Method and device for predicting dynamic mechanical properties of damaged concrete under complex environment

Abstract

This invention belongs to the field of concrete performance prediction technology. Addressing the problems of high cost, long cycle, and low data reliability of traditional SHPB tests, it provides a method and device for predicting the dynamic mechanical properties of damaged concrete under complex environments. The method constructs an independent meta-learning model based on an MLP-SSE-FWW network. In the task construction phase, each task is randomly divided into support and query sets, and these sets are randomly combined across multiple tasks, enabling the model to learn the commonalities between tasks. Secondly, the support set of the test task is forced to include the origin, and static compressive strength embedding (SSE) is introduced to embed prior knowledge into the learning process, improving the accuracy of the prediction results. Finally, its robustness is evaluated by repeated runs with different random seeds. This method solves the problem of traditional models relying on large amounts of high-quality experimental data in predicting the dynamic constitutive curves of freeze-thaw damaged concrete under sulfate conditions, demonstrating high stability and accuracy under conditions of small sample sizes.

Description

Technical Field

[0001] This application relates to the field of dynamic mechanical property prediction technology for concrete, and in particular to methods and apparatus for predicting the dynamic mechanical properties of damaged concrete under complex environments. Background Technology

[0002] Concrete, as a crucial building material, is used in residential, commercial, and industrial buildings, as well as various projects such as bridges and tunnels, water conservancy projects, infrastructure, and national defense facilities. However, many concrete structures have already been damaged due to insufficient mechanical properties before reaching their designed service life.

[0003] Freeze-thaw damage and sulfate attack are significant factors affecting concrete durability and pose a serious threat to the safety of concrete structures. Saline soils are widely distributed in my country, with the western region being the main concentrated area, and these soils are rich in sulfates. This region also features low precipitation and high evaporation, making concrete structures susceptible to severe mechanical property degradation due to wet-dry cycles. Furthermore, my country has a large cold region coverage, with a high proportion of seasonally cold areas. Due to these unique climatic conditions, cold regions experience severe cold and large seasonal temperature differences, making concrete structures vulnerable to freeze-thaw damage. In the sulfate-rich soil environment of the cold western regions, concrete structures are not only affected by freeze-thaw cycles but also by sulfate attack, leading to even more rapid deterioration of concrete performance.

[0004] During their service life, concrete structures bear not only static loads but also frequent dynamic loads such as impacts and explosions. For critical national defense facilities, in addition to considering ordinary dynamic loads, impact and blast loads must also be taken into account. Concrete is a rate-sensitive material; its dynamic strength increases compared to its static strength after being subjected to dynamic loads, thus making it a material of significant research importance.

[0005] Currently, there is considerable research on the static mechanical properties of concrete damaged by freeze-thaw cycles and sulfate attack, but research on its dynamic mechanical properties is lacking. Since the dynamic mechanical properties of concrete are an important indicator for evaluating the safety of concrete structures, this invention proposes a meta-learning-based method for predicting the dynamic mechanical properties of damaged concrete under complex environments. This method addresses the problem that traditional models rely on large amounts of high-quality experimental data to predict the dynamic constitutive curves of freeze-thaw damaged concrete under sulfate conditions. By balancing experimental costs and time, it achieves accurate predictions with less data without sacrificing accuracy, providing a new approach for predicting concrete performance under complex environments. Summary of the Invention

[0006] In view of the above problems, this application provides a method and device for predicting the dynamic mechanical properties of damaged concrete under complex environments. It aims to obtain eight dynamic constitutive curves from the impact test of concrete damaged by freeze-thaw in sulfate environment, combine the static compressive strength after damage with external environmental factors (number of freeze-thaw cycles, number of sulfate wet-dry cycles), and integrate the MLP-SSE-FWW composite model and MAML algorithm to comprehensively and accurately evaluate the dynamic constitutive curves of concrete.

[0007] In a first aspect, embodiments of this application provide a method for predicting the dynamic mechanical properties of damaged concrete under complex environments, the method comprising:

[0008] Multiple dynamic constitutive curves and static compressive strengths under different working conditions were obtained based on impact tests and uniaxial compression tests. A damage feature database was established based on the obtained dynamic constitutive curves and preprocessed. Each working condition corresponds to a task, and all tasks are divided into training tasks and testing tasks.

[0009] Based on the damage feature database, the batch size is set, and multiple training tasks are randomly combined to determine the support set and query set of the training tasks;

[0010] An independent meta-learning model based on the MLP-SSE-FWW network is constructed, and the support set, query set, and static stress strength of the training task are input into the independent meta-learning model to obtain a set of optimal meta-parameters.

[0011] The support set of the test task is input into the irrelevance meta-learning model, the optimal meta-parameters are quickly fine-tuned to obtain the fast weights of the test phase that are adapted to the test task, and predictions are made on the query set of the test task based on the fast weights of the test phase, outputting the prediction results of the dynamic mechanical properties of concrete.

[0012] Multiple random seed experiments were designed, each corresponding to a predicted dynamic constitutive curve. The RMSE results of all random seed experiments were statistically analyzed, and the RMSE results were summarized into a bar chart with error bars to evaluate the stability of the predicted dynamic mechanical properties of concrete.

