversion within 90%. Completed internally. Temperature time-history data of each sensor node are as follows: Figure 2 As shown, Figure 2 Showing the 0 after pouring Up to 168 Temperature evolution curves of the inner core area and representative nodes on the surface, and comparison of predicted temperature field data with measured temperature data.

[0136] In step S05, the dynamic tensile strength of each node is calculated using the temperature equivalent age as an index. The alarm threshold is dynamically adjusted using the Bayesian confidence interval, with an initial alarm threshold of 0.80. After 72... After data accumulation, a Bayesian update adjusts the alarm threshold to 0.78, and the alarm threshold adaptively converges as construction progresses. The spatial distribution of the stress intensity ratio is as follows: Figure 3 As shown, Figure 3 Showing the 60th day after pouring The three-dimensional spatial distribution of stress intensity ratios at various nodes within the cast-in-place structure at all times, and the coordinates of abnormal heat source spatial distribution areas are displayed as superimposed points.

[0137] 58 days after pouring The gradient tracing crack source tracing algorithm for the inverse heat conduction problem identified an abnormally concentrated area of ​​heat source spatial distribution in the upper part of the core area of ​​the cast-in-place structure. The stress intensity ratio of the nodes in the abnormal area reached 0.83, triggering a level 2 warning. The coordinates of the abnormal area were pushed to the on-site management platform, and control instructions to strengthen top surface covering and local water cooling were immediately issued on-site. Subsequent continuous monitoring showed that after the intervention of maintenance and control, the stress intensity ratio of the abnormal area decreased to the 72nd level. The temperature dropped to 0.74, the Level 3 emergency warning was not triggered, and no through-temperature cracks were ultimately produced in the poured body.

[0138] The key parameters for each step are summarized in Table 1:

[0139] Table 1 Summary of Key Monitoring and Model Parameters

[0140]

[0141] The timing sequence of each stage of the early warning process is as follows: Figure 4 As shown, Figure 4 It shows the process from the start of pouring to the 168th The early warning process sequence includes the timing annotations for parameter convergence time, peak hydration heat release interval, adaptive time step shortening trigger time, early warning trigger time, and maintenance control issuance time.

[0142] Compared to traditional methods, the technological advancements of this invention are reflected in the following aspects: Traditional methods drive finite element calculations with fixed laboratory calibration parameters, failing to track the dynamic evolution of hydration heat release in the field. This invention, however, utilizes an extended Kalman filter online identification mechanism, continuously correcting the finite element heat source terms as construction progresses. Systematic biases in temperature stress calculations are eliminated as parameters converge, a feat inherently impossible with fixed-parameter methods. Traditional methods rely on interpolation from a limited number of temperature measurement points to assess crack risk, lacking physical constraints within the blind zone. This invention embeds physical constraints into a spatiotemporal graph attention model, forcibly embedding the heat conduction equation in both attention priors and physical residuals, ensuring physical rationality in the blind zone temperature field prediction—something purely data-driven methods cannot guarantee when observations are sparse. Traditional methods assess crack risk through surface temperature gradients. When cracks initiate internally, surface features are not yet significant, inevitably leading to delayed warnings. This invention, through the inverse problem of heat conduction, inverts the spatial anomaly distribution of internal heat sources, locating the initiation region before cracks penetrate the surface, achieving dual spatial and temporal advance warning—a capability inherently lacking in surface observation methods.

[0143] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0144] Table 2. Variable Explanation Table (Part 1)

[0145]

[0146] Table 3. Variable Explanation Table (Part Two)

[0147]

[0148] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for early warning of temperature cracks in large-volume concrete construction, characterized in that, Includes the following steps: Temperature and strain sensors were arranged in a three-dimensional grid on the body and surface of the large-volume concrete casting to collect temperature data, strain data, age data and three-dimensional coordinate data of each node; An online parameter identification framework for extended Kalman filtering is constructed. The ultimate heat release and time parameters are used as state vector components and jointly estimated with the temperature state. Temperature data is used to drive the dynamic convergence of parameters, and the converged ultimate heat release and time parameters are output. Substitute the converged ultimate heat release and time parameters into the Kelvin chain model ordinary differential equations, and use the parallel finite element solver of the graphics processor to output the temperature stress data of each node. Temperature data, strain data, age data, and three-dimensional coordinate data are input into a physical constraint-embedded spatiotemporal graph attention model to output predicted temperature field data. Simultaneously, the predicted temperature field data is input into a gradient tracing crack source tracing algorithm for the inverse problem of heat conduction to output coordinates of anomaly regions in the spatial distribution of heat sources. A dynamic tensile strength evolution model is established using temperature equivalent age as an index, and the dynamic tensile strength data of each node is output. The stress intensity ratio is obtained by calculating the ratio of temperature stress data to dynamic tensile strength data, and the alarm threshold is dynamically adjusted in combination with the Bayesian confidence interval. When the stress intensity ratio exceeds the alarm threshold, a graded early warning is triggered, the coordinates of the abnormal heat source spatial distribution area are pushed to the on-site management platform, and maintenance control instructions are issued.

