Method for predicting residual compressive strength of frp reinforced nano-modified uhpc after high temperature exposure

Abstract

The application relates to a residual compressive strength prediction method of FRP reinforced nano-modified UHPC after high-temperature exposure. The residual compressive strength prediction method of FRP reinforced nano-modified UHPC after high-temperature exposure provided by the application comprises the following steps: UHPC characteristic vectors are obtained by respectively preprocessing UHPC proportioning parameters and nano material parameters, and the UHPC characteristic vectors are respectively subjected to feature extraction together with FRP reinforcement parameters to respectively obtain UHPC features and FRP reinforcement features, and then the features are sequentially subjected to feature fusion, and deep feature extraction is performed together with a high-temperature exposure parameter to obtain high-temperature exposure UHPC features, which are respectively connected with the UHPC features and the FRP reinforcement features to respectively generate residual compressive features of high-temperature UHPC and high-temperature FRP; and the residual compressive features of the high-temperature UHPC and the high-temperature FRP are subjected to feature fusion and prediction to obtain a residual strength value. The residual compressive strength prediction method of FRP reinforced nano-modified UHPC provided by the application improves the residual strength prediction accuracy of FRP reinforced nano-modified UHPC after high-temperature exposure.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of performance prediction of concrete, and particularly relates to a residual compressive strength prediction method and model for FRP reinforced nano-modified UHPC after high temperature exposure, an electronic device and a computer storage medium. BACKGROUND

[0002] As the most commonly used building material at present, concrete is widely used in various civil engineering fields such as roads, bridges, high-rise buildings, and water structures due to its abundant raw materials, low cost, mature construction technology, and good performance. Under the new development concept of green environmental protection in the new era, ordinary concrete has been unable to meet the higher and stronger engineering requirements. In this case, ultra-high performance concrete (UHPC) has been increasingly applied in engineering construction due to its ultra-high strength and excellent durability.

[0003] In order to further improve the bearing capacity and crack resistance of UHPC components, the existing technology introduces fiber reinforced composite materials (FRP) to externally wrap and reinforce UHPC. FRP materials generally have high strength, high elastic modulus, strong corrosion resistance, and light weight, and common types include carbon fiber composite materials (CFRP) and glass fiber composite materials (GFRP). The enhancement principle is that when UHPC expands laterally during compression (Poisson effect), the FRP layer wrapped on its surface provides lateral restraint, thereby forming a triaxial stress state inside the UHPC, effectively inhibiting the propagation of microcracks and improving the confined compressive strength and ductility of UHPC. Therefore, FRP reinforcement technology is widely used in building structure reinforcement, seismic retrofitting, and post-disaster repair engineering practices.

[0004] Although FRP can effectively improve the confined compressive strength of UHPC, its enhancement effect is easily degraded in high temperature exposure environments (such as fires). The main reason is that the enhancement effect of FRP depends on the bonding performance between the resin matrix and the fibers, and the resin matrix (especially epoxy resin) will soften and lose adhesion properties in the range of 150-250 degrees Celsius, causing the FRP layer to delaminate, delaminate, and even burn damage. At this time, the lateral restraint provided by FRP to UHPC rapidly decays, causing UHPC to return to a uniaxial compression state, resulting in strength degradation, crack propagation, and deformation intensification under the combined action of high temperature and structural thermal damage, ultimately leading to a significant reduction in residual bearing capacity, forming the so-called residual compressive strength problem.

[0005] In order to alleviate the problem of residual compressive strength, the prior art introduces nano materials such as nano silicon dioxide (Nano-SiO2), nano titanium dioxide (Nano-TiO2), nano aluminum oxide or carbon nanotubes into UHPC to form nano modified UHPC to improve the microstructure density and crack resistance of the original UHPC. Nano materials have the characteristics of small particle size, large specific surface area and high chemical reactivity. After appropriate mixing, it can promote the cement hydration reaction, fill the micro-pore structure and increase the amount of hydration product, thereby improving the thermal stability and structural integrity of FRP reinforced nano modified UHPC after thermal damage, and then enhancing the residual compressive strength of FRP reinforced nano modified UHPC after high temperature exposure.

[0006] However, the change of residual strength of FRP reinforced nano modified UHPC is jointly affected by multiple factors such as original UHPC ratio, FRP type and layer number, nano material type and dosage, temperature exposure level and duration, etc. The variables show a highly nonlinear coupling characteristic. In particular, different dosages of nano materials and different ratios of original UHPC, combined with different layers and different types of FRP, the influence effect on residual strength shows cross enhancement, inhibition or saturation, etc. It is difficult to predict the residual strength by linear model or simple nonlinear model.

[0007] And the prior art generally uses machine learning methods such as artificial neural network (ANN) and other machine learning models to predict the residual strength of UHPC reinforced by nano materials after high temperature environment, but ignores the thermal degradation behavior and interface failure mechanism of FRP reinforced system; Or, based on material parameters, geometric characteristics and FRP characteristics, a model is constructed to predict the limited compressive strength of FRP reinforced UHPC under normal temperature conditions, but the residual strength of FRP reinforced nano modified UHPC after high temperature exposure is not considered.

[0008] Therefore, the prior art cannot uniformly model the "thermal degradation process of FRP reinforced system" and "nonlinear improvement effect of nano materials on UHPC thermal performance", and does not establish a residual strength estimation under the three factors of "thermal environment, FRP parameters and nano dosage". Therefore, the prior art has the problem of low accuracy of estimating the residual compressive strength of FRP reinforced nano modified UHPC after high temperature action. SUMMARY

[0009] Based on this, the purpose of the present application is to provide a method for predicting the residual compressive strength of FRP reinforced nano modified UHPC after high temperature exposure.

[0010] A method for predicting the residual compressive strength of FRP reinforced nano modified UHPC after high temperature exposure, comprising the following steps:

[0011] S1: pre-process the UHPC proportioning parameter and the nano material parameter to obtain a UHPC feature vector;

[0012] S2: respectively perform feature extraction on the UHPC feature vector and the FRP reinforcement parameter to obtain UHPC features and FRP reinforcement features;

[0013] S3: perform feature fusion on the UHPC features and the FRP reinforcement features to obtain a strength fusion feature of the UHPC;

[0014] S4: perform deep feature extraction on the strength fusion feature of the UHPC and the high-temperature exposure parameter to obtain a high-temperature feature of the UHPC;

[0015] S5: respectively perform residual connection on the high-temperature feature of the UHPC and the UHPC features and the FRP reinforcement features to obtain residual compressive strength features of the high-temperature UHPC and the high-temperature FRP;

[0016] S6: perform feature fusion on the residual compressive strength features of the high-temperature UHPC and the high-temperature FRP to obtain a comprehensive feature of the residual compressive strength;

[0017] S7: predict the comprehensive feature of the residual compressive strength to obtain a residual strength value.

[0018] The residual compressive strength prediction method of the FRP reinforced nano modified UHPC after high-temperature exposure provided by the application, compared with the prior art, by performing feature extraction and fusion on the UHPC proportioning parameter, the nano material parameter and the FRP reinforcement parameter before introducing the high-temperature exposure parameter, constructing the strength fusion feature of the FRP reinforced nano modified UHPC, effectively enhancing the feature representation ability of the original physical parameter before being heated, making the model can express the structure performance of the FRP reinforced nano modified UHPC in the normal state, so that when the high-temperature exposure parameter is fused, the extracted feature has good “basic feature bearing capacity”, significantly improving the stability and generalization ability of the residual compressive strength prediction.

[0019] In addition, by introducing the residual connection mode, the features under high-temperature erosion are effectively fused with the original structure features, while the degradation features are retained, the model's memory ability for the pre-heat strength is enhanced, thereby improving the generalization ability and prediction accuracy of the residual compressive strength prediction under high-temperature action, solving the problem of inaccurate residual compressive strength prediction of the FRP reinforced nano modified UHPC under high-temperature action in the prior art.

[0020] Further, the feature extraction of the step S2 is performed by a UHPC feature extraction module and a FRP reinforcement feature extraction module,

[0021] The UHPC feature extraction module is configured to perform nonlinear feature extraction on the UHPC feature vector to obtain UHPC features, wherein the nonlinear feature extraction comprises a stacked fully connected layer and a nonlinear activation function, and is specifically represented as follows:

[0022] Feature UHPC =(ReLU(FC(UHPC))) n

[0023] In the formula, Feature UHPC represents the UHPC features; UHPC represents the input UHPC feature vector; FC represents the fully connected layer; ReLU is the nonlinear activation function; and n represents the number of stacks.

[0024] The FRP reinforcement feature extraction module is configured to perform preliminary feature extraction on the FRP reinforcement parameters to obtain FRP reinforcement features, wherein the preliminary feature extraction comprises an embedding layer, a fully connected layer and a linear layer, and is specifically represented as follows:

[0025] Feature FRP =Linear(FC(EM(FRP)))

[0026] In the formula, Feature FRP represents the FRP reinforcement features; FRP represents the input FRP reinforcement parameters; EM represents the embedding layer; and Linear represents the linear layer.

