Urban bridge updating construction parameter optimization method and system based on deep learning

Abstract

This invention discloses a method and system for optimizing construction parameters for urban bridges based on deep learning, belonging to the field of data processing technology. The method includes: acquiring measured values ​​of construction parameters and concrete performance to form a historical dataset; training a deep neural network to obtain a prediction model; calculating the prediction uncertainty of the prediction model for each combination of construction parameters, and filtering out parameter regions with high uncertainty; generating new construction parameters and simulated concrete performance values ​​based on the parameter update model within these regions, forming a synthetic sample set, and adding it to the historical dataset to retrain the prediction model; repeating the above steps until the prediction model is verified to converge in the entire construction parameter space, obtaining a converged prediction model; setting target performance requirements, using the converged prediction model as a fitness function evaluator, and searching for construction parameters that meet the target performance requirements through an optimization algorithm to obtain the optimal parameter combination. This invention achieves precise matching between urban bridge construction parameters and target concrete performance.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for optimizing construction parameters for urban bridge updates based on deep learning. Background Technology

[0002] In the construction of existing urban bridges, concrete performance is affected by a variety of construction parameters, exhibiting highly nonlinear coupling characteristics. Current construction parameter ratios largely rely on engineering experience and on-site trial mix tests. Some technologies employ shallow machine learning models to fit the mapping relationship between construction parameters and concrete performance, achieving preliminary predictions of indicators such as compressive strength and bond performance.

[0003] However, shallow models have limited ability to fit high-dimensional coupled construction parameters, have a large number of regions with high prediction uncertainty, and have poor generalization ability. The number of measured samples is insufficient and unevenly distributed, making it difficult to cover complex construction scenarios, resulting in large prediction errors for concrete performance. Construction parameter optimization is mainly based on manual trial and error, which cannot accurately match the preset target performance requirements. Problems such as insufficient bonding between new and old concrete and excessive internal and external temperature differences in hydration heat are prone to occur, resulting in poor control of construction quality and efficiency. Summary of the Invention

[0004] This invention provides a method and system for optimizing construction parameters of urban bridges based on deep learning. The purpose is to solve the technical problems in the prior art where the model has low accuracy in weak areas of the parameter space due to limited measured data, and it is difficult to reliably search for the optimal construction parameters.

[0005] In view of the above problems, the present invention provides a method and system for optimizing construction parameters for urban bridge updates based on deep learning.

[0006] In a first aspect, this invention provides a method for optimizing construction parameters for urban bridge updates based on deep learning, including: S1: Obtain multiple sets of urban bridge construction parameters and corresponding measured values ​​of concrete performance to form a historical dataset, and use the historical dataset to train a deep neural network to obtain a concrete performance prediction model. S2: In the construction parameter space, calculate the prediction uncertainty of the concrete performance prediction model for each combination of construction parameters, and screen out the parameter regions with high uncertainty. S3: Within the region of high uncertainty parameters, new urban bridge construction parameters and their corresponding concrete performance simulation values ​​are generated based on the parameter update model to form a synthetic sample set. S4: Add the synthetic sample set to the historical dataset and retrain the concrete performance prediction model to obtain an updated concrete performance prediction model. S5: Repeat steps S2 to S4 until the updated concrete performance prediction model is verified to converge in the entire construction parameter space, and the converged concrete performance prediction model is obtained. S6: Set target performance requirements, use the converged concrete performance prediction model as a fitness function evaluator, and use an optimization algorithm to search for urban bridge construction parameters that meet the target performance requirements to obtain the optimal combination of construction parameters.

[0007] Secondly, this invention provides a deep learning-based system for optimizing construction parameters for urban bridge updates, including: The initial model training module is used to obtain multiple sets of urban bridge construction parameters and corresponding measured values ​​of concrete performance to form a historical dataset, and to train a deep neural network using the historical dataset to obtain a concrete performance prediction model. The uncertainty assessment module is used to calculate the prediction uncertainty of the concrete performance prediction model for each combination of construction parameters in the construction parameter space, and to screen out the parameter regions with high uncertainty. The synthetic sample generation module is used to generate new urban bridge construction parameters and their corresponding concrete performance simulation values ​​based on the parameter update model within the high uncertainty parameter region, thus forming a synthetic sample set. The incremental model training module is used to add the synthetic sample set to the historical dataset and retrain the concrete performance prediction model to obtain an updated concrete performance prediction model. The iterative convergence judgment module is used to repeatedly call the uncertainty assessment module, the synthetic sample generation module, and the model incremental training module until the updated concrete performance prediction model is verified to converge in the entire construction parameter space, and the converged concrete performance prediction model is obtained. The optimal parameter search module is used to set target performance requirements, use the converged concrete performance prediction model as a fitness function evaluator, and search for urban bridge construction parameters that meet the target performance requirements through optimization algorithms to obtain the optimal combination of construction parameters.

[0008] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides a method and system for optimizing construction parameters for urban bridge renovation based on deep learning. First, a concrete performance prediction model is constructed using deep learning. Then, regions with high uncertainty parameters are located and synthetic samples are generated to iteratively optimize the model, effectively eliminating prediction blind spots and improving the full-space generalization accuracy of the concrete performance prediction model. Finally, guided by the target performance, an optimization algorithm is used to intelligently search for the optimal combination of construction parameters, eliminating the reliance on experience-based trial matching and achieving precise matching between urban bridge construction parameters and the target performance of concrete, thus ensuring the quality and efficiency of bridge renovation construction. Attached Figure Description

[0009] Figure 1 A flowchart illustrating the method for optimizing construction parameters for urban bridge updates based on deep learning, as provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the deep learning-based urban bridge construction parameter optimization system provided in an embodiment of the present invention. The components represented by each number in the attached diagram are explained below: The module includes: initial model training module 11, uncertainty assessment module 12, synthetic sample generation module 13, incremental model training module 14, iterative convergence judgment module 15, and optimal parameter search module 16. Detailed Implementation

[0010] This invention provides a method and system for optimizing construction parameters of urban bridges based on deep learning, which addresses the technical problem in existing technologies where limited measured data leads to low accuracy of models in weak areas of the parameter space and difficulty in reliably searching for optimal construction parameters.

[0011] Example 1, as Figure 1 As shown, this invention provides a method for optimizing construction parameters for urban bridge updates based on deep learning, the method comprising: S1: Obtain multiple sets of urban bridge construction parameters and corresponding measured values ​​of concrete performance to form a historical dataset, and use the historical dataset to train a deep neural network to obtain a concrete performance prediction model.

[0012] In this embodiment of the invention, multiple sets of urban bridge construction parameters and corresponding measured values ​​of concrete performance are obtained to form a historical dataset. A deep neural network is then trained using this historical dataset to obtain a concrete performance prediction model. During urban bridge construction, the workability, mechanical properties, and interfacial bonding properties of concrete are influenced by various construction parameters, including water-cement ratio, curing conditions, and interface treatment. A highly nonlinear coupling relationship exists between these parameters and their performance. Traditional empirical formulas or linear regression models struggle to accurately characterize this high-dimensional nonlinear relationship, leading to significant prediction bias. To subsequently achieve uncertainty assessment, model iterative optimization, and optimal parameter search in the construction parameter space, a standardized dataset must first be established, and a deep neural network adapted to high-dimensional nonlinear characteristics must be built and trained to obtain an initial concrete performance prediction model.

[0013] Step S1 in the method provided in this embodiment of the invention includes: This involves acquiring multiple sets of urban bridge construction parameters and corresponding measured values ​​of concrete performance to form a historical dataset, including: Multiple sets of construction parameters were collected during the urban bridge renovation construction process. These construction parameters included at least the water-cement ratio, sand ratio, admixture dosage, water-reducing agent dosage, pouring temperature, curing humidity, surface roughness grade of old concrete, and interface agent type. For each set of construction parameters, the corresponding measured values ​​of concrete performance are obtained through on-site experiments or quality inspection reports. The measured values ​​of concrete performance include at least 3-day compressive strength, 28-day compressive strength, bond strength between new and old concrete, slump, heat of hydration, internal and external temperature difference, and shrinkage strain. Each set of construction parameters is paired with the corresponding measured values ​​of concrete performance to form a data sample. All data samples are combined to form a historical dataset.

[0014] First, multiple sets of construction parameters were collected during the urban bridge renovation construction process. These parameters included at least the water-cement ratio, sand ratio, admixture dosage, water-reducing agent dosage, pouring temperature, curing humidity, surface roughness grade of the old concrete, and interface agent type. These construction parameters are key controllable variables that directly determine the quality of concrete forming and the bonding effect between the old and new concrete. The selected water-cement ratio, sand ratio, admixture dosage, and water-reducing agent dosage determine the concrete foundation mix proportions and workability; pouring temperature and curing humidity affect concrete hydration and strength development; and the surface roughness grade of the old concrete and the interface agent type directly determine the bonding effect between the old and new concrete. These parameters are essential control parameters for bridge renovation concrete construction.

[0015] For example, data was collected for the reinforcement and renovation project of an elevated bridge in a certain city. A total of 200 sets of construction parameters were collected. One set of construction parameters was as follows: water-cement ratio 0.42, sand ratio 36%, fly ash admixture dosage 15%, polycarboxylate superplasticizer dosage 1.0%, pouring temperature 25℃, curing humidity 90%, old concrete surface roughness grade II, and interface agent is water-based epoxy type.