[0013] Furthermore, the damage feature database includes multiple sets of dynamic features, including the number of wet-dry cycles, the number of freeze-thaw cycles, strain, and stress.

[0014] Furthermore, based on the damage feature database, the batch size is set, and multiple training tasks are randomly combined to determine the support set and query set for the training tasks, including:

[0015] Batch size is determined based on dynamic features in the damage feature database;

[0016] Multiple training tasks are randomly combined based on batch size, and all data under the working conditions corresponding to each training task are used as candidate data.

[0017] The training task is divided into a support set and a query set from the candidate data; and the support set includes at least the data corresponding to the origin of the dynamic constitutive curve.

[0018] Furthermore, the independence meta-learning model based on the MLP-SSE-FWW network is obtained by fusing the MLP-SSE-FWW network and the independence meta-learning model, wherein the MLP-SSE-FWW network includes:

[0019] A regression-type fully connected network, consisting of fully connected network layers and the GELU activation function, is used to capture the relationship between dynamic features and their predicted values.

[0020] The static strength embedding network, consisting of fully connected network layers, GELU activation function, and BatchNorm layer, is used to encode static stress strength into a high-dimensional embedding feature vector.

[0021] An extension function is used to feed fast weights into the outer loop of the irrelevant meta-learning model during meta-learning forward propagation.

[0022] Furthermore, the training task and static stress resistance are input into the irrelevance meta-learning model to obtain a set of optimal meta-parameters, including:

[0023] The static compressive strength is input into the static strength embedding network, and a high-dimensional embedding feature vector is output.

[0024] The high-dimensional embedded feature vector and the support set of a single training task are input into the inner loop of the irrelevant meta-learning model to determine the support loss and calculate the model gradient. Then, the randomly initialized initial parameters of the model are updated based on the learning rate of the inner loop to obtain the fast weights.

[0025] The expansion function is used to input fast weights into the outer loop of the irrelevant meta-learning model during the meta-learning forward propagation;

[0026] The query set and fast weights of the training task are input into a regressive fully connected network for outer loop training. The query loss is determined and the meta-parameters are updated based on the query loss.

[0027] Repeat the above steps until the maximum number of iterations is reached, and obtain a set of optimal meta-parameters.

[0028] Secondly, embodiments of this application provide a device for predicting the dynamic mechanical properties of damaged concrete under complex environments, implemented based on the above-described method. The device includes:

[0029] The database construction module obtains multiple dynamic constitutive curves and static compressive strengths under different working conditions based on impact tests and uniaxial compression tests, and establishes a damage feature database based on the obtained dynamic constitutive curves; each working condition corresponds to a task, and all tasks are divided into training tasks and testing tasks;

[0030] The task combination module is used to set the batch size based on the damage feature database, randomly combine multiple training tasks, and determine the support set and query set of the training tasks.

[0031] The model building module is used to construct an independent meta-learning model based on the MLP-SSE-FWW network, and inputs the support set and query set of the training task as well as the static stress strength into the independent meta-learning model to obtain a set of optimal meta-parameters;

[0032] The prediction module is used to input the support set of the test task into the irrelevance meta-learning model, quickly fine-tune the optimal meta-parameters to obtain the fast weights of the test phase that are adapted to the test task, and make predictions on the query set of the test task based on the fast weights of the test phase, and output the prediction results of the dynamic mechanical properties of concrete.

[0033] The evaluation module is used to conduct multiple random seed experiments, each corresponding to a predicted dynamic constitutive curve, and statistically analyze the RMSE results of all random seed experiments. Based on the RMSE results, the multiple dynamic constitutive curves are summarized into a constitutive curve with intervals, which is used to evaluate the stability of the predicted dynamic mechanical properties of concrete.

[0034] Compared with the prior art, the specific beneficial effects of the present invention are as follows:

[0035] Firstly, the SHPB test is often used in the study of the dynamic mechanical properties of materials under long-term complex environments. However, this method is costly, data acquisition is difficult, and errors are unavoidable. Environmental factors can further amplify these shortcomings, making the test cost high, the cycle long, and the data reliability difficult to guarantee. Therefore, this invention proposes a method for predicting the dynamic mechanical properties of damaged concrete under complex environments. Based on the meta-parameters optimized by MAML (initial parameters sensitive to new tasks), it predicts the dynamic constitutive curve of concrete under new working conditions. Only 1-2 steps of gradient fine-tuning are needed to achieve the expected goal, solving the pain point of traditional fine-tuning: "a large number of samples + multiple iterations".

[0036] Secondly, the independent meta-learning model does not depend on a specific model structure, thus avoiding the limitation of "strong binding between model structure and task". Therefore, this invention constructs an independent meta-learning model (MAML) based on the MLP-SSE-FWW network, integrates MAML with the MLP-SSE-FWW network, and inputs the relatively easy-to-obtain static compressive strength after damage into the static strength embedding network (SSE) to encode it into a high-dimensional feature vector, so that it can fully learn the relationship between static compressive strength and dynamic compressive strength, thereby improving the final prediction accuracy.