[0027] The step S3 comprises the following sub-steps:

[0028] The UHPC features and the FRP reinforcement features are subjected to a splicing operation to generate a comprehensive feature vector, and the comprehensive feature vector is subjected to feature fusion by using a multilayer perceptron and a nonlinear activation function to obtain the strength fusion features of the UHPC, and is specifically represented as follows:

[0029] Feature fusion =ReLU(MLP(Concat(Feature UHPC ,Feature FRP )))

[0030] In the formula, Feature fusion represents the strength fusion features of the UHPC; MLP is a multilayer perceptron, representing a combination of multiple fully connected layers; and Concat represents the splicing operation.

[0031] The deep feature extraction of the step S4 is performed by the temperature feature fusion module and the deep feature extraction module.

[0032] The temperature feature fusion module is configured to splice the UHPC fusion feature and the high-temperature exposure parameter to generate a UHPC spliced feature after high-temperature exposure, and is specifically expressed as follows:

[0033] Feature combined = Concat(Feature fusion , T, t)

[0034] In the formula, Feature combined represents the UHPC spliced feature after high-temperature exposure; T and t are temperature parameters and exposure time parameters in the normalized high-temperature exposure parameter respectively;

[0035] The deep feature extraction module is configured to sequentially perform feature extraction, self-attention weighting and nonlinear activation on the UHPC spliced feature after high-temperature exposure to generate a high-temperature feature of the UHPC, and is specifically expressed as follows:

[0036] Feature temp = ReLU(Attention(MLP(Feature combined )))

[0037] In the formula, Feature temp is the high-temperature feature of the UHPC; and Attention is a self-attention mechanism.

[0038] The step S6 includes the following sub-steps:

[0039] The residual compressive strength features of the high-temperature UHPC and the high-temperature FRP are sequentially subjected to feature fusion, multi-layer perception, self-attention and a nonlinear activation function to obtain a comprehensive feature of the residual compressive strength, and are specifically expressed as follows:

[0040] Feature last = ReLU(Attention(MLP(Concat(Feature tempUHPC , Feature tempFRP ))))

[0041] In the formula, Feature last represents the comprehensive feature of the residual compressive strength; Feature tempUHPC represents the residual compressive strength feature of the high-temperature UHPC; and Feature tempFRP represents the residual compressive strength feature of the high-temperature FRP.

[0042] Accordingly, the application introduces a self-attention mechanism in the UHPC high-temperature feature extraction stage and the comprehensive feature fusion stage of the residual compressive strength, respectively, to dynamically identify and highlight the key factors in the high-temperature exposure parameters and the nano-modified UHPC features and the FRP reinforcement features that have the most significant influence on the residual strength, thereby improving the model's understanding ability of the complex damage mode under high-temperature environment, and further significantly improving the accuracy and adaptability of the residual compressive strength prediction.

[0043] Further, a training method based on a phased heuristic optimization is used to train the trainable parameters in the residual compressive strength prediction method of the FRP reinforced nano-modified UHPC, which specifically includes the following steps:

[0044] T1, collect the compressive strength data set and the residual compressive strength data set of the FRP reinforced nano-modified UHPC;

[0045] T2, divide the compressive strength data set and the residual compressive strength data set to generate the training set and the test set corresponding to the compressive strength and the residual compressive strength, respectively;

[0046] T3, use the PSO heuristic algorithm and the test set of the compressive strength to optimize the weights of the UHPC feature extraction module and the FRP reinforcement feature extraction module in step S2, and the trainable parameters of the multilayer perceptron in step S3, to obtain the initialized UHPC feature extraction module, FRP reinforcement feature extraction module and multilayer perceptron;

[0047] T4, input the training set and the test set of the compressive strength into the initialized UHPC feature extraction module, FRP reinforcement feature extraction module and multilayer perceptron for training, and freeze the weights to obtain the trainable parameters in the preliminary trained residual compressive strength prediction method of the FRP reinforced nano-modified UHPC;

[0048] T5, use the LSO heuristic algorithm to perform secondary weight optimization on the preliminary trained residual compressive strength prediction model of the FRP reinforced nano-modified UHPC to generate the trainable parameters in the secondary initialized residual compressive strength prediction method of the FRP reinforced nano-modified UHPC;

[0049] T6, according to the training set and the test set of the residual compressive strength, train the trainable parameters in the secondary initialized residual compressive strength prediction model of the FRP reinforced nano-modified UHPC to complete the training of the trainable parameters in the residual compressive strength prediction method of the FRP reinforced nano-modified UHPC.

[0050] Accordingly, by means of the PSO heuristic algorithm, the trainable parameters of the FRP reinforced nano-modified UHPC feature extractor under the condition of no high temperature influence are searched to ensure that the most basic and meaningful features of the FRP reinforced nano-modified UHPC can be effectively extracted in the initial stage, prevent the uncertainty interference caused by the introduction of the high temperature exposure parameter in the subsequent stage, thereby significantly improving the learning ability of the model to the compressive strength features of the FRP reinforced nano-modified UHPC at room temperature, and providing a solid foundation for the subsequent training stage of the introduction of the high temperature exposure parameter.

[0051] In addition, after the trainable parameters of the feature extractor without high temperature influence are frozen, the LSO is used to perform global search on the trainable parameters in the whole residual compressive strength prediction to find the optimal initial weight, so that the trainable parameters for the residual compressive strength prediction can better adapt to the feature learning under the complex high temperature environment in the subsequent training, and premature falling into a local optimal solution is avoided, thereby reducing the accumulation of errors in the training process and significantly improving the training precision and convergence speed.

[0052] Further, the PSO heuristic algorithm specifically includes the following steps:

[0053] T31, initializing the particles of the PSO to obtain an initial population; wherein the population is composed of multiple particles, and each particle includes a position and a corresponding speed, the position is used to represent the combination of the trainable parameters corresponding to the current particle, and the speed is used to represent the moving direction and step size of the particle in the parameter search space;

[0054] T32, calculating the fitness of the current population according to the mean square error loss function and the test set of the compressive strength to obtain the fitness corresponding to the current population;

[0055] T33, updating the state of the current population according to the fitness corresponding to all particles in the current population to obtain an updated population;

[0056] T34, judging whether the fitness corresponding to the current population meets an iteration condition: if not, step T32 is continuously executed; if yes, step T35 is executed;

[0057] T35, taking the trainable parameters corresponding to the particle with the lowest fitness in the current population as the optimal trainable parameters, and substituting the trainable parameters into the trainable parameters of the UHPC feature extraction module and the FRP reinforcement feature extraction module in step S2 and the multilayer perceptron in step S3 to obtain the UHPC feature extraction module and the FRP reinforcement feature extraction module in step S2 and the multilayer perceptron in step S3 after initialization;

[0058] The LSO heuristic algorithm specifically comprises the following steps:

[0059] T51, parameters of the LSO heuristic algorithm are initialized to generate a plurality of rays;

[0060] T52, according to the mean square error loss function and the test set of the residual compressive strength, the fitness of all the current rays is calculated to obtain the fitness corresponding to all the current rays;

[0061] T53, whether the fitness corresponding to all the current rays is close to the fitness of the current global optimal solution is judged: if yes, the current global optimal solution is updated, and step T54 is executed; if no, step T54 is directly executed;

[0062] T54, according to the fitness corresponding to all the current rays, the random state of all the current rays is updated to obtain the updated rays;

[0063] T55, whether the updated rays meet an iteration threshold is judged: if no, step T52 is continuously executed; if yes, the current global optimal solution is taken as a trainable parameter in a residual compressive strength prediction method of the secondary FRP reinforced nano-modified UHPC, and secondary weight optimization is completed.

[0064] Accordingly, the PSO (particle swarm optimization algorithm) ensures that the basic features of the FRP reinforced nano-modified UHPC can be accurately extracted in the prediction process without high-temperature interference. Meanwhile, by utilizing the powerful global search characteristics of the LSO (light spectrum optimization algorithm), all the trainable parameters are further optimized, especially in the feature learning process under the influence of high temperature, the local optimal solution is jumped out, so that the optimal weight required for the residual compressive strength prediction of the UHPC and the FRP reinforced system under high-temperature exposure can be more accurately found, the estimation accuracy of the residual compressive strength of the FRP reinforced nano-modified UHPC after high-temperature action is significantly improved, and the convergence speed of the training process is effectively improved.

[0065] A residual compressive strength prediction model of FRP reinforced nano-modified UHPC after high-temperature exposure, comprising a UHPC preprocessing unit, a double-branch feature extraction unit, a UHPC compressive strength feature fusion unit, a UHPC high-temperature feature extraction unit, a double-branch high-temperature feature extraction unit, a UHPC residual compressive strength feature fusion unit and a residual strength prediction unit;

[0066] The UHPC preprocessing unit is used for preprocessing UHPC proportioning parameters and nano material parameters to obtain a UHPC feature vector;

[0067] The double-branch feature extraction unit is configured to perform feature extraction on the UHPC feature vector and the FRP reinforcement parameter respectively, and obtain UHPC features and FRP reinforcement features.

[0068] The UHPC compressive strength feature fusion unit is configured to perform feature fusion on the UHPC features and the FRP reinforcement features, and obtain strength fusion features of the UHPC.

[0069] The UHPC high-temperature feature extraction unit is configured to perform deep feature extraction on the strength fusion features of the UHPC and the high-temperature exposure parameter, and obtain high-temperature features of the UHPC.