[0016] Secondly, for each set of construction parameters, corresponding measured values ​​of concrete performance are obtained through on-site experiments or quality inspection reports. These measured values ​​include at least 3-day compressive strength, 28-day compressive strength, bond strength between new and old concrete, slump, heat of hydration, internal and external temperature difference, and shrinkage strain. These measured concrete performance values ​​are actual indicators reflecting the workability, mechanical properties, interfacial properties, and deformation properties of concrete. They serve as a true indicator for training the concrete performance prediction model and directly reflect the actual construction effect of the combination of construction parameters. For each set of construction parameters, corresponding measured values ​​of 3-day compressive strength, 28-day compressive strength, bond strength between new and old concrete, slump, heat of hydration, internal and external temperature difference, and shrinkage strain are obtained through on-site compressive strength tests, bond strength tests, and slump measurements.

[0017] For example, corresponding to the above construction parameters, the measured values ​​of concrete performance were obtained as follows: 3-day compressive strength 26.5 MPa, 28-day compressive strength 45.5 MPa, bond strength between new and old concrete 2.0 MPa, slump 170 mm, heat of hydration with internal and external temperature difference 10 °C, and shrinkage strain 170 × 10⁻⁶. -6 .

[0018] Next, each set of construction parameters is paired with its corresponding measured concrete performance value to form a data sample. All data samples are combined to form a historical dataset. The historical dataset is a collection of samples composed of single sets of construction parameters paired with their corresponding measured concrete performance values, providing the foundational data for supervised training of the subsequent deep neural network. Each set of construction parameters is paired with its corresponding measured concrete performance value to form an independent data sample; multiple sets of data samples are then aggregated to form a complete historical dataset. For example, pairing the above construction parameters with their performance values ​​results in one data sample. Each data sample contains both the construction parameter and its corresponding measured concrete performance value. The remaining sets are processed similarly, and 200 paired data samples together form the historical dataset.

[0019] The concrete performance prediction model is obtained by training a deep neural network using the historical dataset, including: The construction parameters and measured concrete performance values ​​in the historical dataset are normalized to obtain normalized construction parameters and normalized measured concrete performance values. Construct a multi-layer feedforward deep neural network, which includes an input layer, multiple hidden layers and an output layer, with each hidden layer followed by a batch normalization layer and a dropout layer; Normalized construction parameters are used as input features, and normalized measured concrete performance values ​​are used as output labels. The multilayer feedforward deep neural network is trained in a supervised manner until it is verified to converge, thus obtaining a trained deep neural network. The trained deep neural network is then used as a concrete performance prediction model.

[0020] First, the construction parameters and measured concrete performance values ​​in the historical dataset are normalized to obtain normalized construction parameters and normalized measured concrete performance values. Normalization is used to eliminate differences in the dimensions and numerical ranges of different parameters. Min-Max normalization is employed, with the formula: Normalized value = (Original value - Minimum value of the parameter sample) / (Maximum value of the parameter sample - Minimum value of the parameter sample), mapping all data to the interval [0,1]. The maximum and minimum values ​​of the parameters are statistically obtained from all samples of the corresponding parameters in the historical dataset.

[0021] For example, taking the water-gel ratio as an example, the maximum water-gel ratio in 200 samples is 0.50 and the minimum is 0.38. The original water-gel ratio of this group is 0.42. After normalization, the value is (0.42-0.38) / (0.50-0.38)=0.33, thus completing the water-gel ratio normalization calculation. The other parameters are calculated in the same way.

[0022] Secondly, a multi-layer feedforward deep neural network is constructed, comprising an input layer, multiple hidden layers, and an output layer. Each hidden layer is followed by a batch normalization layer and a dropout layer. The multi-layer feedforward deep neural network is a deep learning network structure with unidirectional signal propagation. The batch normalization layer is used to standardize the hidden layer output, accelerating training. The dropout layer prevents overfitting by randomly disabling some neurons.

[0023] Specifically, a multi-layer feedforward deep neural network is constructed: the input layer has 8 neurons, corresponding to 8 types of construction parameters, including water-cement ratio, sand ratio, admixture dosage, water-reducing agent dosage, pouring temperature, curing humidity, surface roughness grade of old concrete, and interface agent type; there are 3 hidden layers with 64, 32, and 16 neurons respectively, and the ReLU function is used for the activation function of the hidden layers; each hidden layer is followed by a batch normalization layer and a discard layer with a discard rate of 0.2; the output layer has 6 neurons, corresponding to 6 types of concrete performance indicators, including 3-day compressive strength, 28-day compressive strength, bond strength between new and old concrete, slump, heat of hydration, internal and external temperature difference, and shrinkage strain, and the linear activation function is used for the activation function of the output layer.

[0024] Finally, the normalized construction parameters are used as input features, and the normalized measured values ​​of concrete performance are used as output labels to perform supervised training on the multilayer feedforward deep neural network until convergence is verified, resulting in a trained deep neural network. This trained deep neural network is then used as the concrete performance prediction model. Supervised training utilizes labeled sample data to continuously optimize network weights, making the neural network model output approximate the true labels. The concrete performance prediction model is a deep neural network model that, after training convergence, can directly output predicted concrete performance based on the current construction parameters.

[0025] Specifically, normalized construction parameters were used as inputs, and normalized measured performance values ​​were used as labels. The Adam optimizer was employed, with a batch size of 32, and mean squared error (MSE) was used as the loss function. The convergence condition was that the loss on the validation set decreased by less than 1 × 10⁻⁶ for 20 consecutive iterations. -4 Training stops when the maximum number of iterations (500) is reached. The trained network is then the concrete performance prediction model.

[0026] For example, 200 sets of samples are divided into 160 training sets and 40 validation sets in an 8:2 ratio. The Adam optimizer, batch size 32, and mean squared error loss are used for training. The validation set loss decreases by less than 1×10 for 20 consecutive rounds. -4 Training is stopped at this point to obtain the initial concrete performance prediction model.

[0027] In this embodiment of the invention, a standardized dataset is constructed by collecting bridge construction parameters and performance indicators, and dimensional differences are eliminated by normalization; a multi-layer feedforward deep neural network structure is adopted, and batch normalization and discarding layers are used to improve the generalization of the model; normalized training is completed by fixing the loss function, optimizer and convergence conditions to obtain an initial concrete performance prediction model, which can initially fit the nonlinear mapping relationship between construction parameters and concrete performance, providing a reliable basic model for subsequent uncertainty assessment and parameter optimization.

[0028] S2: In the construction parameter space, calculate the prediction uncertainty of the concrete performance prediction model for each combination of construction parameters, and screen out the parameter regions with high uncertainty.

[0029] In this embodiment of the invention, the prediction uncertainty of the concrete performance prediction model for each combination of construction parameters is calculated in the construction parameter space, and high uncertainty parameter regions are screened out. The prediction accuracy of the initial concrete performance prediction model obtained in S1 varies regionally within the construction parameter space. In some parameter regions, due to sparse historical samples and complex parameter coupling, the prediction results fluctuate greatly and have low reliability. The robustness of the model prediction cannot be judged by a single prediction value, but the standard deviation of the prediction output value can directly reflect the degree of fluctuation. The greater the fluctuation, the worse the model's prediction effect on that parameter combination and the higher the uncertainty. To achieve accurate positioning of the weak prediction regions of the concrete performance prediction model, the discard layer needs to be activated during the testing phase to obtain the prediction distribution through multiple forward propagations, quantify the prediction uncertainty of the entire parameter space, and use an adaptive dynamic threshold to divide the high uncertainty parameter regions, providing targeted targets for subsequent directional generation of synthetic samples and iterative optimization of the concrete performance prediction model.

[0030] Step S2 in the method provided in this embodiment of the invention includes: The construction parameter space is determined based on the value range of each construction parameter; Uniform sampling is performed within the construction parameter space to obtain multiple candidate combinations of construction parameters; For each candidate combination of construction parameters, the candidate combination of construction parameters is input into the concrete performance prediction model multiple times, and the output of some neurons in the concrete performance prediction model is randomly discarded each time it is input, so as to obtain multiple predicted output values. Calculate the standard deviation of multiple predicted output values ​​corresponding to each candidate construction parameter combination, and use the standard deviation as a measure of the prediction uncertainty of the candidate construction parameter combination; Calculate the arithmetic mean of the prediction uncertainty measure for all candidate construction parameter combinations, and use the arithmetic mean as the first dynamic threshold. The region covered by candidate construction parameter combinations whose prediction uncertainty metric is higher than the first dynamic threshold is marked as a high uncertainty parameter region.

[0031] First, based on the value range of each construction parameter, the construction parameter space is determined. The construction parameter space is a multi-dimensional continuous space composed of the reasonable value ranges of all adjustable construction parameters in S1, involving eight types of construction parameters, hence an 8-dimensional parameter space. This space forms the basis for subsequent sampling and uncertainty calculations. The maximum and minimum values ​​of the eight types of construction parameters in the historical dataset of S1 are statistically analyzed to determine the upper and lower limits of each construction parameter's value. The combination of the value ranges of all parameters constitutes the complete construction parameter space.