[0037] Finally, the robustness of this invention was comprehensively evaluated by repeatedly running it with different random seeds. This method solves the problem that traditional models rely on a large amount of high-quality experimental data to predict the dynamic constitutive curves of concrete damaged by freeze-thaw cycles in sulfate environments. Experimental results show that this invention exhibits high stability and accuracy under conditions with a small number of samples, providing a new approach for predicting concrete performance in complex environments. Attached Figure Description

[0038] 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.

[0039] Figure 1 The predicted results of the dynamic constitutive curve are as follows: 8 freeze-thaw cycles, 6 wet-dry cycles, and a static compressive strength of 52.29 MPa.

[0040] Figure 2 The predicted results of the dynamic constitutive curve are as follows: 16 freeze-thaw cycles, 12 wet-dry cycles, and a static compressive strength of 51.09 MPa.

[0041] Figure 3 The predicted results of the dynamic constitutive curve are as follows: 24 freeze-thaw cycles, 18 wet-dry cycles, and a static compressive strength of 49.38 MPa.

[0042] Figure 4 The predicted results of the dynamic constitutive curve are as follows: 32 freeze-thaw cycles, 24 wet-dry cycles, and a static compressive strength of 46.13 MPa.

[0043] Figure 5 The predicted results of the dynamic constitutive curve are as follows: 40 freeze-thaw cycles, 30 wet-dry cycles, and a static compressive strength of 41.02 MPa.

[0044] Figure 6The predicted results of the dynamic constitutive curve are as follows: 48 freeze-thaw cycles, 36 wet-dry cycles, and a static compressive strength of 36.29 MPa.

[0045] Figure 7 The predicted results of the dynamic constitutive curve are as follows: 56 freeze-thaw cycles, 42 wet-dry cycles, and a static compressive strength of 30.64 MPa.

[0046] Figure 8 This is the prediction result for working condition 2 after optimizing the support point selection strategy;

[0047] Figure 9 This is the prediction result for working condition 7 after optimizing the support point selection strategy;

[0048] Figure 10 These are the RMSE statistics of the prediction model under various operating conditions, with error bars representing the standard deviation.

[0049] Figure 11 These are the RMSE statistics for working conditions 2 and 7 after optimizing the support point selection strategy. The error bars represent the standard deviation.

[0050] Figure 12 This is a flowchart illustrating the method proposed in this invention. Detailed Implementation

[0051] 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.

[0052] Example 1: As Figure 12 As shown in this embodiment, the method for predicting the dynamic mechanical properties of damaged concrete under complex environments includes the following steps:

[0053] Step 1: Based on impact tests and uniaxial compression tests, obtain multiple dynamic constitutive curves and static compressive strengths under different working conditions. Establish a damage feature database based on the obtained dynamic constitutive curves and perform preprocessing. Each working condition corresponds to a task, and all tasks are divided into training tasks and test tasks.

[0054] Specifically, in this embodiment, eight dynamic constitutive curves under eight working conditions can be obtained through impact testing. These curves describe the stress-strain relationship of a material under dynamic loads and are obtained through experimental testing. They are commonly used to analyze the dynamic mechanical properties of materials. Each dynamic constitutive curve corresponds to a set of data on the number of sulfate wet-dry cycles, the number of freeze-thaw cycles, the static compressive strength, stress, and strain. That is, one dynamic constitutive curve contains all the parameters under one working condition.

[0055] A damage feature database was constructed based on the number of freeze-thaw cycles and the number of sulfate wet-dry cycles corresponding to eight dynamic constitutive curves. This database includes multiple sets of dynamic features, including the number of wet-dry cycles, the number of freeze-thaw cycles, strain, and stress. All data in the damage feature database underwent Z-score normalization preprocessing. Each working condition corresponds to one task, and the eight tasks were divided into seven training tasks and one test task. Each training task and test task includes a support set and a query set.

[0056] Optionally, Z-score normalization preprocessing can be performed on all data in the damage feature database, which may include the following steps:

[0057] Step 1.1: Normalize the number of wet-dry cycles using the Z-score to obtain the normalized number of wet-dry cycles. The calculation formula is as follows:

[0058] ;

[0059] in, It is the normalized number of wet-dry cycles. This is the original value of the i-th wet-dry cycle. It is the average number of all wet and dry cycles. It is the standard deviation of the number of wet-dry cycles;

[0060] Step 1.2: Normalize the number of freeze-thaw cycles using the Z-score to obtain the normalized number of freeze-thaw cycles. The calculation formula is as follows:

[0061] ;

[0062] in, It is the normalized number of freeze-thaw cycles. It is the original value of the i-th freeze-thaw cycle. It is the average number of all freeze-thaw cycles. It is the standard deviation of the number of freeze-thaw cycles;

[0063] Step 1.3: Normalize the strain using the Z-score to obtain the normalized strain. The calculation formula is as follows:

[0064] ;

[0065] in, It is the normalized strain. It is the i-th original strain value. It is the mean strain. It is the standard deviation of strain;

[0066] Step 1.4: Normalize the stress using the Z-score to obtain the normalized stress. The calculation formula is as follows:

[0067] ;

[0068] in, It is the normalized stress. It is the i-th original stress value. It is the average stress. It is the standard deviation of stress.

[0069] Step 2: Based on the damage feature database, set the batch size and randomly combine multiple training tasks to determine the support set and query set of the training tasks;

[0070] Optionally, step 2 may include the following sub-steps:

[0071] Step 2.1: Set the batch size based on the dynamic features in the damage feature database.