[0070] The double-branch high-temperature feature extraction unit is configured to perform residual connection on the high-temperature features of the UHPC and the UHPC features and the FRP reinforcement features respectively, and obtain residual compressive features of the high-temperature UHPC and the high-temperature FRP.

[0071] The UHPC residual compressive strength feature fusion unit is configured to perform feature fusion on the residual compressive features of the high-temperature UHPC and the high-temperature FRP, and obtain comprehensive features of the residual compressive strength.

[0072] The residual strength prediction unit is configured to perform prediction on the comprehensive features of the residual compressive strength, and obtain a residual strength value.

[0073] In order to better understand and implement, the present application is described in detail below in conjunction with the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0074] Figure 1 A simple structure schematic diagram of the residual compressive strength prediction model of the FRP reinforced nano-modified UHPC after high-temperature exposure according to the present application;

[0075] Figure 2 A simple flowchart schematic diagram of the residual compressive strength prediction method of the FRP reinforced nano-modified UHPC after high-temperature exposure according to the present application;

[0076] Figure 3 A simple structure schematic diagram of the residual compressive strength prediction model. DETAILED DESCRIPTION

[0077] To address the issue of low accuracy in estimating the residual compressive strength of FRP-reinforced nano-modified UHPC after high-temperature exposure in existing technologies, this invention preprocesses the UHPC formulation parameters and nanomaterial parameters separately to obtain UHPC feature vectors. These vectors are then combined with FRP reinforcement parameters for feature extraction, yielding UHPC features and FRP reinforcement features respectively. Next, the UHPC and FRP reinforcement features are fused to obtain UHPC strength fusion features. Furthermore, depth feature extraction is performed using high-temperature exposure parameters to obtain high-temperature exposed UHPC features. These features are then fused with both the UHPC and FRP reinforcement features to generate residual compressive strength features for high-temperature UHPC and high-temperature FRP respectively. Finally, the residual compressive strength features of high-temperature UHPC and high-temperature FRP are fused to obtain a comprehensive feature of residual compressive strength. This comprehensive feature is then used to predict the residual strength value.

[0078] Accordingly, this invention models the strength of FRP-reinforced nano-modified UHPC before high-temperature exposure, especially the feature construction of nanomaterial parameters and UHPC ratio, the feature construction of FRP reinforcement, and the fusion of the two. That is, by decoupling the strength characteristics before and after high-temperature exposure, the nonlinear effects of nano-modified UHPC itself and FRP before and after high-temperature exposure are effectively captured, thereby improving the accuracy of predicting the residual strength of FRP-reinforced nano-modified UHPC after high-temperature exposure.

[0079] In addition, after the high-temperature exposure parameters are fused and extracted with the aforementioned fusion features, the features of nano-modified UHPC and FRP are further processed by high-temperature feature modeling to ensure that the potential feature representations of high-temperature exposure features, nano-modified UHPC features, and FRP reinforcement features can be expressed more accurately, thereby optimizing the prediction accuracy of residual strength values.

[0080] In summary, by decoupling the modeling of features before and after high-temperature exposure, and by fusing the secondary modeling of features after and before high-temperature exposure, the estimation accuracy of residual compressive strength of FRP-reinforced nano-modified UHPC under high-temperature conditions is significantly improved.

[0081] Based on the above design, this invention proposes a method for predicting the residual compressive strength of FRP-reinforced nano-modified UHPC after high-temperature exposure, and based on this method, proposes a prediction model for the residual compressive strength of FRP-reinforced nano-modified UHPC after high-temperature exposure.

[0082] Please also refer to Figure 1 , Figure 2 and Figure 3 , Figure 1A simple structure schematic diagram of a residual compressive strength prediction model of the FRP reinforced nano-modified UHPC after high temperature exposure, Figure 2 A simple flowchart schematic diagram of a residual compressive strength prediction method of the FRP reinforced nano-modified UHPC after high temperature exposure, Figure 3 A simple structure schematic diagram of a residual compressive strength prediction model.

[0083] The residual compressive strength prediction model of the FRP reinforced nano-modified UHPC after high temperature exposure comprises an UHPC preprocessing unit 1, a double-branch feature extraction unit 2, an UHPC compressive strength feature fusion unit 3, an UHPC high temperature feature extraction unit 4, a double-branch high temperature feature extraction unit 5, an UHPC residual compressive strength feature fusion unit 6 and a residual strength prediction unit 7.

[0084] The UHPC preprocessing unit 1 is used to perform step S1: pre-processing UHPC proportioning parameters and nano material parameters to obtain an UHPC feature vector.

[0085] Specifically, the UHPC proportioning parameters are used to characterize the initial strength of the UHPC, and the default parameters include the total amount of cementitious materials, the water-binder ratio, the sand-binder ratio and the steel fiber content;

[0086] The total amount of cementitious materials is used to represent the total amount of all cementitious materials in each cubic meter of UHPC, including cement, fine ground slag powder, fly ash and / or silica fume and other materials that can generate silicate gel by hydration, and the total amount of all cementitious materials reflects the basic performance of the UHPC and determines the relative proportion of cement and other ingredients, affecting the strength and durability of the UHPC and other physical properties.

[0087] It should be noted that different proportions of the total amount of cementitious materials will affect the basic performance of the UHPC, especially the mixing ratio of different types of cementitious materials (such as cement, slag powder, fly ash, etc.) will affect the durability, strength and other physical properties of the UHPC. Since the optional materials and proportions are numerous, the present application does not specifically limit all combination methods, but describes the conventional materials and common proportioning methods.

[0088] The water-binder ratio is used to represent the mass ratio of water quantity to total amount of cementitious materials per cubic meter of UHPC, thereby characterizing the flowability and basic strength of the concrete.

[0089] The sand-binder ratio is used to represent the mass ratio of sand quantity to total amount of cementitious materials per cubic meter of UHPC, to reflect the compactness and flowability of the UHPC.

[0090] It should be noted that different gradations of sand will affect the basic performance of UHPC, and since there are many gradation ratios of sand, the present application does not specifically limit all combination modes, but describes common gradation ratios.

[0091] The steel fiber content, which represents the total amount of steel fiber, is usually compared with the total mass of UHPC, and the content has a direct impact on the compressive strength, high temperature resistance and crack resistance of UHPC.

[0092] The nano material parameters include the type and content of nano materials;

[0093] The type of nano material includes nano silicon dioxide, nano titanium dioxide, etc., and different types of nano materials will have different effects on the performance of UHPC, such as crack resistance, compressive strength or thermal stability.

[0094] The content of nano material represents the total amount of nano material, which is usually compared with the total mass of cementitious material in the UHPC ratio parameters, and the content has a direct impact on the compressive strength, high temperature resistance and durability of UHPC, and too high or too low will affect the gain effect of nano material.

[0095] The pretreatment is to use principal component analysis (PCA) to reduce the dimensionality of the UHPC ratio parameters, and extract the principal components in the UHPC ratio parameters, that is, to retain the most influential features for the prediction results;

[0096] At the same time, the type of nano material in the nano material parameter is one-hot encoded, and the content of nano material is standardized, so as to obtain the standardized nano material parameter vector;

[0097] Finally, the principal components in the UHPC ratio parameters are combined with the standardized nano material parameter vector to generate the UHPC feature vector.

[0098] The principal component analysis can reduce data redundancy, remove information that is not important for prediction accuracy, and retain the most important feature information, which helps to improve the calculation efficiency and improve the accuracy of the model; the one-hot encoding converts the category data of nano materials into numerical data, avoiding unnecessary effects of category variables, and ensuring that each type of nano material can independently contribute to the model; the standardization is used to ensure that variables with different units and different scales can be compared in the same scale, so as to ensure the data stability in the training process or inference process, and avoid the influence of some features due to the difference in dimension.

[0099] Accordingly, by respectively preprocessing the UHPC proportioning parameter and the nano material parameter, and generating the corresponding feature vectors, the redundant features are reduced, the overfitting risk is reduced, and the input feature vectors have high learnability and physical interpretability, that is, they can be effectively used for training the model and can reflect the actual influence of the UHPC proportioning parameter and the nano material parameter on the performance of the UHPC. In addition, through preprocessing, the subsequent steps can accurately capture the complex interaction between the UHPC proportioning and the nano material on the compressive strength, especially the compressive strength after high temperature exposure, thereby improving the prediction accuracy and reliability.

[0100] The double-branch feature extraction unit 2 is configured to perform step S2: performing feature extraction on the UHPC feature vector and the FRP reinforcement parameter respectively to obtain UHPC features and FRP reinforcement features.

[0101] Specifically, the double-branch feature extraction unit 2 includes a UHPC feature extraction module and an FRP reinforcement feature extraction module.

[0102] The UHPC feature extraction module is configured to perform nonlinear feature extraction on the UHPC feature vector to obtain UHPC features.