[0032] For example, based on the statistical results of 200 sets of construction parameters in S1, the value ranges of 8 types of parameters are determined: water-cement ratio 0.38~0.50, sand ratio 34%~40%, admixture dosage 10%~20%, water-reducing agent dosage 0.8%~1.2%, pouring temperature 20℃~30℃, curing humidity 85%~95%, surface roughness of old concrete grade I~III, and interface agent divided into epoxy / emulsion types. The above ranges are combined to form an 8-dimensional construction parameter space.

[0033] Secondly, uniform sampling is performed within the construction parameter space to obtain multiple candidate combinations of construction parameters. Uniform sampling refers to selecting numerical points at equal intervals within the value range of each type of construction parameter, and then arranging and combining the numerical points of different parameters to finally obtain candidate combinations that uniformly cover the entire construction parameter space. This avoids sampling being concentrated in a certain local area and ensures the comprehensiveness of uncertainty calculation.

[0034] Specifically, for the eight categories of construction parameters, continuous construction parameters are divided into value points at fixed intervals. Continuous construction parameters include six categories: water-cement ratio, sand ratio, admixture dosage, water-reducing agent dosage, pouring temperature, and curing humidity. For discrete construction parameters, all categories are selected as value points. Discrete construction parameters include two categories: surface roughness grade of old concrete and interface agent type. All value points of the construction parameters are permuted and combined to obtain multiple candidate construction parameter combinations. For example, by dividing each continuous parameter into five equally spaced value points and covering all discrete parameters, the final combination generates... A set of evenly distributed candidate construction parameter combinations, covering the entire construction parameter space.

[0035] Furthermore, for each candidate construction parameter combination, the candidate construction parameter combination is input into the concrete performance prediction model multiple times, and the output of some neurons in the concrete performance prediction model is randomly discarded each time it is input, so as to obtain multiple prediction output values.

[0036] Specifically, for each candidate combination of construction parameters, the candidate combination of construction parameters is input into the concrete performance prediction model multiple times, and the outputs of some neurons in the concrete performance prediction model are randomly discarded each time, resulting in multiple predicted output values, including: After each hidden layer of the concrete performance prediction model, a dropout layer used during training is retained, and the dropout layer is kept active during the testing phase. The dropout layer only temporarily blocks the output of some neurons during forward propagation and does not change the network structure of the concrete performance prediction model. For the same candidate combination of construction parameters, the forward propagation calculation is repeated a preset number of times, and the discard layer randomly discards a different proportion of neurons during each forward propagation to obtain a preset number of predicted output values; The preset number of predicted output values ​​are used as multiple predicted output values.

[0037] First, a dropout layer used during training is retained after each hidden layer of the concrete performance prediction model, and this dropout layer remains active during the testing phase. During forward propagation, the dropout layer only temporarily masks the output of some neurons, without altering the network structure of the concrete performance prediction model. The dropout layer is a built-in layer in the concrete performance prediction model S1, used during training to randomly mask some neurons to prevent overfitting; it is turned off during the regular testing phase. Activating the dropout layer during the testing phase is equivalent to Monte Carlo Dropout, where different neurons are randomly masked for each prediction. This does not modify the model's network structure or weights, but only temporarily changes the computation path for a single prediction, thus simulating the randomness of the model's predictions.

[0038] Specifically, the dropout layer after each hidden layer in the concrete performance prediction model trained in S1 is retained; the automatic test shutdown function of the dropout layer is disabled, ensuring that the dropout layer remains active during the prediction phase; the dropout layer only temporarily masks a portion of the neuron outputs, and the number of layers, neurons, and weights of the model remain unchanged. For example, using the dropout layer with a dropout rate of 0.2 from S1, 20% of the hidden layer neuron outputs are randomly masked during each prediction in the testing phase. The concrete performance prediction model still has an 8-input, 3-hidden-layer, 6-output structure, without any structural modifications.

[0039] Secondly, for the same candidate combination of construction parameters, the forward propagation calculation is repeated a preset number of times. During each forward propagation, a different proportion of neurons are randomly discarded in the discard layer, resulting in a preset number of predicted output values. Forward propagation is the process of calculating predicted concrete performance values ​​layer by layer after the construction parameter combination is input into the model. Repeating forward propagation multiple times for the same candidate combination, due to the different neurons discarded each time, will yield multiple sets of different prediction results. The degree of fluctuation in the prediction results directly reflects the uncertainty.

[0040] Specifically, the preset number of forward propagation iterations is 50, which is a conventionally reproducible number in this field. Any candidate combination of construction parameters is selected and input into the prediction model of the activated dropout layer; the forward propagation calculation is repeated 50 times, with different neurons randomly masked in each dropout layer, ultimately yielding 50 complete sets of concrete performance prediction output values.

[0041] For example, the following candidate combinations were selected: water-cement ratio 0.48, sand ratio 39%, admixture dosage 18%, water-reducing agent dosage 1.1%, pouring temperature 28℃, curing humidity 88%, roughness grade II, and epoxy interface agent. After inputting the data into the model, 50 forward propagation cycles were performed to obtain 50 sets of predicted values, each containing 6 indicators including 3-day strength, 28-day strength, and bond strength.

[0042] Finally, a preset number of predicted output values ​​are used as multiple predicted output values. For example, 50 sets of predicted output values ​​are obtained in the end.

[0043] Furthermore, the standard deviation of multiple predicted output values ​​corresponding to each candidate construction parameter combination is calculated, and this standard deviation is used as a measure of the prediction uncertainty of that candidate construction parameter combination. Standard deviation is a statistical indicator that measures the magnitude of fluctuation in a set of data; here it is used to calculate the dispersion of multiple predicted output values ​​for the same candidate combination. The smaller the standard deviation, the smaller the fluctuation in the prediction result, and the more stable and reliable the model prediction; the larger the standard deviation, the greater the fluctuation, and the higher the prediction uncertainty.

[0044] Specifically, for each candidate combination of construction parameters, the corresponding values ​​of the six concrete performance indicators in the 50 prediction results are extracted, and the independent standard deviation of each indicator is calculated by substituting them into the standard deviation formula. The arithmetic mean of the standard deviations of the six performance indicators is then calculated, and this average standard deviation is used as the measurement value of the prediction uncertainty of the candidate combination of construction parameters.

[0045] For example, the standard deviations of the six performance indicators predicted 50 times for the above candidate combination were calculated to be 0.48, 0.51, 0.49, 0.53, 0.50 and 0.52, respectively. The average value was 0.505, which means that the prediction uncertainty measure of the candidate construction parameter combination is 0.505.

[0046] Subsequently, the arithmetic mean of the prediction uncertainty measures for all candidate construction parameter combinations is calculated, and this arithmetic mean is used as the first dynamic threshold. The first dynamic threshold is the arithmetic mean calculated from the uncertainty measure values ​​of all candidate combinations. It is an adaptive threshold, requiring no manual setting of a fixed value, and can objectively distinguish the boundaries between high and low uncertainties in the construction parameter space.

[0047] Specifically, all prediction uncertainty measures for multiple candidate construction parameter combinations are summarized; all measures are added together and then divided by the total number of candidate combinations to obtain the arithmetic mean; this arithmetic mean is defined as the first dynamic threshold. For example, the calculated... The arithmetic mean of the uncertainty measure of the candidate combinations is 0.45, which is the first dynamic threshold of 0.45.

[0048] Finally, the region covered by candidate construction parameter combinations whose prediction uncertainty exceeds the first dynamic threshold is marked as a high-uncertainty parameter region. The high-uncertainty parameter region is the interval in the construction parameter space covered by parameter combinations whose model predictions exhibit large fluctuations, poor accuracy, and low reliability. This region, due to insufficient historical samples or complex parameter coupling, is a key area for model optimization.

[0049] Specifically, the uncertainty metric of each candidate combination is compared with the first dynamic threshold; all candidate construction parameter combinations with uncertainty metrics greater than the first dynamic threshold are selected; based on the parameter value range corresponding to the selected combinations, continuous intervals are marked in the construction parameter space, which are the high uncertainty parameter regions. For example, all candidate combinations with standard deviations greater than 0.45 MPa are selected, and they are all concentrated in the range of water-cement ratio 0.47~0.50 and sand ratio 38%~40%, and this parameter range is marked as the high uncertainty parameter region.

[0050] In this embodiment of the invention, by activating the discard layer during the testing phase and calculating the standard deviation through multiple predictions, the predictive stability of the initial concrete performance prediction model in different parameter regions can be intuitively and quantitatively determined, and the high uncertainty parameter regions with large prediction fluctuations and low accuracy can be accurately located. The mean value is used as a dynamic threshold, eliminating the need for manual parameter tuning. The division results are objective and accurate, providing a clear target for subsequent directional generation of synthetic samples and iterative optimization of the concrete performance prediction model, thus avoiding ineffective operations.

[0051] S3: Within the region of high uncertainty parameters, new urban bridge construction parameters and their corresponding concrete performance simulation values ​​are generated based on the parameter update model, forming a synthetic sample set.