[0072] In this embodiment, based on the level of dry-wet cycle count, freeze-thaw cycle count, strain, and stress (i.e., feature data) contained in the damage feature database, the number of random combinations of training tasks is set, i.e., the batch size. Specifically, the damage feature database of this application corresponds to 8 tasks, and the 8 tasks correspond to 8 working conditions. The batch size can be set to 3. If it is too large, the individual features will be lost; if it is too small, the commonalities of the tasks will not be learned.

[0073] Step 2.2: Randomly combine multiple training tasks based on batch size, and use all data under the working conditions corresponding to each training task as candidate data;

[0074] Step 2.3: Divide the candidate data into a support set and a query set for the training task; and the support set shall include at least the data corresponding to the origin of the dynamic constitutive curve;

[0075] Since there is no strain without stress, all dynamic constitutive curves are curves passing through the origin (0,0) of the coordinate axis. Therefore, the support set of this invention is forced to include the origin, thereby adding the origin to the physical constraint, which is consistent with the physical meaning of the constitutive curve.

[0076] Specifically, randomly selecting m data points from the candidate data as the support set, and forcibly including the origin, and using the remaining data points as the query set, can prevent overfitting and enhance generalization ability, making more efficient use of the data.

[0077] For example, the training task x can be divided into a support set and a query set. The support set is constructed by randomly selecting 8 samples (including the origin). If the training task x has n data samples, it can be divided into... The support set consists of one set of data. The query set consists of the remaining samples. All test tasks follow the same partitioning strategy.

[0078] The training and testing tasks are completely isolated; the testing tasks do not participate in training. If there is more than one testing task, the number of random combinations of testing tasks, i.e., the batch size, is preset to 3. During the inner loop, the support sets of the testing tasks run independently to obtain fast weights. Then, the fast weights of the three tasks are passed to the outer loop for prediction. The loss of each query set under its corresponding fast weight is calculated, and these losses are summed. The average loss of the three query losses is then calculated, and the meta-parameters are optimized to learn the commonalities between the tasks.

[0079] Step 3: Construct an independent meta-learning model based on the MLP-SSE-FWW network, and input the support set and query set of the training task as well as the static stress strength into the independent meta-learning model to obtain a set of optimal meta-parameters;

[0080] The independent meta-learning model based on the MLP-SSE-FWW network is obtained by fusing the MLP-SSE-FWW network and the independent meta-learning model. The MLP-SSE-FWW network includes a regressive fully connected network (MLP), a static strength embedding network (SSE), and a spread function (FWW), wherein:

[0081] The regression-type fully connected network (MLP) consists of fully connected network layers and the GELU activation function. It is suitable for small sample learning and reduces the statistical error caused by small sample data. It is used to capture the relationship between dynamic features (number of wet-dry cycles, number of freeze-thaw cycles, strain, stress features) and predicted values ​​(dynamic compressive strength).

[0082] The Static Strength Embedding Network (SSE) consists of a fully connected network layer, a GELU activation function, and a BatchNorm layer. It is used to encode static stress strength into a high-dimensional embedding feature vector. The GELU activation function provides a constant input distribution for the BatchNorm layer, which helps BatchNorm to perform normalization work more effectively, thereby accelerating training and improving performance.

[0083] The extension function (FWW) is used to input fast weights into the outer loop of MAML during meta-learning forward propagation.

[0084] Optionally, step 3 may include the following sub-steps:

[0085] Step 3.1: Input the static compressive strength into the static strength embedding network and output a high-dimensional embedding feature vector;

[0086] It is worth noting that in this embodiment, the static compressive strength is kept high-dimensional because the material has a stress rate effect under impact behavior. Therefore, the static compressive strength is, in principle, a relatively important parameter, and it is necessary to fully understand the relationship between the static compressive strength and the dynamic compressive strength.

[0087] Step 3.2: Input the high-dimensional embedding feature vector and the support set of a single training task into the inner loop of the irrelevant meta-learning model, determine the support loss and calculate the model gradient, and then update the randomly initialized initial parameters of the model based on the learning rate of the inner loop to obtain the fast weights;

[0088] Specifically, a regression fully connected network (MLP) is used to capture the correlation between high-dimensional embedded feature vectors (dynamic stress resistance) and dynamic features. K samples are selected to calculate the support loss. The model gradient is obtained by taking the derivative of the support loss. The initial parameters of the randomly initialized model are then updated based on the inner loop learning rate, thereby updating the fast weights through a fast weight generator.

[0089] The formula for calculating the support loss is as follows:

[0090] ;

[0091] Where Loss is the support loss, K is the number of samples, pred_j is the predicted value of the j-th sample, and target_j is the actual value of the j-th sample;

[0092] The formula for updating fast weights is:

[0093] ;

[0094] Where θ' is the updated fast weight, and θ is the initial weight. The gradient is the model gradient, and α is the model gradient coefficient.

[0095] Optionally, the learning rate of the inner loop is between 0.5 and 0.001, and is generally set to 0.01.