[0103] The nonlinear feature extraction includes stacked fully connected layers and a nonlinear activation function, and the number of stacks is 3 by default. The nonlinear feature extraction is stacked to capture the nonlinear relationship between the UHPC proportioning and the nano material, thereby improving the prediction ability of the UHPC strength, and is specifically represented as follows:

[0104] Feature UHPC = (ReLU (FC (UHPC))) n n = 3

[0105] In the formula, Feature UHPC represents the UHPC features, which are used to represent the UHPC strength-related features obtained after multiple nonlinear feature extractions; UHPC represents the input UHPC feature vector; FC represents the fully connected layer, which extracts the deep information in the UHPC feature vector through multiple layers of mapping; ReLU is a nonlinear activation function, which is used to enhance the nonlinear expression ability of the model on the UHPC features; n represents the number of stacks, which enables the model to capture the deep nonlinear relationship between the UHPC proportioning parameter and the nano material parameter, and ensures that the model can fit complex UHPC performance influencing factors.

[0106] The FRP reinforcement feature extraction module is configured to perform preliminary feature extraction on the FRP reinforcement parameter to obtain FRP reinforcement features.

[0107] The FRP reinforcing parameter is used to represent a fiber reinforced composite material parameter for wrapping and reinforcing the UHPC, and by default includes the type of FRP and the number of FRP reinforcing layers; the type of FRP is processed by one-hot encoding to represent the reinforcing effect of different FRP types on the UHPC; and the number of FRP reinforcing layers is processed by standardization to reflect the strength of FRP reinforcement, and generally the more the number of layers, the higher the compressive strength of the UHPC after FRP reinforcement, but there is a marginal effect.

[0108] The preliminary feature extraction includes an embedding layer, a fully connected layer and a linear layer, and is specifically represented as follows:

[0109] Feature FRP =Linear(FC(EM(FRP)))

[0110] In the formula, Feature FRP represents an FRP reinforcing feature, which reflects the influence of the FRP type and the number of layers on the strength of the UHPC; FRP represents the input FRP reinforcing parameter; EM represents the embedding layer, which is used to convert the FRP type and the number of FRP reinforcing layers into a dense low-dimensional vector; and Linear represents the linear layer, which is used to perform linear mapping on the data processed by the embedding layer, and further improve the feature representation ability of the FRP reinforcing feature.

[0111] Accordingly, the present application performs independent feature extraction operations on the UHPC feature vector and the FRP reinforcing parameter through the double-branch network structure, so as to ensure that the model can perform targeted feature extraction on different types of data (such as numerical data of the nano-modified UHPC ratio and category data of the FRP), thereby improving the prediction ability of the model for the residual compressive strength of the FRP reinforced nano-modified UHPC.

[0112] The UHPC compressive strength feature fusion unit 3 is used to perform step S3: performing feature fusion on the UHPC feature and the FRP reinforcing feature to obtain the strength fusion feature of the UHPC.

[0113] Specifically, by performing a splicing operation on the UHPC feature and the FRP reinforcing feature, a comprehensive feature vector is generated, and a multilayer perceptron and a nonlinear activation function are used to perform feature fusion on the comprehensive feature vector to obtain the strength fusion feature of the UHPC, which is specifically represented as follows:

[0114] Feature fusion =ReLU(MLP(Concat(Feature UHPC ,Feature GRP )))

[0115] In the formula, Feature gusionFeature combined represents the strength fusion feature of the UHPC, which is used to represent the comprehensive characterization of the UHPC compression strength after the UHPC feature and the FRP reinforcement feature are fused; MLP is a multi-layer perception, which is used to represent the combination of multiple fully connected layers, and the present application defaults to two fully connected layers, and the higher order and nonlinear relationship of the feature are extracted by adjusting the dimension or feature size and performing feature extraction, so as to enhance the learning ability of the model to complex data patterns; and Concat represents a splicing operation.

[0116] Accordingly, in the feature fusion, the information of the UHPC feature and the FRP reinforcement feature is reserved by the splicing operation, so that the effective integration of the UHPC and the FRP reinforcement feature information is ensured; at the same time, through the combination of MLP and ReLU, the model can learn the nonlinear interaction relationship between the UHPC feature and the FRP reinforcement feature, and thus the prediction accuracy of the UHPC compression strength is improved.

[0117] The high-temperature feature extraction unit 4 of the UHPC is used to perform step S4: deep feature extraction on the strength fusion feature of the UHPC and the high-temperature exposure parameter to obtain the high-temperature feature of the UHPC.

[0118] Specifically, the high-temperature feature extraction unit 4 of the UHPC includes a temperature feature fusion module and a deep feature extraction module.

[0119] The temperature feature fusion module is used to splice the UHPC fusion feature and the high-temperature exposure parameter to generate the UHPC splicing feature after high-temperature exposure, which is specifically represented as follows:

[0120] Feature combined =Concat(Feature fusion ,T,t)

[0121] In the formula, Feature combined represents the UHPC splicing feature after high-temperature exposure; T and t are the temperature parameter and the exposure time parameter in the normalized high-temperature exposure parameter respectively.

[0122] The deep feature extraction module is used to sequentially perform feature extraction, self-attention (Self-Attention) weighting and nonlinear activation on the UHPC splicing feature after high-temperature exposure to generate the high-temperature feature of the UHPC, which is specifically represented as follows:

[0123] Feature temp =ReLU(Attention(MLP(Feature combined )))

[0124] In the formula, Feature tempAttention is a self-attention mechanism for weighting each position in the input features to capture more detailed and valuable information, so that the model can effectively identify which features (such as temperature and exposure time) have a more significant impact on the compressive strength of UHPC, thereby enhancing the model's prediction ability for the residual strength under high-temperature exposure.

[0125] Therefore, the temperature feature fusion module and the deep feature extraction module help the model to accurately learn the influence of temperature and exposure time on the strength of FRP reinforced nano-modified UHPC, and the self-attention mechanism enables the model to identify key features under different high-temperature exposure parameters, thereby effectively capturing the nonlinear relationship between high-temperature exposure parameters and the strength of FRP reinforced nano-modified UHPC, and significantly improving the prediction accuracy of the residual strength of FRP reinforced nano-modified UHPC after high-temperature exposure.

[0126] The double-branch high-temperature feature extraction unit 5 is configured to perform step S5: performing residual connection of the high-temperature features of UHPC with UHPC features and FRP reinforcement features respectively to obtain residual compressive features of high-temperature UHPC and high-temperature FRP.

[0127] Specifically, the double-branch high-temperature feature extraction unit 5 includes a UHPC feature residual module and an FRP reinforcement feature residual module.

[0128] The UHPC feature residual module is configured to perform residual connection (SkipConnection) of the high-temperature features of UHPC with UHPC features to generate residual compressive features of high-temperature UHPC, and is specifically expressed as follows:

[0129] Feature tempUHPC =Feature temp +Feature UHPC

[0130] In the formula, Feature tempUHPC represents the residual compressive features of high-temperature UHPC; and + represents residual connection.

[0131] The FRP reinforcement feature residual module is configured to perform residual connection of the high-temperature features of UHPC with FRP reinforcement features to generate residual compressive features of high-temperature FRP, and is specifically expressed as follows:

[0132] Feature tempFRP =Feature temp +Feature FRP

[0133] In the formula, Feature tempFRP represents the residual compressive features of high-temperature FRP.

[0134] Accordingly, the application ensures different high-temperature UHPC residual compressive strength features and high-temperature FRP residual compressive strength features by connecting the high-temperature features of UHPC with the original nano-modified UHPC features and the original FRP reinforcement features respectively, thereby effectively enhancing the feature learning capability, significantly improving the information flow, and ensuring that the model can consider the strength changes of UHPC and FRP reinforcement at the same time.

[0135] It should be noted that the aforementioned MLP (double-layer fully connected layer) unifies the dimensions of the spliced features, thereby ensuring that the feature vectors after residual connection have uniform dimensions and avoiding the situation that the dimensions of the spliced feature vectors do not match, guaranteeing the stability and effectiveness of the model in the training process or inference process.

[0136] The residual compressive strength feature fusion unit 6 is configured to perform step S6: feature fusion of the residual compressive strength features of high-temperature UHPC and high-temperature FRP to obtain comprehensive features of residual compressive strength.

[0137] Specifically, the residual compressive strength features of high-temperature UHPC and high-temperature FRP are sequentially subjected to splicing, multi-layer perception, self-attention, and nonlinear activation function for feature fusion to obtain comprehensive features of residual compressive strength, which are specifically represented as follows:

[0138] Feature last =ReLU(Attention(MLP(Concat(Feature tempUHPC ,Feature tempFRP ))))

[0139] In the formula, Feature last represents the comprehensive features of residual compressive strength, which fuse the residual compressive strength features of high-temperature UHPC and high-temperature FRP and combine the self-attention mechanism and ReLU activation function in the fusion process to enhance the key features related to the influence of UHPC and FRP on residual compressive strength under high-temperature exposure.

[0140] Accordingly, the application further weights the features after preliminary fusion by MLP by using the self-attention mechanism to improve the expression ability of the features and enhance the sensitivity of the model to high-temperature exposure, thereby more accurately capturing the influence of nano-modified UHPC and FRP reinforcement features on residual strength under high-temperature exposure, and effectively improving the prediction accuracy of the residual strength of FRP reinforced nano-modified UHPC after high-temperature exposure.

[0141] The residual strength prediction unit 7 is configured to perform step S7: predicting the comprehensive features of residual compressive strength to obtain the residual strength value.

[0142] Specifically, a linear output layer is used to predict the comprehensive feature of the residual compressive strength, and a residual strength value is obtained, which is specifically represented as follows:

[0143] Result=Linear(Feature last )

[0144] In the formula, Result represents the residual strength value.