[0052] In this embodiment of the invention, within the high-uncertainty parameter region, new urban bridge construction parameters and their corresponding simulated concrete performance values ​​are generated based on the parameter update model, forming a synthetic sample set. S2 has accurately located the high-uncertainty parameter region. Due to the sparse historical samples and complex parameter coupling in this region, model predictions exhibit large fluctuations and low accuracy. Model optimization cannot be achieved solely by relying on existing historical datasets; therefore, it is necessary to selectively supplement effective samples in this region. Conditional Generative Adversarial Networks (CGANs) can generate new construction parameter combinations that conform to engineering realities and have a reasonable distribution, based on the existing historical sample distribution and combined with the characteristics of the high-uncertainty region and construction scenario constraints. Then, the corresponding simulated values ​​are obtained through the trained concrete performance prediction model, forming a synthetic sample set. This provides targeted supplementary samples for subsequent model iterations and updates, addressing the problems of insufficient and unevenly distributed historical samples.

[0053] Step S3 in the method provided in this embodiment of the invention includes: A conditional generative adversarial network is constructed as a parameter generation model, wherein the conditional generative adversarial network includes a generator and a discriminator; Using the construction parameters and corresponding measured values ​​of concrete performance in the historical dataset as training samples, the parameter generation model is pre-trained to obtain the trained parameter generation model. The region identifier of the high uncertainty parameter region and the corresponding urban bridge construction feature code are input into the trained parameter generation model to generate multiple new construction parameter combinations. The urban bridge construction feature code includes at least the surface roughness level of old concrete, the type of interface agent, and the degree of spatial constraint of urban bridge construction. Multiple new construction parameter combinations are sequentially input into the concrete performance prediction model, and the corresponding concrete performance prediction values ​​are output. The multiple new construction parameter combinations are paired with the corresponding concrete performance prediction values ​​as concrete performance simulation values. A composite sample set is formed by combining all the new construction parameters and their corresponding simulated concrete performance values.

[0054] First, a Conditional Generative Adversarial Network (CGAN) is constructed as the parameter generation model. This CGAN comprises a generator and a discriminator. CGAN is a generative model built upon Generative Adversarial Networks (GANs) with added conditional constraints. It consists of two sub-networks: a generator and a discriminator. Based on input conditional information, it generates target data that meets specific requirements. The generator generates simulated construction parameter combinations, while the discriminator distinguishes between real historical samples and generated samples. The two networks train adversarially until the discriminator can no longer distinguish between real and fake samples, at which point the model training is complete. This CGAN is used to selectively generate new construction parameter combinations within regions of high uncertainty.

[0055] Specifically, the generator is first constructed as follows: the input layer has 10 neurons, corresponding to 4D high uncertainty region labeling + 3D construction feature encoding + 3D random noise; the hidden layer has 3 layers, with the number of neurons being 64, 32 and 16 respectively, and the activation function is ReLU for all of them; the output layer has 8 neurons, corresponding to 8 types of construction parameters; the activation function of the output layer is adjusted according to the parameter type, with linear activation for continuous parameters and softmax activation for discrete parameters.

[0056] Secondly, a discriminator is constructed: the input layer has 11 neurons, containing 8 types of construction parameters + 3-dimensional construction feature encoding; the hidden layer has 2 layers, with 32 and 16 neurons respectively, and the activation function is ReLU for both layers; the output layer has 1 neuron, and the activation function is Sigmoid, which is used to determine whether the combination of input construction parameters is a real historical sample or a generated sample.

[0057] For example, CGAN with the above structure is built as a parameter generation model. The generator input includes high uncertainty region identifiers, construction feature codes and random noise, and outputs 8 types of construction parameters. The discriminator inputs construction parameters and feature codes, and outputs probability values ​​between 0 and 1. Those close to 1 are real samples, and those close to 0 are generated samples.

[0058] Secondly, the parameter generation model is pre-trained using the construction parameters and corresponding measured concrete performance values ​​from the historical dataset as training samples, resulting in a trained parameter generation model. Pre-training refers to using real samples from the historical dataset to initially train the parameter generation model (CGAN), enabling the model to learn the true mapping relationship between construction parameters and concrete performance and construction characteristics. This ensures that the newly generated parameter combinations conform to the actual engineering situation and are consistent with the distribution of historical samples.

[0059] Specifically, a historical dataset of 200 sets was constructed using S1. The construction parameters of each set were used as the generation target, while the corresponding measured concrete performance values ​​and construction feature codes served as constraints. These parameters were then input into the model. The Adam optimizer was used with a learning rate of 0.0002 and a batch size of 32. The generator loss function was cross-entropy loss, and the discriminator loss function was binary classification cross-entropy loss. Training iterations continued until both the discriminator and generator losses stabilized between 0.6 and 0.7, and the loss decreased by less than 1 × 10⁻⁶ for 20 consecutive iterations. -4 Stop pre-training and obtain the trained parameter generation model.

[0060] For example, 200 sets of historical samples are input into CGAN and pre-trained according to the above settings. After 300 iterations, the loss of both the discriminator and the generator stabilizes at around 0.65, and training is stopped, resulting in a parameter generation model that can be used to generate new parameters.

[0061] Next, the region identifier of the high uncertainty parameter area and the corresponding urban bridge construction feature code are input into the trained parameter generation model to generate multiple new combinations of construction parameters. The urban bridge construction feature code includes at least the surface roughness level of the old concrete, the type of interface agent, and the level of spatial constraint in urban bridge construction. The high uncertainty region identifier is information obtained by digitally encoding the high uncertainty parameter area marked by S2, used to constrain the generator to generate parameters only within this region. The urban bridge construction feature code is information obtained by digitally processing the key constraints of the construction scenario, ensuring that the generated parameters conform to the actual construction conditions, including the surface roughness level of the old concrete, the type of interface agent, and the level of spatial constraint in construction.

[0062] Specifically, the parameter range of high-uncertainty regions is encoded as a 4-dimensional vector; the surface roughness levels of old concrete (I-III) are encoded as [1,0,0], [0,1,0], and [0,0,1]; the interface agent types (epoxy and emulsion) are encoded as [1,0] and [0,1]; and the levels of construction space constraint (mild, moderate, and severe) are encoded as [1,0,0], [0,1,0], and [0,0,1]. These three are combined to form a 3-dimensional construction feature encoding. The above high-uncertainty region identifiers and construction feature encodings are combined with a parameter generation model trained using random noise input, generating 500 sets of new construction parameter combinations that meet the conditions. Invalid combinations that exceed the reasonable range of the construction parameter space are discarded, and valid combinations are retained.

[0063] For example, the high uncertainty area is coded as [0.47, 0.50, 0.38, 0.40], and the construction feature is coded as [0, 1, 0], [1, 0], and [0, 1, 0]. After combining them, they are input into the parameter generation model to generate 500 new construction parameter combinations. 12 invalid combinations that are out of reasonable range are removed, and finally 488 valid new construction parameter combinations are obtained.

[0064] Furthermore, multiple new combinations of construction parameters are sequentially input into the concrete performance prediction model, and corresponding predicted concrete performance values ​​are output. These new combinations of construction parameters are then paired with the corresponding predicted concrete performance values ​​to form simulated concrete performance values. The simulated concrete performance values ​​are the predicted output values ​​obtained by inputting the generated new combinations of construction parameters into the concrete performance prediction model trained by S1. These simulated values ​​are not actual field measurements but are used to replace actual measurements and are paired with the new combinations of construction parameters to form a synthetic sample. This eliminates the need for additional field tests, reducing costs and improving efficiency.

[0065] For example, a newly generated set of construction parameters is selected: water-cement ratio 0.49, sand ratio 39%, admixture dosage 19%, water-reducing agent dosage 1.15%, pouring temperature 29℃, curing humidity 87%, roughness grade II, epoxy interface agent. These are input into the concrete performance prediction model, and the output simulation values ​​are: 3-day compressive strength 25.8MPa, 28-day compressive strength 44.2MPa, bond strength between new and old concrete 1.95MPa, slump 168mm, heat of hydration and internal / external temperature difference 11.5℃, shrinkage strain 172×10⁻⁶. -6 The parameters are then paired with the simulation values ​​to form a valid synthetic sample.

[0066] Finally, a synthetic sample set is constructed by combining all new construction parameter combinations with their corresponding simulated concrete performance values. This synthetic sample set is specifically designed to fill sample gaps in areas of high uncertainty, providing targeted training data for subsequent model iterations and updates.

[0067] Specifically, all valid synthetic samples after verification are aggregated to form a complete synthetic sample set. The sample format is consistent with the historical dataset of S1, and each set contains 8 types of construction parameters and 6 performance indicators, which facilitates subsequent merging with the historical dataset. For example, the above 488 sets of valid synthetic samples are aggregated, each set containing complete construction parameters and corresponding simulation values, and the format is consistent with the historical dataset of S1, forming a synthetic sample set for model updates.

[0068] In this embodiment of the invention, the parameter generation model constructed by CGAN can generate new construction parameter combinations that conform to engineering reality and have a reasonable distribution in the region of high uncertainty parameters. Combined with the concrete performance prediction model, simulation values ​​are obtained without the need for additional field tests, reducing sample collection costs and improving efficiency. The generated synthetic sample set specifically fills the sample gap in the region of high uncertainty and is in the same format as the historical dataset, providing targeted and effective supplementary data for subsequent iterative updates of the S4 concrete performance prediction model, ensuring that the concrete performance prediction model can specifically optimize the prediction accuracy of weak areas.