[0096] Step 3.3: Use the spread function (FWW) to input fast weights into the outer loop of the irrelevant meta-learning model during the meta-learning forward propagation;

[0097] FWW (Forward with Weight) is an extension of the forward propagation process. Its core lies in introducing a weight adjustment mechanism to the input data or intermediate layer outputs to enhance the model's sensitivity to specific features or to implement more complex computational logic. It can provide initial loss signals for rapid weight updates in the inner loop and provide adaptation feedback for meta-parameter optimization in the outer loop.

[0098] Step 3.4: Input the query set and fast weights of the training task into the regressive fully connected network, perform outer loop training, determine the query loss, and update the meta-parameters based on the query loss;

[0099] Specifically, the query set and fast weights are input into a regression fully connected network (MLP), the predicted value is obtained through the MLP, the difference between the predicted value and the test value is used as the query loss, the query loss on each task is summed, and the average value is calculated to obtain the average loss. Finally, the meta-parameters are updated based on the average loss.

[0100] Step 3.5: Repeat steps 3.2-3.4 until the maximum number of iterations is reached, and obtain a set of optimal meta-parameters.

[0101] Step 4: Input the support set of the test task into the irrelevance meta-learning model, quickly fine-tune the optimal meta-parameters to obtain the fast weights of the test phase that are adapted to the test task, and make predictions on the query set of the test task based on the fast weights of the test phase, and output the prediction results of the dynamic mechanical properties of concrete.

[0102] Specifically, the support set of the test task is fixed at 8 sets of data, and the origin is forced to be inside. The support set of the test task is input into the inner loop of the MAML model to quickly update the optimized meta-parameters. The updated meta-parameters are then propagated to the outer loop through FWW. The query set of the test task is input into the outer loop to obtain the predicted value. This process does not update the meta-parameters, but only obtains the final result (i.e., the prediction result of the dynamic mechanical properties of concrete).

[0103] Step 5: Design multiple random seed experiments, each corresponding to a dynamic constitutive curve. Statistically analyze the RMSE results of all random seed experiments and summarize the RMSE results into an RMSE bar chart with error bars to evaluate the stability of the predicted dynamic mechanical properties of concrete.

[0104] Furthermore, to evaluate the robustness of the model, five randomized seed experiments were designed, each corresponding to a dynamic constitutive curve. The entire "meta-training + meta-testing" process was repeated, selecting different test tasks and collecting the RMSE results of all randomized seed experiments. Statistical indicators were calculated, and the results were finally summarized into an RMSE bar chart with error bars. The error bars correspond to the standard deviation of the RMSE, visually demonstrating the model's accuracy and fluctuation across different test tasks. The predicted stress arrays from the five experiments were collected, and the mean and standard deviation of the RMSE at each strain point were calculated. A mean curve was plotted, and the range of mean ± standard deviation was indicated by shading, resulting in a constitutive curve with intervals to reflect the dynamic mechanical properties of concrete. The RMSE results include the RMSE mean, RMSE standard deviation, and RMSE 95% confidence interval.

[0105] Simulation Experiment: To verify the effectiveness of the proposed method for predicting the dynamic mechanical properties of damaged concrete under complex environments, this invention considers the effects of temperature acceleration coefficient, erosion solution concentration acceleration coefficient, and freeze-thaw degradation ratio coefficient in the freeze-thaw-sulfate wet-dry cycle alternating test. Calculations were performed in a typical western region, and the freeze-thaw-sulfate wet-dry cycle alternating test regime was ultimately determined.

[0106] 1) Temperature acceleration coefficient

[0107] ;

[0108] in, T1 is the temperature acceleration coefficient; E / R is the activation energy, which is taken as 14242 here; T1 is the average temperature of the natural environment; T2 is the average temperature of the experiment.

[0109] 2) Acceleration coefficient of erosion solution concentration

[0110] ;

[0111] in, The acceleration factor for the concentration of the erosion solution is... The concentration of corrosive ions used in indoor accelerated testing This represents the concentration of corrosive ions in the actual service environment of concrete.

[0112] 3) Freeze-thaw degradation ratio coefficient

[0113] The ratio between the number of freeze-thaw cycles in the laboratory and the number of freeze-thaw cycles under natural conditions is approximately 1:10 to 1:15, with a value of 1:12. This means that one freeze-thaw cycle in the laboratory is equivalent to 12 freeze-thaw cycles under natural conditions. Therefore, the freeze-thaw degradation ratio coefficient is set to 12.

[0114] The number of freeze-thaw cycles per year was determined based on the field environment in my country, and it was assumed that one freeze-thaw cycle occurred per day. The occurrence of sulfate wet-dry cycles was calculated as follows: Number of sulfate wet-dry cycles = Number of days in the year - Number of days with freeze-thaw cycles. A freeze-thaw-sulfate wet-dry cycle test regime was adopted to simulate a real-world environment for one year. Freeze-thaw cycles accounted for accelerated freeze-thaw degradation, while sulfate wet-dry cycles accounted for accelerated temperature and accelerated concentration of the corrosive solution.

[0115] The formula for calculating the number of freeze-thaw cycles per year under test conditions is as follows:

[0116] ;

[0117] in, The number of freeze-thaw cycles per year under the test environment. This represents the number of freeze-thaw cycles per year in a real-world environment.