[0145] According to the present application, by using a linear output layer, the fusion result of the model after nonlinear feature extraction is mapped to a continuous numerical output, thereby ensuring the interpretability of the residual strength value and the accuracy of the prediction.

[0146] In view of the high cost and long cycle of collecting experimental data of the high-temperature residual compressive strength of FRP reinforced nano-modified UHPC, resulting in a relatively scarce amount of data available for training, in order to effectively improve the training efficiency and prediction accuracy of the residual compressive strength prediction model of FRP reinforced nano-modified UHPC after high-temperature exposure, the present application proposes a training method based on phased heuristic optimization to train the residual compressive strength prediction model of FRP reinforced nano-modified UHPC after high-temperature exposure, which specifically includes the following steps:

[0147] T1, collecting a compressive strength data set and a residual compressive strength data set of FRP reinforced nano-modified UHPC.

[0148] Specifically, the compressive strength data set of FRP reinforced nano-modified UHPC includes UHPC proportioning parameters, nano material parameters, FRP reinforcement parameters and corresponding actual compressive strength values.

[0149] The compressive strength value generally refers to the maximum bearing capacity of a cubic meter of FRP reinforced nano-modified UHPC before failure or damage under axial compression, which can be obtained by testing the cubic meter of FRP reinforced nano-modified UHPC using a concrete compression testing machine. The cubic meter of FRP reinforced nano-modified UHPC is specifically constructed by configuring different UHPC proportioning parameters and nano material parameters to construct nano-modified UHPC, and using FRP reinforcement parameters to reinforce the nano-modified UHPC to obtain FRP reinforced nano-modified UHPC.

[0150] It should be noted that the compressive strength value can also be obtained by non-destructive testing methods such as rebound method and ultrasonic method, and the specific measurement method can be selected according to the accuracy requirements of the experiment, and the present application does not specifically limit it.

[0151] Then, the compressive performance test of the FRP reinforced nano-modified UHPC is carried out, and the compressive strength is obtained.

[0152] a residual compressive strength dataset of the FRP reinforced nano-modified UHPC, comprising the FRP reinforced nano-modified UHPC, a high-temperature exposure parameter, and a corresponding real residual strength value.

[0153] The residual strength value is obtained by testing the compressive performance of the FRP reinforced nano-modified UHPC after high-temperature treatment according to the high-temperature exposure parameter, and obtaining the residual compressive strength of the FRP reinforced nano-modified UHPC after the treatment.

[0154] The high-temperature treatment is to perform high-temperature treatment on the FRP reinforced nano-modified UHPC according to a preset temperature and corresponding duration, and wait for it to cool naturally to room temperature to obtain the treated FRP reinforced nano-modified UHPC.

[0155] T2, dividing the compressive strength dataset and the residual compressive strength dataset to generate training sets and test sets corresponding to the compressive strength and the residual compressive strength, respectively.

[0156] Specifically, the compressive strength dataset is divided into a training set and a test set according to an 8:2 ratio; at the same time, the residual compressive strength dataset is divided into a training set and a test set according to an 8:2 ratio.

[0157] T3, using a particle swarm optimization (PSO) heuristic algorithm and the test set of the compressive strength to optimize the weights of the double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3, and obtain the initialized double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3.

[0158] Specifically, the PSO heuristic algorithm is used to initialize the trainable parameters of the model, which specifically includes the following steps:

[0159] T31, initializing the particles of PSO to obtain an initial population.

[0160] The population is composed of a plurality of particles, and the particles include positions and corresponding speeds. The positions are used to represent a combination of a set of neural network trainable parameters corresponding to the current particle, that is, a numerical set of all weight matrices and bias matrices of the double-branch feature extraction unit 2 and the UHPC compressive strength fusion unit 3, representing the solution of the current particle in the search space.

[0161] The speed is used to represent the direction and step size of the particle moving in the parameter search space, and is used to update the particle position in the next round of iteration.

[0162] The initialization means that the positions and velocities of the particles are initialized in a uniform or normal random distribution in the search space, so as to ensure that the search space is fully explored and local optimal solution is avoided.

[0163] T32, according to the mean square error loss function and the test set of compressive strength, the fitness of the current population is calculated to obtain the fitness corresponding to the current population.

[0164] The mean square error loss (MSE) function is specifically represented as follows:

[0165]

[0166] In the formula, N represents the number of the test set of compressive strength; MSE (j) represents the fitness corresponding to the jth particle; is the compressive strength value obtained by inputting the i th sample into the double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3 corresponding to the trainable parameters of the jth particle for feature extraction and feature fusion, and then inputting the fused features into a linear layer for forward propagation prediction; represents the real compressive strength value corresponding to the i th sample corresponding to the test set of compressive strength.

[0167] T33, according to the fitness corresponding to all particles in the current population, the state of the current population is updated to obtain the updated population.

[0168] Specifically, the state update is used to guide the particles to iteratively update in the optimal direction in the parameter space through the trade-off between the local optimal solution and the global optimal solution of the particles in the current search space, so as to find a better neural network weight combination. The specific update expression of any one particle i is as follows:

[0169] v i (t+1)=ωv i (t)+c1r1(p i -x i (t))+c2r2(g-x i (t))

[0170] x i (t+1)=x i (t)+v i (t+1)

[0171] In the formula, v i (t) represents the speed of the i th particle at the t th iteration; x i(t) represents the position of the i-th particle at the t-th iteration, that is, the current trainable parameter; ω represents an inertia weight for controlling the change in the speed of the particle in advancing or searching; c1 and c2 are learning factors, which are usually set to positive numbers, for controlling the ability of the particle curve to find the global optimal solution and the optimal solution of itself; r1 and r2 are random numbers for introducing randomness to improve the ability of the particle to jump out of the local optimal solution; p i represents the historical optimal position of the i-th particle in all iterations, that is, the position corresponding to the lowest fitness value that the i-th particle has ever obtained; g represents the position corresponding to the lowest fitness value obtained by all particles in all iterations.

[0172] It should be noted that, in order to facilitate the particle to correspond to the model output prediction value, an additional linear layer is used to regress and predict the fusion features of the double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3, and the prediction result is used for fitness calculation.

[0173] T34, judge whether the fitness corresponding to the current population meets an iteration condition: if not, continue to execute step T32; if yes, execute step T35.

[0174] The iteration condition includes a preset maximum iteration number or a preset minimum fitness. Since the convergence speed of the fitness of the particle swarm optimization algorithm is affected by various factors such as the specific network structure, the training sample size, and the loss function, the present application does not limit the specific values of the maximum iteration number or the minimum fitness, but adjusts them according to the actual training effect in experiments.

[0175] T35, set the numerical set of the weight matrix and the bias matrix corresponding to the particle obtaining the lowest fitness in the current population as the optimal trainable parameter, and substitute it into the double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3 to obtain the initialized double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3.

[0176] T4, input the training set and the test set of the compressive strength into the initialized double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3 for training, and freeze the weights of the trained trainable parameters to obtain the preliminary trained residual compressive strength prediction model of the FRP reinforced nano-modified UHPC.

[0177] Specifically, the step T4 further includes the following sub-steps:

[0178] T41, input the training samples of the training set of the compressive strength into the initialized double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3 for feature extraction and feature fusion to obtain the strength fusion features of the UHPC.

[0179] T42, input the strength fusion feature of the UHPC into a linear layer for regression prediction to obtain the current compressive strength value of the FRP reinforced nano-modified UHPC.

[0180] T43, using a mean square error loss function, loss calculation is performed on the current compressive strength value of the FRP reinforced nano-modified UHPC and the real compressive strength value corresponding to the training sample of the training set of the compressive strength, to obtain the current loss value.

[0181] T44, according to the current loss value, gradient calculation is performed on the trainable parameters of the initialized double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3, and the optimizer is used to update the trainable parameters, to obtain the updated double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3.

[0182] The gradient calculation is realized by a back propagation algorithm, which is used to propagate error information in reverse according to the loss value, so as to calculate the partial derivative of each layer trainable parameter to the loss, and obtain the gradient value; the optimizer can use the Adam optimizer, which can adaptively adjust the learning rate of each trainable parameter, to improve the convergence speed and stability of the training.

[0183] T45, using the test set of the compressive strength to determine whether the updated double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3 meet the preset training condition: if yes, the training is completed, and the weight of the corresponding trainable parameters of the double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3 is frozen, to obtain the residual compressive strength prediction model of the FRP reinforced nano-modified UHPC after preliminary training; if not, step T41 is continued.

[0184] Specifically, by inputting the samples in the compressive strength test set into the updated double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3 for feature extraction and feature fusion, and inputting the fused features into a linear layer for regression prediction, and error calculation is performed on the real compressive strength value corresponding to the sample, to evaluate the prediction performance of the model.

[0185] The training condition includes the maximum training round or the error threshold, which is used to set the stopping condition of the model training; when the model reaches the preset standard on the test set, the weight at this time is frozen, to ensure that the double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3 can stably output the features highly related to the compressive strength in the subsequent residual strength prediction task, thereby improving the generalization ability and prediction consistency of the model.