[0069] S4: Add the synthetic sample set to the historical dataset and retrain the concrete performance prediction model to obtain an updated concrete performance prediction model.

[0070] In this embodiment of the invention, the synthetic sample set is added to the historical dataset, and the concrete performance prediction model is retrained to obtain an updated concrete performance prediction model. The synthetic sample set generated by S3 specifically fills the sample gaps in the high uncertainty parameter region located by S2, while the original historical dataset only contains a small number of conventional parameter samples, and the samples in this weak region are sparse and cannot support model optimization. Therefore, it is necessary to merge the synthetic sample set with the historical dataset to expand the dataset size, improve the sample distribution, and specifically make up for the sample defects in the high uncertainty region. Then, through normalized retraining, the network weights of the concrete performance prediction model are optimized, improving the prediction accuracy and robustness of the model in the entire construction parameter space, and providing a more reliable model foundation for subsequent iterative convergence and parameter optimization.

[0071] First, the effective synthetic sample set obtained in S3 is merged with the historical dataset constructed in S1 to unify the sample format. Each group contains 8 types of construction parameters and 6 concrete performance indicators, resulting in a complete merged training dataset. For the merged dataset, the Min-Max normalization method, identical to that used in S1, is applied to map all construction parameters and performance indicators to the [0,1] interval, ensuring consistent preprocessing and avoiding the impact of data scale differences on model training.

[0072] Secondly, the multi-layer feedforward deep neural network structure built using S1 is adopted, with the same training settings and training termination condition as S1, namely, the validation set loss decreases by less than 1×10 for 20 consecutive iterations. -4 Training can be stopped once the maximum number of iterations reaches 500 rounds. The resulting deep neural network is the updated concrete performance prediction model.

[0073] In this embodiment of the invention, by merging historical samples and synthetic samples, the sample gaps in high uncertainty parameter regions are effectively filled, the dataset size is expanded, and the sample distribution is optimized, solving the problems of sparse and unevenly distributed samples in the original dataset. By adopting the same network structure and training settings as the initial training, the training process is ensured to be reproducible and the results are stable. The retrained concrete performance prediction model shows reduced prediction fluctuations and improved accuracy in high uncertainty regions, and its generalization ability across the entire construction parameter space is significantly enhanced. At the same time, the updated concrete performance prediction model maintains the same input and output format as the initial concrete performance prediction model, which can be directly used for iterative convergence verification of S5 and subsequent parameter optimization, providing reliable model support for the smooth progress of the entire method.

[0074] S5: Repeat steps S2 to S4 until the updated concrete performance prediction model is verified to converge in the entire construction parameter space, and the converged concrete performance prediction model is obtained.

[0075] In this embodiment of the invention, steps S2 to S4 are repeated until the updated concrete performance prediction model is verified to converge in the entire construction parameter space, thus obtaining a converged concrete performance prediction model. After a single supplementation of synthetic samples and model update, there may still be some areas in the construction parameter space where the prediction uncertainty does not meet the requirements for engineering applications. A single iteration cannot completely eliminate the model's prediction blind spots. To ensure that the concrete performance prediction model has stable and reliable prediction capabilities throughout the entire construction parameter space, it is necessary to iteratively execute the uncertainty assessment, synthetic sample generation, and model retraining process. By monitoring whether the maximum prediction uncertainty in the entire space is lower than a preset convergence threshold, this is used as the criterion for model convergence until the model is robust in the entire parameter space with no weak areas, ultimately obtaining a converged concrete performance prediction model that can be directly used for parameter optimization.

[0076] Step S5 in the method provided in this embodiment of the invention includes: After each completion of step S4, step S2 is re-executed within the construction parameter space to calculate the prediction uncertainty measure of the current updated concrete performance prediction model for all candidate construction parameter combinations in the entire construction parameter space. Calculate the maximum value of the prediction uncertainty measure for all candidate construction parameter combinations, and use the maximum value as the current maximum prediction uncertainty; The maximum prediction uncertainty is compared with a preset convergence threshold. When the maximum prediction uncertainty is less than the convergence threshold, it is determined that the updated concrete performance prediction model has converged, and the current updated concrete performance prediction model is used as the converged concrete performance prediction model. When the maximum prediction uncertainty is greater than or equal to the convergence threshold, return to step S3, continue to generate new synthetic samples and update the concrete performance prediction model again.

[0077] First, after each step S4 is completed, step S2 is re-executed within the construction parameter space to calculate the prediction uncertainty measure of the updated concrete performance prediction model for all candidate construction parameter combinations in the entire construction parameter space. Repeating the entire S2 process means that, based on the latest updated concrete performance prediction model, the uncertainty of all candidate parameter combinations in the entire construction parameter space is quantified again to obtain the prediction fluctuation distribution of the model after this iteration, and to determine whether weak areas have been optimized.

[0078] Specifically, based on the updated concrete performance prediction model obtained in step S4, the entire process of S2 is repeated: the construction parameter space is determined, candidate construction parameter combinations are uniformly sampled, the discard layer is activated during the testing phase to complete 50 forward propagations, and the average of the standard deviations of the six performance indicators is calculated as the uncertainty measure to obtain the prediction uncertainty measure value of all candidate combinations in this round.

[0079] For example, the updated concrete performance prediction model obtained from the current S4 training is substituted into S2 to... Uncertainty measures were calculated for each candidate construction parameter combination. The measure value in the original high uncertainty area decreased from 0.505MPa to 0.31MPa, but the measure value in some areas was still too high.

[0080] Secondly, the maximum value of the prediction uncertainty metric for all candidate construction parameter combinations is calculated, and this maximum value is taken as the current maximum prediction uncertainty. The maximum prediction uncertainty refers to the maximum value of the uncertainty metric corresponding to all candidate construction parameter combinations in this iteration. This value reflects the model's prediction robustness at its worst position in the entire construction parameter space and serves as the basis for convergence judgment.

[0081] Specifically, all prediction uncertainty measures for candidate combinations are iterated through, and the maximum value is selected by numerical comparison. This maximum value is defined as the maximum prediction uncertainty for the current iteration. For example, after iterating through all uncertainty measures in this round, the maximum value is found to be 0.32 MPa, which is the maximum prediction uncertainty of the updated model.

[0082] Next, the maximum prediction uncertainty is compared with a preset convergence threshold. When the maximum prediction uncertainty is less than the convergence threshold, the updated concrete performance prediction model is determined to have converged, and the current updated concrete performance prediction model is taken as the converged concrete performance prediction model. The convergence threshold is a preset upper limit of the maximum uncertainty that meets the requirements of engineering applications. This threshold is determined based on the accuracy requirements of urban bridge renovation construction. When the maximum prediction uncertainty is lower than this value, it means that the model's full-space prediction accuracy meets the standard.

[0083] Specifically, a convergence threshold of 0.2 MPa is preset. The current maximum prediction uncertainty is compared with this convergence threshold, and the model convergence is determined based on the comparison result. For example, if the current maximum prediction uncertainty is 0.32 MPa and the preset convergence threshold is 0.2 MPa, and the comparison shows that 0.32 MPa > 0.2 MPa, the model is determined not to have met the convergence condition.

[0084] Finally, when the maximum prediction uncertainty is greater than or equal to the convergence threshold, return to step S3 to continue generating new synthetic samples and updating the concrete performance prediction model again. If the maximum prediction uncertainty is greater than or equal to the convergence threshold, return to step S3, generate supplementary synthetic samples in a new round of high uncertainty parameter regions, and execute S3 and S4 sequentially to complete the model update again; if the maximum prediction uncertainty is less than the convergence threshold, terminate the iteration process and determine the currently updated model as the converged concrete performance prediction model.

[0085] For example, since the current model has not converged, return to S3 to generate 320 new supplementary synthetic samples. After merging and retraining in S4, execute S2 again to calculate that the maximum prediction uncertainty is 0.18MPa. Since 0.18MPa < 0.2MPa, the convergence condition is met, the iteration is terminated, and the converged concrete performance prediction model is output.

[0086] In this embodiment of the invention, a closed-loop iterative optimization mechanism is used to continuously target and eliminate weak prediction areas in the construction parameter space, gradually reducing the maximum prediction uncertainty in the entire space. A preset convergence threshold is used as a quantitative judgment standard to ensure that the model convergence process is objective and reproducible. The final converged concrete performance prediction model has small prediction fluctuations, high accuracy, and strong generalization in the entire construction parameter space. It can be used as a stable and reliable fitness evaluator, providing solid support for subsequent intelligent optimization of construction parameters.

[0087] S6: Set target performance requirements, use the converged concrete performance prediction model as a fitness function evaluator, and use an optimization algorithm to search for urban bridge construction parameters that meet the target performance requirements to obtain the optimal combination of construction parameters.