2. The method for early warning of temperature cracks in large-volume concrete construction according to claim 1, characterized in that, The temperature and strain sensors are arranged in a three-dimensional grid, and the spacing between them is determined through statistical analysis of historical data from similar projects and iterative on-site pre-embedded tests.

3. The method for early warning of temperature cracks in large-volume concrete construction according to claim 2, characterized in that, The extended Kalman filter online parameter identification framework specifically involves performing a first-order Taylor expansion of the nonlinear system equations to approximate the nonlinear problem as linear, thereby recursively estimating the mean and covariance of the state vector. The initial prior values ​​of the ultimate heat release and time parameters are obtained by analyzing the mineral composition test report of the same batch of cement combined with laboratory adiabatic temperature rise test data.

4. The method for early warning of temperature cracks in large-volume concrete construction according to claim 3, characterized in that, The Kelvin chain model's set of ordinary differential equations is specifically obtained by expressing the creeping flexibility function as the sum of several exponential functions through Dirichlet series expansion. Each unit only needs to maintain an ordinary differential equation with one internal variable, transforming the historical integral constitutive relation into a recursive format that depends only on the current internal variable.

5. The method for early warning of temperature cracks in large-volume concrete construction according to claim 4, characterized in that, Specifically, the parallel finite element solver of the graphics processor distributes the assembly and solution of the overall stiffness matrix to multiple computing units of the graphics processor for synchronous execution. The element stiffness matrix is ​​stored in the graphics processor's video memory in blocks according to the node number. The matrix assembly stream and the equation solving stream are distributed to two independent CUDA streams for alternating execution.

6. The method for early warning of temperature cracks in large-volume concrete construction according to claim 5, characterized in that, The physical constraint embedding spatiotemporal graph attention model specifically involves constructing all sensor nodes as a dynamic graph. The attention weights of the graph attention layer are jointly determined by the node feature similarity and the physical correlation coefficient derived from the heat conduction equation. The physical correlation coefficient is superimposed on the standard attention score as an attention prior correction term.

7. The method for early warning of temperature cracks in large-volume concrete construction according to claim 6, characterized in that, The physical constraint embedding time modeling of the spatiotemporal graph attention model specifically adopts causal dilated convolution. After each graph attention layer, a physical residual layer based on the finite difference operator is connected in parallel. The physical residual layer calculates the prediction increment based on the current predicted temperature field data and the heat conduction equation and superimposes it onto the output of the graph attention layer in a residual manner.

8. The method for early warning of temperature cracks in large-volume concrete construction according to claim 7, characterized in that, The training of the physical constraint-embedded spatiotemporal graph attention model specifically involves a training loss function composed of three weighted components: data fitting loss, physical residual loss of the heat conduction equation, and temperature gradient continuity regularization term. The learning rate is adjusted according to the comprehensive adjustment function of the thermal field state.

9. The method for early warning of temperature cracks in large-volume concrete construction according to claim 8, characterized in that, The dynamic tensile strength evolution model specifically uses the temperature equivalent age as the time index. The dynamic tensile strength data increases with the temperature equivalent age in accordance with the time-varying function form recommended by the FIB specification. The temperature equivalent age is calculated based on maturity theory.

10. The method for early warning of temperature cracks in large-volume concrete construction according to claim 9, characterized in that, The gradient tracing crack source tracing algorithm for the inverse heat conduction problem is specifically based on the Tikhonov regularization framework. It uses the three-dimensional unsteady heat conduction equation as the forward problem operator, sets the spatial distribution of internal hydration heat sources as the variable to be inverted, and uses the conjugate gradient method to solve iteratively. The regularization parameter is automatically determined by the L-curve method.

Images