[0186] T5, the residual compressive strength prediction model of the preliminary trained FRP reinforced nano modified UHPC is secondarily weight optimized by adopting a light spectrum optimizer (LSO) heuristic algorithm, to obtain trainable parameters in the secondary initialized residual compressive strength prediction model of the FRP reinforced nano modified UHPC.

[0187] Specifically, the LSO heuristic algorithm is used to initialize the trainable parameters of the preliminary trained residual compressive strength prediction model of the FRP reinforced nano modified UHPC, which specifically includes the following steps:

[0188] T51, the parameters of the LSO heuristic algorithm are initialized to generate a plurality of light rays.

[0189] The light rays represent a combination of trainable parameters of a group of neural networks, that is, a numerical set of all weight matrices and bias matrices of the preliminary trained residual compressive strength prediction model of the FRP reinforced nano modified UHPC, used to represent solutions in the search space.

[0190] T52, according to the mean square error loss function and the test set of the residual compressive strength, the fitness of all current light rays is calculated to obtain the fitness corresponding to all current light rays.

[0191] Specifically, according to the trainable parameters corresponding to all light rays, the trainable parameters of the preliminary trained residual compressive strength prediction model of the FRP reinforced nano modified UHPC are initialized, and the samples of the test set of the residual compressive strength are used to regress and predict the residual strength value.

[0192] Then, the residual strength value and the true residual strength value are loss calculated by using the mean square error loss function, to obtain the fitness corresponding to the light rays.

[0193] T53, whether the fitness corresponding to all current light rays is close to the fitness of the current global optimal solution is judged: if yes, the current global optimal solution is updated, and step T54 is executed; if not, step T54 is directly executed.

[0194] T54, according to the fitness corresponding to all current light rays, the random state of all current light rays is updated to obtain the updated light rays.

[0195] Specifically, the random state update includes an exploration phase and a development phase which are alternately executed, that is, the current light ray is iterated to update the state by using the exploration phase, and the next iteration is updated by using the development phase. Conversely, when the current light ray is iterated to update the state by using the development phase, the next iteration is updated by using the exploration phase.

[0196] Wherein, the exploration phase is used to enhance the global search ability of the current light, and the specific calculation expression is as follows:

[0197]

[0198] In the formula, Indicates the candidate solution corresponding to the i-th light of the t+1th iteration, and is used to represent the solution vector and contains all trainable parameters; Indicates the internal refraction vector corresponding to the i-th light, which is calculated according to the incident light and the internal refraction vector, so as to control the refraction of the light in the search space, and the specific calculation expression is as follows:

[0199]

[0200] In the formula, Indicates the incident light, which is used to represent the direction of the current solution, and is obtained by normalizing the average value of the solution corresponding to all lights; Indicates the normal vector of internal refraction, which is used to describe the interaction of light and medium (such as water droplets), and is obtained by normalizing the solution randomly selected from the light of the tth iteration; k r Indicates the refractive index, which is in the refractive index interval of red spectrum to purple spectrum, and the specific calculation expression is as follows:

[0201] k r =k red +RV1(k violet -k red )

[0202] In the formula, k red Indicates the refractive index of red spectrum, which is 1.331 by default; k violet Indicates the refractive index of purple spectrum, which is 1.344 by default; RV k Indicates the kth uniformly generated random number, and the uniformly generated random number RV k for any k has a value interval of [0, 1].

[0203] Indicates the internal reflection vector of the i-th light, and the specific calculation expression is as follows:

[0204]

[0205] In the formula, Indicates the normal vector of internal reflection of the i-th light, which is used to reflect the direction of light reflection in the medium, and is obtained by normalizing the solution of the current light of the tth iteration.

[0206] Indicates the external refraction vector of the i-th light, and the specific calculation expression is as follows:

[0207]

[0208] wherein, is the normal vector of the external refraction of the ith light ray, used to guide the refraction direction of the light ray by the global optimal solution, obtained by normalizing the normal vector of the global optimal solution.

[0209] represents a proportional factor, used to control the step size of each update, which gradually decreases with the number of iterations, ensuring that the optimization process can both jump out of the local optimum and quickly converge to the global optimum, and its specific calculation is as follows:

[0210]

[0211] wherein, is the kth random number vector with a dimension of n and subject to a normal distribution, and n is the dimension of the candidate solution of the light ray, that is, consistent with the dimension of the trainable parameters; a represents an iteration attenuation factor, which decreases with the number of iterations, and its specific calculation is as follows:

[0212]

[0213] wherein, T max represents a preset maximum number of iterations.

[0214] GI represents an adaptive control factor, used to adjust the amplitude and direction of the update of the light ray, ensuring the global search capability of the exploration stage, and its specific calculation is as follows:

[0215] GI = a x r -1 x P -1 (a, 1)

[0216] wherein, r -1 represents an inverse proportional scaling factor, used to amplify the jump range of the light ray in the exploration process, and r is a random number with a value range of [0, 1]; P -1 (a, 1) is the inverse function of the incomplete gamma function, used to introduce a control factor of nonlinear disturbance, which is usually used to model random behavior with skewed distribution, further improving the randomness and direction uncertainty of the jump.

[0217] and respectively represent the r1th solution, the r2th solution, the r3th solution and the r4th solution randomly selected from other light rays, used to determine the direction of the current light ray update, and promote the algorithm to explore new solution space; p and q respectively represent randomly generated numerical values, used to determine the update method of the current light ray.

[0218] wherein the development stage is used to enhance the local optimization ability of the light by local perturbation, so as to perform a slight update within the neighborhood of the current solution, so that the light can quickly converge to a more accurate optimal solution, which is specifically represented as follows:

[0219]

[0220] wherein R k represents the kth completely randomly generated random number, and for any kth completely randomly generated random number, its value interval is [0, 1]; represents a random unit vector containing 0 and 1 values, which is used as a random mask; F (i)′ represents the relative difference degree between the current ith light solution and the global optimal solution, which is used to help the algorithm make local fine-tuning when the fitness is close to the global optimum, otherwise make more extensive exploration, and its specific calculation is represented as follows:

[0221]

[0222] wherein F (i) represents the fitness of the current ith light; F B the fitness corresponding to the current global optimal solution; F W the fitness corresponding to the current global worst solution.

[0223] temp represents a quadratic update judgment expression, which is specifically represented as follows:

[0224]

[0225] wherein I(R<β) is an indicator function, which outputs 1 when R<β is true, otherwise outputs 0, and β is a preset constant, taking a value of [0, 1]; represents a cosine contraction term, which is used to control the amplitude of the update by cosine attenuation, so that the attraction between the current solution and the global optimal solution changes with the number of iterations, and r1 represents a randomly generated value in [0, 1], which is used to adjust the cosine contraction term; P s and P e are both preset probabilities, which are respectively used to control whether to execute the quadratic update judgment expression or the cosine contraction.

[0226] T55, judge whether the updated light meets an iteration threshold: if not, continue to execute step T52; if yes, take the current global optimal solution as a trainable parameter in the residual compressive strength prediction model of the secondary initialized FRP enhanced nano-modified UHPC, and complete the secondary weight optimization.

[0227] T6, training the trainable parameters in the residual compressive strength prediction model of the twice-initialized FRP-reinforced nano-modified UHPC according to the training set and the test set of the residual compressive strength until the trainable parameters meet the expected performance, then completing the training and obtaining the high-accuracy residual compressive strength prediction model of the FRP-reinforced nano-modified UHPC.

[0228] Specifically, the expected performance includes that the loss value on the training set and the test set reaches a preset threshold, or the training reaches a set maximum number of rounds, ensuring the stability and generalization ability of the model.

[0229] It should be noted that the training process of step T6 is substantially consistent with step T4, and the difference lies in that step T4 trains the double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3 using the compressive strength data set of the FRP-reinforced nano-modified UHPC, while step T6 trains the overall FRP-reinforced nano-modified UHPC residual compressive strength prediction model using the residual compressive strength data set of the FRP-reinforced nano-modified UHPC. Therefore, the present application will not repeat the T6 training process here.

[0230] Compared with the prior art, by introducing the high-temperature exposure parameter, the present application enhances the original FRP reinforcement features and the nano-modified UHPC features respectively through residual connection, ensures that the model can effectively capture the nonlinear change relationship between the compressive strength features of the FRP-reinforced nano-modified UHPC from normal temperature to the compressive strength features after high-temperature exposure, thereby significantly improving the modeling ability of the high-temperature influence and ensuring that the prediction of the compressive strength of the FRP-reinforced nano-modified UHPC under high-temperature exposure is more accurate.

[0231] In addition, by using the segmented training strategy and the PSO heuristic algorithm to initialize and train the feature extractor without high-temperature interference, the model can accurately learn the basic feature extraction ability of the FRP-reinforced nano-modified UHPC at normal temperature; then, the trained part of the parameters is frozen, and the LSO heuristic algorithm is introduced to further initialize and train the overall model, thereby effectively avoiding the training instability under high-temperature interference, improving the prediction ability of the model for the residual compressive strength of the FRP-reinforced nano-modified UHPC under high-temperature exposure, and finally significantly improving the residual strength prediction accuracy of the FRP-reinforced nano-modified UHPC after high-temperature exposure.