[0088] In this embodiment of the invention, target performance requirements are set, and a converged concrete performance prediction model is used as a fitness function evaluator. An optimization algorithm searches for urban bridge construction parameters that meet the target performance requirements to obtain the optimal combination of construction parameters. The converged concrete performance prediction model has stable and accurate predictive capabilities across the entire construction parameter space and can serve as a reliable performance evaluation tool. In urban bridge renovation construction, specific concrete target performances, such as bond strength and compressive strength, need to be set according to specific engineering requirements. Traditional construction parameter proportioning relies on manual experience and trial mixing, which suffers from low efficiency, inaccurate matching of target performance, and potential quality problems. Therefore, a converged model is used as a fitness evaluator, and a particle swarm optimization (PSO) algorithm is used to intelligently search within the construction parameter space to efficiently find the construction parameter combination that meets all target performance requirements and has the lowest fitness value. This achieves intelligent and precise optimization of construction parameters, replacing traditional experience-based trial mixing and improving construction efficiency and quality.

[0089] Step S6 in the method provided in this embodiment of the invention includes: Set target performance requirements, which include target values ​​or target ranges for various concrete performance indicators. Among them, the concrete performance indicators include at least 3-day compressive strength, 28-day compressive strength, bond strength between new and old concrete, slump, heat of hydration, internal and external temperature difference, and shrinkage strain. The converged concrete performance prediction model is used as a fitness function evaluator, and the combination of construction parameters is used as the input of the fitness function evaluator. The deviation between the predicted concrete performance value output by the converged concrete performance prediction model and the target performance requirement is used as the fitness value. With the minimum fitness value as the optimization objective, the particle swarm optimization algorithm is used to search for the optimal combination of construction parameters in the construction parameter space.

[0090] First, target performance requirements are set, including target values ​​or target ranges for various concrete performance indicators. These concrete performance indicators include at least 3-day compressive strength, 28-day compressive strength, bond strength between new and old concrete, slump, heat of hydration, internal and external temperature difference, and shrinkage strain. These target performance requirements are specific target values ​​or reasonable target ranges set for concrete performance indicators based on the actual quality needs of urban bridge renovation construction, and serve as the basis for judging whether the combination of construction parameters is qualified.

[0091] Specifically, the target requirements for various concrete performance indicators should be clearly defined, including at least 3-day compressive strength, 28-day compressive strength, bond strength between new and old concrete, slump, heat of hydration, internal and external temperature difference, and shrinkage strain. Among these, 3-day compressive strength, 28-day compressive strength, and bond strength between new and old concrete are strength-related indicators and require minimum target values; heat of hydration, internal and external temperature difference, and shrinkage strain are control-related indicators and require maximum limits; slump is a fluidity indicator and requires a reasonable range. For example, based on the needs of a city's elevated bridge reinforcement and renovation project, the target performance requirements are set as follows: 3-day compressive strength ≥ 25 MPa, 28-day compressive strength ≥ 45 MPa, bond strength between new and old concrete ≥ 2.0 MPa, slump 160~180 mm, heat of hydration, internal and external temperature difference ≤ 12℃, and shrinkage strain ≤ 180 × 10⁻⁶. -6 .

[0092] Secondly, the converged concrete performance prediction model is used as a fitness function evaluator, with the construction parameter combination as input. The deviation between the predicted concrete performance value output by the converged concrete performance prediction model and the target performance requirement is used as the fitness value. The fitness function evaluator is a tool used to evaluate whether the construction parameter combination meets the target performance requirement; in this case, it is the converged concrete performance prediction model. The fitness value is a quantitative value that measures the magnitude of the deviation between the predicted performance of the construction parameter combination and the target performance. The smaller the deviation, the smaller the fitness value, and the better the parameter combination.

[0093] Specifically, the converged concrete performance prediction model is used as the fitness function evaluator. Any combination of construction parameters is taken as input, and the concrete performance prediction model outputs six corresponding predicted concrete performance values. The sum of the absolute values ​​of the deviations of each performance indicator is used as the fitness value calculation rule. If the predicted value of a certain performance indicator exceeds the target range, an additional penalty term is added, with a penalty coefficient set to 10, to ensure that the parameter combination meets all target requirements.

[0094] The fitness value is calculated using the following formula: Fitness Value = |3-day compressive strength prediction - target value| + |28-day compressive strength prediction - target value| + |bond strength prediction - target value| + |slump prediction - target range median| + |hydration heat temperature difference prediction - target limit| + |shrinkage strain prediction - target limit| + (out-of-range items × 10). The slump value is calculated using the median value of the target range to account for deviations.

[0095] For example, a set of construction parameter combinations is selected, and after inputting into the convergent concrete performance prediction model, the predicted values ​​are: 3-day compressive strength 26MPa, 28-day compressive strength 44MPa, bond strength between new and old concrete 1.9MPa, slump 170mm, heat of hydration and internal / external temperature difference 11℃, and shrinkage strain 175×10⁻⁶. -6 The 28-day compressive strength is 44MPa < 45MPa and the bond strength is 1.9MPa < 2.0MPa, which are outside the target range. Substituting these values ​​into the formula, the fitness value is calculated to be 28.1.

[0096] Next, with the minimum fitness value as the optimization objective, the particle swarm optimization algorithm is used to search for the optimal combination of construction parameters in the construction parameter space.

[0097] Among them, taking the minimum fitness value as the optimization objective, the particle swarm optimization algorithm is used to search for the optimal combination of construction parameters in the construction parameter space, including: Initialize a particle swarm, which contains multiple particles, each corresponding to a combination of construction parameters. The position of each particle is randomly initialized within the construction parameter space, and the velocity of each particle is randomly initialized within a preset velocity range. The construction parameters corresponding to the position of each particle are input into the converged concrete performance prediction model, the predicted concrete performance value of each particle is output, and the fitness value is calculated in combination with the target performance requirements. The current fitness value of each particle is compared with the historical best fitness value of that particle. If the current fitness value is less than the historical best fitness value, the current position of the particle is updated to the individual best position of that particle. The position corresponding to the smallest fitness value among all the individual optimal positions of all particles is taken as the global optimal position, and the fitness value corresponding to the global optimal position is taken as the global optimal fitness value. According to the preset velocity update rules and position update rules, the velocity and position of each particle are updated, and the particles are guided to gather towards the individual optimal position and the global optimal position until the change of the global optimal fitness value in multiple consecutive iterations is lower than the preset change threshold or the preset maximum number of iterations is reached. The combination of construction parameters corresponding to the final globally optimal position is output as the optimal combination of construction parameters.

[0098] First, the particle swarm is initialized. The particle swarm contains multiple particles, each corresponding to a combination of construction parameters. The position of each particle is randomly initialized within the construction parameter space, and the velocity of each particle is randomly initialized within a preset velocity range. The particle swarm consists of multiple particles, each corresponding to a combination of construction parameters. The particle position is the parameter combination, and the particle velocity controls the search direction and step size of the particle within the construction parameter space. Initialization determines the initial position and velocity of the particles.

[0099] Specifically, the particle swarm size is set to 50 particles to balance search efficiency and accuracy. The position of each particle corresponds to the value of 8 types of construction parameters, which are randomly generated within the construction parameter space defined by S2 to ensure that all parameter values ​​are within a reasonable range. The velocity of each particle corresponds to the adjustment step size of the 8 types of construction parameters, which is randomly generated within a preset velocity range, and the velocity direction is random.

[0100] For example, 50 particles are initialized, where the initial position of particle 1 is: water-cement ratio 0.43, sand ratio 37%, admixture dosage 13%, water-reducing agent dosage 1.0%, pouring temperature 24℃, curing humidity 91%, roughness grade II, and epoxy interface agent; the initial velocity is: water-cement ratio 0.02, sand ratio 0.03, admixture dosage 0.02, water-reducing agent dosage 0.01, pouring temperature 0.04, curing humidity 0.03, roughness 0, and interface agent 0.

[0101] Secondly, the construction parameter combination corresponding to the position of each particle is input into the converged concrete performance prediction model, and the predicted concrete performance value of each particle is output. Combined with the target performance requirements, the fitness value is calculated. Each particle corresponds to a set of construction parameter combinations. By calculating its fitness value, the deviation between the parameter combination and the target performance requirements can be evaluated, providing a basis for subsequent particle position updates and optimal position selection.

[0102] Specifically, the initial positions of 50 particles are substituted one by one into the converged concrete performance prediction model, and the predicted values ​​of 6 performance items corresponding to each set of parameters are output. According to the fitness value calculation rules, the predicted value of each particle is compared with the target performance requirements one by one, and the fitness value of each particle is calculated.

[0103] For example, the initial position of particle 1 is input into the converged concrete performance prediction model, and the predicted values ​​are: 3-day compressive strength 27 MPa, 28-day compressive strength 46 MPa, bond strength between new and old concrete 2.1 MPa, slump 172 mm, heat of hydration and internal / external temperature difference 10.5℃, and shrinkage strain 170 × 10⁻⁶. -6 All indicators meet the target requirements, and the calculated fitness value is 16.6.

[0104] Next, the current fitness value of each particle is compared with its historical best fitness value. If the current fitness value is less than the historical best fitness value, the particle's current position is updated to its individual best position. The individual best position is the position with the minimum fitness value for a single particle throughout all iterations. The individual best fitness value, which corresponds to this position, is used to guide the particle to search in its optimal direction.

[0105] Specifically, each particle's individual optimal fitness value is initialized to its initial fitness value, and its individual optimal position is its initial position. After each calculation of a particle's current fitness value, it is compared with the particle's historical individual optimal fitness value. If the current fitness value is less than the historical individual optimal fitness value, it means that the current position is better, so the particle's individual optimal position is updated to the current position, and the individual optimal fitness value is updated to the current fitness value. If the current fitness value is greater than or equal to the historical individual optimal fitness value, the individual optimal position and fitness value remain unchanged.