[0118] The formula for calculating the number of sulfate wet-dry cycles per year under experimental conditions is as follows:

[0119] ;

[0120] in, This refers to the number of sulfate wet-dry cycles per year under the test environment. This represents the number of sulfate wet-dry cycles per year in a real-world environment. The acceleration coefficient is the temperature. The acceleration factor for the concentration of the erosion solution;

[0121] The present invention simulates a one-year test environment with 8 freeze-thaw cycles and 6 freeze-thaw-sulfate dry-wet cycles, for a total of 7 large cycles.

[0122] 4) Specimen fabrication:

[0123] The concrete cylinder test mold used a 75mm diameter PVC pipe with an inner diameter of 70mm and a height of 140mm. The PVC pipe was cut vertically to facilitate subsequent demolding, and then waterproof tape was applied to prevent grout leakage.

[0124] The mixed concrete was poured into the mold in sequence, then placed on a vibrating table to vibrate it fully, removing the surface laitance and smoothing the surface. Finally, it was covered with plastic wrap and moved to the designated location. The poured specimens were placed on a plastic board and left to stand for 24 hours before being demolded. Then, they were naturally cured to the predetermined age.

[0125] After the specimens are demolded, about 10mm is cut off from one end of each large cylindrical specimen. Then, they are cut into three small cylinders with a height of 35mm. After cutting, a fine grinder is used to finely grind both ends of all the small cylinders to ensure that the non-parallelism between the two ends is within 0.02mm. The specimens are then made.

[0126] 5) Freeze-thaw cycle test:

[0127] One freeze-thaw cycle consists of a spraying phase, a cooling phase, a low-temperature isothermal phase, a heating phase, and a high-temperature isothermal phase. The duration of each phase is 5 min, 2 h, 2 h, 0.5 h, and 1.5 h, respectively, totaling 6 hours and 5 min. The low temperature is set at -20℃ to ensure the specimen can complete the freezing process at this temperature. The high temperature is set at +25℃. After the high-temperature phase, the environmental chamber system automatically sprays three times, each time lasting 1 minute, to keep the specimen in a moisture-retaining state. After one cycle is completed, the system automatically begins the next cycle.

[0128] Before the first freeze-thaw cycle test, all specimens were immersed in a 5% sodium sulfate solution for 4 days, with the solution level 20 mm above the specimens to ensure they were fully saturated. From the second major cycle onwards, the specimens were immersed in the 5% sodium sulfate solution for 1 day. After immersion, the specimens were removed and placed sequentially on the wooden strips laid in the freeze-thaw test chamber to prevent the bottom of the specimens from freezing due to residual water at the bottom of the chamber and to ensure that each surface of the specimen was evenly exposed to the temperature changes within the chamber. The specimens were then placed under the spray nozzles of the environmental chamber, and water was replenished before each freeze-thaw cycle.

[0129] 6) Sulfate wet-dry cycle test:

[0130] Immerse the specimens in a plastic tank containing a 5% Na₂SO₄ solution, maintaining a certain distance between specimens and between specimens and the tank walls to prevent them from sticking together. Cover the tank with a plastic film to prevent the sulfate solution from evaporating, and replenish or replace the solution periodically. The immersion time is 16 hours from the start of immersion until the end of the immersion process.

[0131] After the soaking process, remove the specimens and place them in a designated location to air dry for 1 hour. After air drying, place the specimens in a forced-air drying oven, paying attention to their placement to avoid stacking two specimens together. Turn on the oven and maintain the temperature at approximately 65°C for 6 hours. After drying, cool the specimens for 1 hour until the surface temperature reaches room temperature, then continue soaking them in the solution.

[0132] It is worth noting that ettringite, a sulfate erosion product in concrete, may decompose at 70°C, thereby altering the erosion mechanism. Therefore, it is necessary to ensure a drying temperature of 65°C ± 5°C; the drying temperature in this invention is 65°C.

[0133] 7) SHPB impact test:

[0134] SHPB testing technology is a mature method for studying the dynamic mechanical properties of materials, and its instrumentation and basic principles are widely recognized. Therefore, this invention will only briefly describe the specific instrument parameters and core calculation methods involved in this experiment.

[0135] This impact test employed a 74mm diameter straight-cone variable cross-section SHPB testing apparatus, which mainly includes a projectile, a compression bar system (incident bar, transmission bar, and absorbing bar), a damper, a hyperdynamic strain gauge, and a data processing system. The test equipment parameters are as follows: the elastic modulus of the compression bar is 210 GPa, the elastic wave velocity in the compression bar is 5172 m / s, and the density of the compression bar is 7850 kg / m³. 3 The incident rod is 3.2m long and the transmission rod is 1.8m long.

[0136] The SHPB test technique for studying the dynamic mechanical properties of quasi-brittle materials like concrete is based on the following two fundamental assumptions:

[0137] (1) One-dimensional stress wave assumption: The compression bar is always in a plane section elastic state, and the stress wave is a one-dimensional stress wave that propagates unidirectionally in the incident bar and the transmission bar.

[0138] (2) Assumption of uniform stress distribution in specimen: The internal stress of the specimen is uniformly distributed.