[0232] Based on the same inventive concept, the present application also provides an electronic device, which can be a server, a desktop computing device or a mobile computing device (e.g., a laptop computer, a handheld computing device, a tablet computer, a netbook, etc.) or a terminal device. The device includes one or more processors and a memory, wherein the processor is configured to execute a program to implement the method for predicting the residual compressive strength of the FRP reinforced nano-modified UHPC after high temperature exposure according to the embodiments of the present application; and the memory is configured to store the computer program executable by the processor.

[0233] Based on the same inventive concept, the present application also provides a computer readable storage medium, which corresponds to the method for predicting the residual compressive strength of the FRP reinforced nano-modified UHPC after high temperature exposure according to the above-mentioned embodiments of the present application. The computer readable storage medium has a computer program stored thereon, which, when executed by a processor, implements the steps of the method for predicting the residual compressive strength of the FRP reinforced nano-modified UHPC after high temperature exposure according to any of the above-mentioned embodiments.

[0234] The present application can be in the form of a computer program product implemented on one or more storage media (including but not limited to a disk memory, a CD-ROM, an optical memory, etc.) containing program code. The computer usable storage media include permanent and non-permanent, removable and non-removable media, which can be implemented by any method or technology to store information. The information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage device, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0235] The above-mentioned embodiments only express several embodiments of the present application, which are described in detail and specifically, but should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the inventive concept, several modifications and improvements can be made, and the present application also intends to include these modifications and improvements.

Claims

1. A method for predicting the residual compressive strength of FRP-reinforced nanomodified UHPC after high temperature exposure, characterized in that, The method comprises the following steps: S1: preprocessing UHPC proportioning parameters and nano material parameters to obtain UHPC feature vectors; S2: feature extraction is performed on the UHPC feature vectors and FRP reinforcement parameters respectively to obtain UHPC features and FRP reinforcement features; S3: feature fusion is performed on the UHPC features and FRP reinforcement features to obtain strength fusion features of the UHPC; S4: deep feature extraction is performed on the strength fusion features of the UHPC and high-temperature exposure parameters to obtain high-temperature features of the UHPC; S5: residual connection is performed on the high-temperature features of the UHPC and the UHPC features and the FRP reinforcement features respectively to obtain residual compressive strength features of high-temperature UHPC and high-temperature FRP; S6: feature fusion is performed on the residual compressive strength features of the high-temperature UHPC and the high-temperature FRP to obtain comprehensive features of the residual compressive strength; S7: prediction is performed on the comprehensive features of the residual compressive strength to obtain a residual strength value.

2. The method of predicting residual compressive strength of FRP-reinforced nanomodified UHPC according to claim 1, wherein The feature extraction of the step S2 is performed by a UHPC feature extraction module and an FRP reinforcement feature extraction module, The UHPC feature extraction module is configured to perform nonlinear feature extraction on the UHPC feature vectors to obtain UHPC features, wherein the nonlinear feature extraction comprises a stacked fully connected layer and a nonlinear activation function, and is specifically represented as follows: Feature UHPC = (ReLU (FC (UHPC) ) ) n In the formula, Feature UHPC represents the UHPC feature; UHPC represents the input UHPC feature vector; FC represents the fully connected layer; ReLU is a nonlinear activation function; n represents the number of stacks; The FRP reinforcement feature extraction module is configured to perform preliminary feature extraction on the FRP reinforcement parameters to obtain FRP reinforcement features, wherein the preliminary feature extraction comprises an embedding layer, a fully connected layer and a linear layer, and is specifically represented as follows: Feature FRP = Linear(FC(EM(FRP))) In the formula, Feature fRP represents the FRP reinforcement feature; FRP represents the input FRP reinforcement parameter; EM represents the embedding layer; Linear represents the linear layer; The step S3 comprises the following sub-steps: Splicing operation is performed on the UHPC features and the FRP reinforcement features to generate comprehensive feature vectors, and multi-layer perception and nonlinear activation function are used to perform feature fusion on the comprehensive feature vectors to obtain strength fusion features of the UHPC, which are specifically represented as follows: Feature fusion = ReLU(MLP(Concat(Feature UHPC ,Feature FRP ))) In the formula, Feature fusion represents the strength fusion feature of UHPC; MLP is a multi-layer perception, representing a combination of multiple fully connected layers; and Concat represents a concatenation operation. The deep feature extraction of the step S4 is performed by a temperature feature fusion module and a deep feature extraction module; The temperature feature fusion module is configured to splice the UHPC fusion features and the high-temperature exposure parameters to generate UHPC splicing features after high-temperature exposure, which are specifically represented as follows: Feature combined = Concat(Feature fusion , T, t) where Feature combined represents the splicing feature of UHPC after high temperature exposure; T and t are the temperature parameter and the exposure time parameter in the normalized high temperature exposure parameter, respectively; The deep feature extraction module is configured to perform feature extraction, self-attention weighting and nonlinear activation on the UHPC splicing features after high-temperature exposure in sequence to generate high-temperature features of the UHPC, which are specifically represented as follows: Feature temp = ReLU(Attention(MLP(Feature combined ))) In the formula, Feature temp is the high temperature feature of UHPC; Attention is a self-attention mechanism; The step S6 comprises the following sub-steps: The residual compressive strength features of the high-temperature UHPC and the high-temperature FRP are sequentially subjected to splicing, multi-layer perception, self-attention and nonlinear activation function for feature fusion to obtain comprehensive features of the residual compressive strength, which are specifically represented as follows: Feature last = ReLU(Attention(MLP(Concat(Feature tempUHPC , Feature tempFRP )))) where Feature last represents a comprehensive feature of residual compressive strength; Feature tempUHPC represents a residual compressive feature of high-temperature UHPC; Feature tempFRP represents a residual compressive feature of high-temperature FRP.

3. The method of predicting residual compressive strength of FRP reinforced nanomodified UHPC according to claim 2, characterized in that, A training method based on a phased heuristic optimization is used to train all trainable parameters in the residual compressive strength prediction method of the FRP reinforced nano modified UHPC, which specifically comprises the following steps: T1: collecting a compressive strength data set and a residual compressive strength data set of the FRP reinforced nano modified UHPC; T2, divide the compressive strength data set and the residual compressive strength data set to generate a training set and a test set corresponding to the compressive strength and the residual compressive strength respectively; T3, using the PSO heuristic algorithm and the test set of the compressive strength, the weight of the trainable parameters of the UHPC feature extraction module and the FRP reinforcement feature extraction module in step S2 and the multi-layer perceptron in step S3 is optimized to obtain the initialized UHPC feature extraction module, the FRP reinforcement feature extraction module and the multi-layer perceptron; T4, input the training set and the test set of the compressive strength into the initialized UHPC feature extraction module, the FRP reinforcement feature extraction module and the multi-layer perceptron for training, and freeze the weight to obtain the trainable parameters of the preliminary trained residual compressive strength prediction method of the FRP enhanced nano modified UHPC; T5, using the LSO heuristic algorithm, the trainable parameters of the preliminary trained residual compressive strength prediction method of the FRP enhanced nano modified UHPC are secondarily optimized to generate the trainable parameters of the second initialized residual compressive strength prediction method of the FRP enhanced nano modified UHPC; T6, according to the training set and the test set of the residual compressive strength, the trainable parameters of the second initialized residual compressive strength prediction method of the FRP enhanced nano modified UHPC are trained until the trainable parameters meet the expected performance, and the training is completed.

4. The method of predicting residual compressive strength of FRP-reinforced nanomodified UHPC according to claim 3, characterized in that, The PSO heuristic algorithm specifically includes the following steps: T31, initialize the particles of PSO to obtain an initial population; wherein the population is composed of a plurality of particles, the particles include positions and corresponding speeds, the positions are used to represent the combination of the trainable parameters corresponding to the current particles; the speed is used to represent the direction and step size of the particle moving in the parameter search space; T32, according to the mean square error loss function and the test set of the compressive strength, the fitness of the current population is calculated to obtain the fitness corresponding to the current population; T33, according to the fitness corresponding to all particles in the current population, the state of the current population is updated to obtain the updated population; T34, judge whether the fitness corresponding to the current population meets an iteration condition: if not, continue to execute step T32; if yes, execute step T35; T35, the trainable parameters corresponding to the particle obtaining the lowest fitness in the current population are taken as the optimal trainable parameters, which are substituted into the trainable parameters of the UHPC feature extraction module and the FRP reinforcement feature extraction module in step S2 and the multi-layer perceptron in step S3 to obtain the initialized UHPC feature extraction module and the FRP reinforcement feature extraction module in step S2 and the multi-layer perceptron in step S3; The LSO heuristic algorithm specifically includes the following steps: T51, initialize the parameters of the LSO heuristic algorithm to generate a plurality of rays; T52, according to the mean square error loss function and the test set of the residual compressive strength, the fitness of all the current rays is calculated to obtain the fitness corresponding to all the current rays; T53, judging whether the fitness corresponding to the current all light rays is close to the fitness of the current global optimal solution: if yes, updating the current global optimal solution, and executing step T54; if no, directly executing step T54; T54, performing random state update on the current all light rays according to the fitness corresponding to the current all light rays, to obtain updated light rays; T55, judging whether the updated light rays meet an iteration threshold: if no, continuing to execute step T52; if yes, taking the current global optimal solution as a trainable parameter in a secondary initialization FRP enhanced nano modified UHPC residual compressive strength prediction method, to complete secondary weight optimization.