[0106] For example, particle 1 has an initial fitness value of 16.6 and its optimal position is the initial position. After one position update, particle 1's current fitness value is 14.2. Since 14.2 < 16.6, particle 1's optimal position is updated to the current position, and its optimal fitness value is updated to 14.2.

[0107] Furthermore, the position with the smallest fitness value among all the individual optimal positions of all particles is taken as the global optimal position, and the fitness value corresponding to this global optimal position is taken as the global optimal fitness value. The global optimal position is the position with the smallest fitness value among all the individual optimal positions of all particles. The global optimal fitness value, i.e., the fitness value corresponding to this position, is the optimal parameter combination currently searched for by the entire particle swarm, used to guide all particles to gather towards the global optimal direction.

[0108] Specifically, the individual optimal fitness values ​​of all particles are iterated, and the minimum value is selected. The optimal position of the particle corresponding to this minimum value is found and determined as the global optimal position for the current iteration, with the corresponding fitness value being the global optimal fitness value. For example, after iterating through the individual optimal fitness values ​​of 50 particles, the minimum value of 12.8 is selected. This value corresponds to the individual optimal position of particle 23 with the following characteristics: water-cement ratio 0.42, sand ratio 36%, admixture dosage 15%, water-reducing agent dosage 1.0%, pouring temperature 25℃, curing humidity 90%, roughness grade II, and epoxy interface agent. Therefore, this position is determined as the global optimal position, with a global optimal fitness value of 12.8.

[0109] Then, according to preset velocity and position update rules, the velocity and position of each particle are updated, guiding the particles to converge towards the individual optimal position and the global optimal position, until the change in the global optimal fitness value over multiple consecutive iterations is lower than a preset change threshold or the preset maximum number of iterations is reached. The velocity update rule is used to adjust the particle's search direction and step size, guiding the particles to converge towards the individual optimal position and the global optimal position. The position update rule is used to adjust the particle's position in the parameter space based on the new velocity, achieving optimization of the parameter combination.

[0110] The speed update rule is the standard PSO formula: ;in The inertia weight is set to 0.8. , As an acceleration factor, all are set to 2.0. , Let v(t) be a random number between 0 and 1, v(t) be the current velocity, and t(t) be the current position. For the optimal position of an individual, This is the globally optimal position. Position update rule: After the update, particle positions need to be verified to ensure that all parameters remain within a reasonable range in the construction parameter space. If they exceed the range, they should be adjusted to the boundary values. Convergence condition: The change in the global optimal fitness value over 30 consecutive iterations is <1×10. -3 If the preset maximum number of iterations of 300 is reached, the iteration stops; otherwise, the above steps are repeated to continue updating the particle velocity and position.

[0111] For example, taking particle 1 as an example, its current velocity v(t) = 0.02, current position t(t) = 0.43, and its optimal position... Global optimal position , Substituting into the formula, we get the new velocity v(t+1)=0; the new position t(t+1)=0.43+0=0.43; repeating the iteration, when the global optimal fitness value drops from 12.8 to 12.79, the change range is <1×10 for 30 consecutive rounds. -3 Stop iterating.

[0112] Finally, the combination of construction parameters corresponding to the final globally optimal position is output as the optimal construction parameter combination. The optimal construction parameter combination is the combination of construction parameters corresponding to the globally optimal position after iterative convergence. This combination has the smallest fitness value, its prediction performance is closest to the target performance requirement, and all performance indicators meet the preset target.

[0113] Specifically, after iterative convergence, the final global optimal position is extracted, and the values ​​of the eight types of construction parameters corresponding to this position are converted into a parameter format that can be directly used in actual engineering. This parameter combination is then input into the converged concrete performance prediction model to verify whether its predicted performance fully meets the target performance requirements, ensuring the effectiveness of the parameter combination. The verified construction parameter combination is output as the optimal construction parameter combination and used in actual urban bridge renovation construction.

[0114] For example, after iterative convergence, the final globally optimal location corresponds to the following combination of construction parameters: water-cement ratio 0.42, sand ratio 36%, admixture dosage 15%, water-reducing agent dosage 1.0%, pouring temperature 25℃, curing humidity 90%, old concrete surface roughness grade II, and epoxy interface agent. Inputting this combination into the concrete performance prediction convergence model, the predicted performance meets the target requirements. This combination is then output as the optimal construction parameter combination for on-site construction.

[0115] In this embodiment of the invention, the converged prediction model is used as the fitness evaluator to ensure the accuracy and stability of the parameter combination performance evaluation. The particle swarm optimization algorithm enables intelligent search of the construction parameter space, replacing traditional manual experience-based trial mixing and improving parameter optimization efficiency. During the iteration process, individual optimality and global optimality guidance ensure accurate search direction, and the final optimal construction parameter combination fully meets the preset target performance requirements. This achieves precise matching between concrete performance and construction parameters, effectively avoiding construction quality problems caused by unreasonable parameters, reducing construction costs and trial mixing workload, and providing a scientific and reliable parameter basis for urban bridge renovation construction.

[0116] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects: This invention provides a method and system for optimizing construction parameters for urban bridge renovation based on deep learning. It constructs a standardized historical dataset and trains a deep learning prediction model, quantifies the prediction uncertainty across the entire construction parameter space, generates synthetic samples in high-uncertainty regions, iteratively optimizes the model until the entire space prediction converges, and finally intelligently searches for the optimal combination of construction parameters based on a particle swarm optimization algorithm combined with the convergence model. This achieves accurate fitting of the nonlinear relationship between concrete construction parameters and performance indicators, effectively compensating for prediction defects in sparse sample regions, improving the stability and generalization accuracy of concrete performance prediction across the entire space, and eliminating the traditional experience-based trial-and-match model. It realizes intelligent and precise optimization of construction parameters, ensuring that all concrete performance indicators meet the engineering requirements of urban bridge renovation construction, thus improving construction quality and efficiency.

[0117] Example 2, as Figure 2 As shown, this invention provides a deep learning-based system for optimizing construction parameters for urban bridge upgrades. The system includes: The initial model training module 11 is used to obtain multiple sets of urban bridge construction parameters and corresponding measured values ​​of concrete performance to form a historical dataset, and to train a deep neural network using the historical dataset to obtain a concrete performance prediction model. Uncertainty assessment module 12 is used to calculate the prediction uncertainty of the concrete performance prediction model for each combination of construction parameters in the construction parameter space, and to screen out the parameter region with high uncertainty. The synthetic sample generation module 13 is used to generate new urban bridge construction parameters and their corresponding concrete performance simulation values ​​based on the parameter update model within the high uncertainty parameter region, thereby forming a synthetic sample set. The model incremental training module 14 is used to add the synthetic sample set to the historical dataset and retrain the concrete performance prediction model to obtain an updated concrete performance prediction model. The iterative convergence judgment module 15 is used to repeatedly call the uncertainty assessment module, the synthetic sample generation module and the model incremental training module until the updated concrete performance prediction model is verified to converge in the entire construction parameter space, and the converged concrete performance prediction model is obtained. The optimal parameter search module 16 is used to set target performance requirements, use the converged concrete performance prediction model as a fitness function evaluator, and search for urban bridge construction parameters that meet the target performance requirements through optimization algorithms to obtain the optimal combination of construction parameters.

[0118] In one embodiment, the initial model training module 11 is further configured to: This involves acquiring multiple sets of urban bridge construction parameters and corresponding measured values ​​of concrete performance to form a historical dataset, including: Multiple sets of construction parameters were collected during the urban bridge renovation construction process. These construction parameters included at least the water-cement ratio, sand ratio, admixture dosage, water-reducing agent dosage, pouring temperature, curing humidity, surface roughness grade of old concrete, and interface agent type. For each set of construction parameters, the corresponding measured values ​​of concrete performance are obtained through on-site experiments or quality inspection reports. The measured values ​​of concrete performance include at least 3-day compressive strength, 28-day compressive strength, bond strength between new and old concrete, slump, heat of hydration, internal and external temperature difference, and shrinkage strain. Each set of construction parameters is paired with the corresponding measured values ​​of concrete performance to form a data sample. All data samples are combined to form a historical dataset.

[0119] The concrete performance prediction model is obtained by training a deep neural network using the historical dataset, including: The construction parameters and measured concrete performance values ​​in the historical dataset are normalized to obtain normalized construction parameters and normalized measured concrete performance values. Construct a multi-layer feedforward deep neural network, which includes an input layer, multiple hidden layers and an output layer, with each hidden layer followed by a batch normalization layer and a dropout layer; Normalized construction parameters are used as input features, and normalized measured concrete performance values ​​are used as output labels. The multilayer feedforward deep neural network is trained in a supervised manner until it is verified to converge, thus obtaining a trained deep neural network. The trained deep neural network is then used as a concrete performance prediction model.