[0139] Based on the above assumptions, the derivation of the formulas for stress, strain, and strain rate is performed below. Let F1 be the force on the specimen near the incident rod, and F2 be the force on the specimen near the transmission rod. Then we have:

[0140] ;

[0141] ;

[0142] ;

[0143] In the formula: E is the elastic modulus of the compression member, GPa; A is the cross-sectional area of ​​the compression member, mm. 2 ;ε I For the incident wave; ε T For transmitted waves; ε R The reflected wave is σ; the dynamic compressive strength of the specimen is σ. It is the stress near the end face of the specimen on the incident rod. It is the stress near the end face of the specimen with the transmission rod. It is the area of ​​the specimen.

[0144] To evaluate the model's generalization ability, the model was systematically tested under different operating conditions (condition 1, condition 3, condition 4, condition 5, and condition 6) with 1, 3, 4, 5, and 6 iterations. The results correspond to... Figure 1 , Figure 3, Figure 4 , Figure 5 , Figure 6 The results show that, under conditions of limited supporting data, the average curve predicted by the model can accurately reproduce the actual dynamic stress-strain relationship. Furthermore, the narrow range of the standard deviation of the predicted stress indicates low uncertainty in the model predictions, making the results stable and reliable. Quantitative evaluation results are as follows: Figure 10 As shown, the root mean square error (RMSE) of the predictions for the above working conditions is less than 3 MPa in terms of standard deviation; referring to Table 1, especially under working condition 3, the prediction error of the peak dynamic compressive strength is only 1.04%. The above indicators fully demonstrate that the model meets the accuracy requirements for engineering applications.

[0145] Table 1. Prediction error results under operating condition 3:

[0146] ;

[0147] Figure 7 The results for condition 7 after 7 large cycles are shown in the figure. As can be seen from the figure, there is a significant error between the predicted results after 7 large cycles and the experimental data. The main reason is that the sample size of the corresponding high-cycle data is relatively small, making it difficult to fully capture the complex characteristics of the dynamic stress-strain curve at this stage. However, compared with existing technologies, this invention can still accurately predict the dynamic constitutive curve of concrete under different large cycle numbers based on limited effective data, and its accuracy is superior to existing methods.

[0148] Figure 2 The results for Condition 2 after two large cycles are shown in the figure. As can be seen, the standard deviation range of the model's predicted stress is wider in Condition 2 compared to other conditions. Analysis reveals this is mainly due to the uneven distribution of support points: only one support point is selected in the critical stage before peak strain, while subsequent random support points are concentrated after peak strain. This distribution leads to insufficient model learning before the peak strain, resulting in weaker predictive constraints and higher uncertainty in this interval. Although its mean RMSE (2.5 MPa) is still below the engineering threshold, it results in a larger standard deviation. This phenomenon indicates that systematically optimizing the support point selection strategy (such as adding key points before the peak strain) can effectively enhance model constraints, reduce the standard deviation, and thus improve the stability and reliability of predictions. The prediction results for Condition 2 and Condition 7 after support point adjustment are shown below. Figure 8 and Figure 9 As shown in the figure, the optimized prediction results are significantly improved. After optimizing the support point selection strategy, from... Figure 11 The RMSE statistics for conditions 2 and 7 show that the model prediction accuracy and stability have been significantly improved, and the significant shortening of the error bars reveals the improvement in model prediction convergence.

[0149] Example 2: This example provides a device for predicting the dynamic mechanical properties of damaged concrete under complex environments, based on a method for predicting the dynamic mechanical properties of damaged concrete under complex environments. The device includes:

[0150] The database construction module obtains multiple dynamic constitutive curves and static compressive strengths under different working conditions based on impact tests and uniaxial compression tests, and establishes a damage feature database based on the obtained dynamic constitutive curves; each working condition corresponds to a task, and all tasks are divided into training tasks and testing tasks;

[0151] The task combination module is used to set the batch size based on the damage feature database, randomly combine multiple training tasks, and determine the support set and query set of the training tasks.

[0152] The model building module is used to construct an independent meta-learning model based on the MLP-SSE-FWW network, and inputs the support set and query set of the training task as well as the static stress strength into the independent meta-learning model to obtain a set of optimal meta-parameters.

[0153] The prediction module is used to input the support set of the test task into the irrelevance meta-learning model, quickly fine-tune the optimal meta-parameters to obtain the fast weights of the test phase that are adapted to the test task, and make predictions on the query set of the test task based on the fast weights of the test phase, and output the prediction results of the dynamic mechanical properties of concrete.

[0154] The evaluation module is used to conduct multiple random seed experiments, each corresponding to a predicted dynamic constitutive curve, and statistically analyze the RMSE results of all random seed experiments. Based on the RMSE results, the multiple dynamic constitutive curves are summarized into a constitutive curve with intervals, which is used to evaluate the stability of the predicted dynamic mechanical properties of concrete.

[0155] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0156] Finally, it should be noted that in this text, relational terms such as "first" and "second" are used merely 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.