5. A model for predicting the residual compressive strength of FRP-reinforced nanomodified UHPC after high temperature exposure, characterized in that, The UHPC pretreatment unit, the double-branch feature extraction unit, the UHPC compressive strength feature fusion unit, the UHPC high-temperature feature extraction unit, the double-branch high-temperature feature extraction unit, the UHPC residual compressive strength feature fusion unit, and the residual strength prediction unit are included. The UHPC pretreatment unit is configured to pretreat UHPC proportioning parameters and nano material parameters to obtain a UHPC feature vector. The double-branch feature extraction unit is configured to perform feature extraction on the UHPC feature vector and FRP reinforcement parameters respectively to obtain UHPC features and FRP reinforcement features. The UHPC compressive strength feature fusion unit is configured to perform feature fusion on the UHPC features and the FRP reinforcement features to obtain strength fusion features of the UHPC. The UHPC high-temperature feature extraction unit is configured to perform deep feature extraction on the strength fusion features of the UHPC and high-temperature exposure parameters to obtain high-temperature features of the UHPC. The double-branch high-temperature feature extraction unit is configured to perform residual connection on the high-temperature features of the UHPC and the UHPC features and the FRP reinforcement features respectively to obtain residual compressive features of high-temperature UHPC and high-temperature FRP. The UHPC residual compressive strength feature fusion unit is configured to perform feature fusion on the residual compressive features of the high-temperature UHPC and the high-temperature FRP to obtain comprehensive features of the residual compressive strength. The residual strength prediction unit is configured to predict the comprehensive features of the residual compressive strength to obtain a residual strength value. 6.The residual compressive strength prediction model of FRP-reinforced nanomodified UHPC according to claim 5, wherein, The double-branch feature extraction unit includes a UHPC feature extraction module and an FRP reinforcement feature extraction module. The UHPC feature extraction module is configured to perform nonlinear feature extraction on the UHPC feature vector to obtain UHPC features; wherein the nonlinear feature extraction includes stacked fully connected layers and a nonlinear activation function, and is specifically represented as follows: Feature UHPC = (ReLU (FC (UHPC) ) ) n where Feature UHPC represents the UHPC features; UHPC represents the input UHPC feature vector; FC represents a fully connected layer; ReLU is a nonlinear activation function; and n represents the number of stacks. The FRP reinforcement feature extraction module is configured to perform preliminary feature extraction on the FRP reinforcement parameters to obtain FRP reinforcement features; wherein the preliminary feature extraction includes an embedding layer, a fully connected layer, and a linear layer, and is specifically represented as follows: Feature FRP = Linear(FC(EM(FRP))) In the formula, Feature FRP represents the FRP reinforcement feature; FRP represents the input FRP reinforcement parameter; EM represents an embedding layer; Linear represents a linear layer; The UHPC compressive strength feature fusion unit is further configured to perform the following sub-steps: The UHPC characteristic and the FRP reinforcing characteristic are spliced to generate a comprehensive feature vector, and a multi-layer perception and a nonlinear activation function are used for feature fusion of the comprehensive feature vector to obtain a strength fusion feature of the UHPC, and the strength fusion feature of the UHPC is specifically expressed as follows: Feature fusion = ReLU(MLP(Concat(Feature UHPC , Feature FRP ))) where Feature fusion represents the strength blending feature of UHPC; MLP is a multi-layer perceptron representing a combination of multiple fully connected layers; and Concat represents a concatenation operation. The high-temperature feature extraction unit of the UHPC comprises a temperature feature fusion module and a deep feature extraction module; The temperature feature fusion module is configured to splice the UHPC fusion feature and the high-temperature exposure parameter to generate a UHPC splicing feature after high-temperature exposure, and the UHPC splicing feature after high-temperature exposure is specifically expressed as follows: Feature combined = Concat(Feature fusion , T, t) where Feature combined represents the splicing feature of UHPC after high temperature exposure; T and t are the temperature parameter and the exposure time parameter in the normalized high temperature exposure parameter, respectively; The deep feature extraction module is configured to sequentially perform feature extraction, self-attention weighting and nonlinear activation on the UHPC splicing feature after high-temperature exposure to generate a high-temperature feature of the UHPC, and the high-temperature feature of the UHPC is specifically expressed as follows: Feature temp = ReLU(Attention(MLP(Feature combined ))) In the formula, Feature temp is the high temperature feature of UHPC; Attention is a self-attention mechanism; The residual compressive strength feature fusion unit of the UHPC is further configured to perform the following sub-steps: The residual compressive strength features of the high-temperature UHPC and the high-temperature FRP are sequentially fused through splicing, multi-layer perception, self-attention and a nonlinear activation function to obtain a comprehensive feature of the residual compressive strength, and the comprehensive feature of the residual compressive strength is specifically expressed as follows: Feature last = ReLU(Attention(MLP(Concat(Feature tempUHPC , Feature tempFRP )))) where Feature last represents the compressive residual strength of the composite; Feature tempUHPC represents the compressive residual strength of the high-temperature UHPC; Feature tempFRP represents the compressive residual strength of the high-temperature FRP.

7. The residual compressive strength prediction model of FRP-reinforced nanomodified UHPC according to claim 6, characterized in that, The training method based on the staged heuristic optimization is used to train the residual compressive strength prediction model of the FRP reinforced nano-modified UHPC after high-temperature exposure, and the training method specifically comprises the following steps: T1, collect the compressive strength dataset and the residual compressive strength dataset of the FRP reinforced nano-modified UHPC; T2, divide the compressive strength dataset and the residual compressive strength dataset to generate training sets and test sets corresponding to the compressive strength and the residual compressive strength, respectively; T3, use the PSO heuristic algorithm and the test set of the compressive strength to optimize the weights of the double-branch feature extraction unit and the UHPC compressive strength feature fusion unit to obtain the initialized double-branch feature extraction unit and the UHPC compressive strength feature fusion unit; T4, input the training set and the test set of the compressive strength into the initialized double-branch feature extraction unit and the UHPC compressive strength feature fusion unit for training, and freeze the weights to obtain trainable parameters in the preliminary trained residual compressive strength prediction model of the FRP reinforced nano-modified UHPC; T5, use the LSO heuristic algorithm to perform secondary weight optimization on the preliminary trained residual compressive strength prediction model of the FRP reinforced nano-modified UHPC to generate trainable parameters in the secondary initialized residual compressive strength prediction model of the FRP reinforced nano-modified UHPC; T6, train the trainable parameters in the secondary initialized residual compressive strength prediction model of the FRP reinforced nano-modified UHPC according to the training set and the test set of the residual compressive strength until the trainable parameters meet the expected performance, and then the training is completed to obtain the high-accuracy residual compressive strength prediction model of the FRP reinforced nano-modified UHPC. 8.The residual compressive strength prediction model of FRP-reinforced nanomodified UHPC according to claim 7, wherein, The PSO heuristic algorithm specifically comprises the following steps: T31, initializing the particles of the PSO to obtain an initial population; wherein the population is composed of a plurality of particles, and each particle includes a position and a corresponding velocity, the position being used to represent a combination of the trainable parameters corresponding to the current particle; and the velocity being used to represent the direction and step size of the particle moving in the parameter search space; T32, performing fitness calculation on the current population according to the mean square error loss function and the test set of the compressive strength to obtain the fitness corresponding to the current population; T33, performing state updating on the current population according to the fitness corresponding to all particles in the current population to obtain an updated population; T34, judging whether the fitness corresponding to the current population meets an iteration condition: if not, continuing to perform step T32; and if yes, performing step T35; T35, taking the numerical set of the weight matrix and the bias matrix corresponding to the particle obtaining the lowest fitness in the current population as the optimal trainable parameters, and substituting them into the double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3 to obtain the initialized double-branch feature extraction unit 2 and the UHPC compressive strength feature fusion unit 3; The LSO heuristic algorithm specifically includes the following steps: T51, initializing the parameters of the LSO heuristic algorithm to generate a plurality of rays; T52, performing fitness calculation on all current rays according to the mean square error loss function and the test set of the residual compressive strength to obtain the fitness corresponding to all current rays; T53, judging whether the fitness corresponding to all current rays is close to the fitness of the current global optimal solution: if yes, updating the current global optimal solution and performing step T54; and if not, directly performing step T54; T54, performing random state updating on all current rays according to the fitness corresponding to all current rays to obtain updated rays; T55, judging whether the updated rays meet an iteration threshold: if not, continuing to perform step T52; and if yes, taking the current global optimal solution as the trainable parameters in the residual compressive strength prediction model of the secondary FRP reinforced nano-modified UHPC to complete secondary weight optimization.

9. An electronic device comprising: The memory, the processor, and the computer program stored in the memory and executable on the processor, characterized in that the processor implements the residual compressive strength prediction method of the FRP reinforced nano-modified UHPC after high temperature exposure according to any one of claims 1-4 when executing the computer program.

10. A computer-readable storage medium storing computer-executable instructions, characterized in that, The computer executable instructions are executed by the processor to implement the residual compressive strength prediction method of the FRP reinforced nano-modified UHPC after high temperature exposure according to any one of claims 1-4.

Images