[0157] The above provides a detailed description of the method and apparatus for predicting the dynamic mechanical properties of damaged concrete under complex environments. 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 its core ideas. 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 method for predicting the dynamic mechanical properties of damaged concrete under complex environments, characterized in that, The method includes: Multiple dynamic constitutive curves and static compressive strengths under different working conditions were obtained based on impact tests and uniaxial compression tests. A damage feature database was established based on the obtained dynamic constitutive curves and preprocessed. Each working condition corresponds to a task, and all tasks are divided into training tasks and testing tasks. Based on the damage feature database, the batch size is set, and multiple training tasks are randomly combined to determine the support set and query set of the training tasks; An independent meta-learning model based on the MLP-SSE-FWW network is constructed, and the support set, query set, and static stress strength of the training task are input into the independent meta-learning model to obtain a set of optimal meta-parameters. The support set of the test task is input into the irrelevance meta-learning model, the optimal meta-parameters are quickly fine-tuned to obtain the fast weights of the test phase that are adapted to the test task, and predictions are made on the query set of the test task based on the fast weights of the test phase, outputting the prediction results of the dynamic mechanical properties of concrete. Multiple random seed experiments were designed, each corresponding to a predicted dynamic constitutive curve. The RMSE results of all random seed experiments were statistically analyzed, and the RMSE results were summarized into a bar chart with error bars to evaluate the stability of the predicted dynamic mechanical properties of concrete. The independence meta-learning model based on the MLP-SSE-FWW network is obtained by fusing the MLP-SSE-FWW network and the independence meta-learning model. The MLP-SSE-FWW network includes: A regression-type fully connected network, consisting of fully connected network layers and the GELU activation function, is used to capture the relationship between dynamic features and their predicted values. The static strength embedding network, consisting of fully connected network layers, GELU activation function, and BatchNorm layer, is used to encode static stress strength into a high-dimensional embedding feature vector. An extension function is used to feed fast weights into the outer loop of the irrelevant meta-learning model during meta-learning forward propagation.

2. The method according to claim 1, characterized in that, The damage feature database includes multiple sets of dynamic features, including the number of wet-dry cycles, the number of freeze-thaw cycles, strain, and stress.

3. The method according to claim 2, characterized in that, Based on the damage feature database, a batch size is set, and multiple training tasks are randomly combined to determine the support set and query set for each training task, including: Batch size is determined based on dynamic features in the damage feature database; Multiple training tasks are randomly combined based on batch size, and all data under the working conditions corresponding to each training task are used as candidate data. The training task is divided into a support set and a query set from the candidate data; and the support set includes at least the data corresponding to the origin of the dynamic constitutive curve.

4. The method according to claim 3, characterized in that, The training task and static compressive strength are input into the irrelevant meta-learning model to obtain a set of optimal meta-parameters, including: The static compressive strength is input into the static strength embedding network, and a high-dimensional embedding feature vector is output. The high-dimensional embedded feature vector and the support set of a single training task are input into the inner loop of the irrelevant meta-learning model to determine the support loss and calculate the model gradient. Then, the randomly initialized initial parameters of the model are updated based on the learning rate of the inner loop to obtain the fast weights. The expansion function is used to input fast weights into the outer loop of the irrelevant meta-learning model during the meta-learning forward propagation; The query set and fast weights of the training task are input into a regressive fully connected network for outer loop training. The query loss is determined and the meta-parameters are updated based on the query loss. Repeat the above steps until the maximum number of iterations is reached, and obtain a set of optimal meta-parameters.

5. A device for predicting the dynamic mechanical properties of damaged concrete under complex environments, characterized in that, The apparatus, implemented based on the method according to any one of claims 1-4, comprises: The database construction module obtains multiple dynamic constitutive curves and static compressive strengths under different working conditions based on impact tests and uniaxial compression tests, and establishes a damage feature database based on the obtained dynamic constitutive curves; each working condition corresponds to a task, and all tasks are divided into training tasks and testing tasks; The task combination module is used to set the batch size based on the damage feature database, randomly combine multiple training tasks, and determine the support set and query set of the training tasks. The model building module is used to construct an independent meta-learning model based on the MLP-SSE-FWW network, and inputs the support set and query set of the training task as well as the static stress strength into the independent meta-learning model to obtain a set of optimal meta-parameters; The irrelevance meta-learning model based on the MLP-SSE-FWW network is obtained by fusing the MLP-SSE-FWW network and the irrelevance meta-learning model. The MLP-SSE-FWW network includes: a regression fully connected network, consisting of fully connected network layers and a GELU activation function, used to capture the relationship between dynamic features and predicted values ​​of dynamic features; a static strength embedding network, including fully connected network layers, a GELU activation function, and a BatchNorm layer, used to encode static compressive strength into a high-dimensional embedded feature vector; and an expansion function, used to input fast weights into the prediction module of the outer loop of the irrelevance meta-learning model during the meta-learning forward propagation, used to input the support set of the test task into the irrelevance meta-learning model, to quickly fine-tune the optimal meta-parameters, to obtain fast weights for the test phase that are adapted to the test task, and to make predictions on the query set of the test task based on the fast weights of the test phase, outputting the prediction results of the dynamic mechanical properties of concrete. The evaluation module is used to conduct multiple random seed experiments, each corresponding to a predicted dynamic constitutive curve, and statistically analyze the RMSE results of all random seed experiments. Based on the RMSE results, the multiple dynamic constitutive curves are summarized into a constitutive curve with intervals, which is used to evaluate the stability of the predicted dynamic mechanical properties of concrete.

Images