[0120] In one embodiment, the uncertainty assessment module 12 is further configured to: The construction parameter space is determined based on the value range of each construction parameter; Uniform sampling is performed within the construction parameter space to obtain multiple candidate combinations of construction parameters; For each candidate combination of construction parameters, the candidate combination of construction parameters is input into the concrete performance prediction model multiple times, and the output of some neurons in the concrete performance prediction model is randomly discarded each time it is input, so as to obtain multiple predicted output values. Calculate the standard deviation of multiple predicted output values ​​corresponding to each candidate construction parameter combination, and use the standard deviation as a measure of the prediction uncertainty of the candidate construction parameter combination; Calculate the arithmetic mean of the prediction uncertainty measure for all candidate construction parameter combinations, and use the arithmetic mean as the first dynamic threshold. The region covered by candidate construction parameter combinations whose prediction uncertainty metric is higher than the first dynamic threshold is marked as a high uncertainty parameter region.

[0121] Specifically, for each candidate combination of construction parameters, the candidate combination of construction parameters is input into the concrete performance prediction model multiple times, and the outputs of some neurons in the concrete performance prediction model are randomly discarded each time, resulting in multiple predicted output values, including: After each hidden layer of the concrete performance prediction model, a dropout layer used during training is retained, and the dropout layer is kept active during the testing phase. The dropout layer only temporarily blocks the output of some neurons during forward propagation and does not change the network structure of the concrete performance prediction model. For the same candidate combination of construction parameters, the forward propagation calculation is repeated a preset number of times, and the discard layer randomly discards a different proportion of neurons during each forward propagation to obtain a preset number of predicted output values; The preset number of predicted output values ​​are used as multiple predicted output values.

[0122] In one embodiment, the synthetic sample generation module 13 is further configured to: A conditional generative adversarial network is constructed as a parameter generation model, wherein the conditional generative adversarial network includes a generator and a discriminator; Using the construction parameters and corresponding measured values ​​of concrete performance in the historical dataset as training samples, the parameter generation model is pre-trained to obtain the trained parameter generation model. The region identifier of the high uncertainty parameter region and the corresponding urban bridge construction feature code are input into the trained parameter generation model to generate multiple new construction parameter combinations. The urban bridge construction feature code includes at least the surface roughness level of old concrete, the type of interface agent, and the degree of spatial constraint of urban bridge construction. Multiple new construction parameter combinations are sequentially input into the concrete performance prediction model, and the corresponding concrete performance prediction values ​​are output. The multiple new construction parameter combinations are paired with the corresponding concrete performance prediction values ​​as concrete performance simulation values. A composite sample set is formed by combining all the new construction parameters and their corresponding simulated concrete performance values.

[0123] In one embodiment, the iterative convergence determination module 15 is further configured to: After each call to the model incremental training module 14, the uncertainty assessment module 12 is called again in the construction parameter space to calculate the prediction uncertainty measure of the current updated concrete performance prediction model for all candidate construction parameter combinations in the entire construction parameter space. Calculate the maximum value of the prediction uncertainty measure for all candidate construction parameter combinations, and use the maximum value as the current maximum prediction uncertainty; The maximum prediction uncertainty is compared with a preset convergence threshold. When the maximum prediction uncertainty is less than the convergence threshold, it is determined that the updated concrete performance prediction model has converged, and the current updated concrete performance prediction model is used as the converged concrete performance prediction model. When the maximum prediction uncertainty is greater than or equal to the convergence threshold, the synthetic sample generation module 13 is invoked to continue generating new synthetic samples and update the concrete performance prediction model again.

[0124] In one embodiment, the optimal parameter search module 16 is further configured to: Set target performance requirements, which include target values ​​or target ranges for various concrete performance indicators. Among them, the concrete performance indicators include at least 3-day compressive strength, 28-day compressive strength, bond strength between new and old concrete, slump, heat of hydration, internal and external temperature difference, and shrinkage strain. The converged concrete performance prediction model is used as a fitness function evaluator, and the combination of construction parameters is used as the input of the fitness function evaluator. The deviation between the predicted concrete performance value output by the converged concrete performance prediction model and the target performance requirement is used as the fitness value. With the minimum fitness value as the optimization objective, the particle swarm optimization algorithm is used to search for the optimal combination of construction parameters in the construction parameter space.

[0125] Among them, taking the minimum fitness value as the optimization objective, the particle swarm optimization algorithm is used to search for the optimal combination of construction parameters in the construction parameter space, including: Initialize a particle swarm, which contains multiple particles, each corresponding to a combination of construction parameters. The position of each particle is randomly initialized within the construction parameter space, and the velocity of each particle is randomly initialized within a preset velocity range. The construction parameters corresponding to the position of each particle are input into the converged concrete performance prediction model, the predicted concrete performance value of each particle is output, and the fitness value is calculated in combination with the target performance requirements. The current fitness value of each particle is compared with the historical best fitness value of that particle. If the current fitness value is less than the historical best fitness value, the current position of the particle is updated to the individual best position of that particle. The position corresponding to the smallest fitness value among all the individual optimal positions of all particles is taken as the global optimal position, and the fitness value corresponding to the global optimal position is taken as the global optimal fitness value. According to the preset velocity update rules and position update rules, the velocity and position of each particle are updated, and the particles are guided to gather towards the individual optimal position and the global optimal position until the change of the global optimal fitness value in multiple consecutive iterations is lower than the preset change threshold or the preset maximum number of iterations is reached. The combination of construction parameters corresponding to the final globally optimal position is output as the optimal combination of construction parameters.

[0126] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing construction parameters of urban bridge renewal based on deep learning, characterized in that, The method includes: S1: Obtain multiple sets of urban bridge construction parameters and corresponding measured values ​​of concrete performance to form a historical dataset, and use the historical dataset to train a deep neural network to obtain a concrete performance prediction model. S2: In the construction parameter space, calculate the prediction uncertainty of the concrete performance prediction model for each combination of construction parameters, and screen out the parameter regions with high uncertainty. S3: Within the region of high uncertainty parameters, new urban bridge construction parameters and their corresponding concrete performance simulation values ​​are generated based on the parameter update model to form a synthetic sample set. S4: Add the synthetic sample set to the historical dataset and retrain the concrete performance prediction model to obtain an updated concrete performance prediction model. S5: Repeat steps S2 to S4 until the updated concrete performance prediction model is verified to converge in the entire construction parameter space, and the converged concrete performance prediction model is obtained. S6: Set target performance requirements, use the converged concrete performance prediction model as a fitness function evaluator, and use an optimization algorithm to search for urban bridge construction parameters that meet the target performance requirements to obtain the optimal combination of construction parameters. 2.The deep learning-based urban bridge renewal construction parameter optimization method of claim 1, wherein, Multiple sets of urban bridge construction parameters and corresponding measured values ​​of concrete performance were obtained to form a historical dataset, including: Multiple sets of construction parameters were collected during the urban bridge renovation construction process. These construction parameters included at least the water-cement ratio, sand ratio, admixture dosage, water-reducing agent dosage, pouring temperature, curing humidity, surface roughness grade of old concrete, and interface agent type. For each set of construction parameters, the corresponding measured values ​​of concrete performance are obtained through on-site experiments or quality inspection reports. The measured values ​​of concrete performance include at least 3-day compressive strength, 28-day compressive strength, bond strength between new and old concrete, slump, heat of hydration, internal and external temperature difference, and shrinkage strain. Each set of construction parameters is paired with the corresponding measured values ​​of concrete performance to form a data sample. All data samples are combined to form a historical dataset. 3.The deep learning-based urban bridge renewal construction parameter optimization method of claim 2, wherein, A deep neural network is trained using the historical dataset to obtain a concrete performance prediction model, including: The construction parameters and measured concrete performance values ​​in the historical dataset are normalized to obtain normalized construction parameters and normalized measured concrete performance values. Construct a multi-layer feedforward deep neural network, which includes an input layer, multiple hidden layers and an output layer, with each hidden layer followed by a batch normalization layer and a dropout layer; Normalized construction parameters are used as input features, and normalized measured concrete performance values ​​are used as output labels. The multilayer feedforward deep neural network is trained in a supervised manner until it is verified to converge, thus obtaining a trained deep neural network. The trained deep neural network is then used as a concrete performance prediction model. 4.The deep learning-based urban bridge renewal construction parameter optimization method of claim 1, wherein, In the construction parameter space, the prediction uncertainty of the concrete performance prediction model for each combination of construction parameters is calculated, and the region of high uncertainty parameters is screened out, including: The construction parameter space is determined based on the value range of each construction parameter; Uniform sampling is performed within the construction parameter space to obtain multiple candidate combinations of construction parameters; For each candidate combination of construction parameters, the candidate combination of construction parameters is input into the concrete performance prediction model multiple times, and the output of some neurons in the concrete performance prediction model is randomly discarded each time it is input, so as to obtain multiple predicted output values. Calculate the standard deviation of multiple predicted output values ​​corresponding to each candidate construction parameter combination, and use the standard deviation as a measure of the prediction uncertainty of the candidate construction parameter combination; Calculate the arithmetic mean of the prediction uncertainty measure for all candidate construction parameter combinations, and use the arithmetic mean as the first dynamic threshold. The region covered by candidate construction parameter combinations whose prediction uncertainty metric is higher than the first dynamic threshold is marked as a high uncertainty parameter region. 5.The deep learning-based urban bridge renewal construction parameter optimization method of claim 4, wherein, For each candidate combination of construction parameters, the candidate combination of construction parameters is input into the concrete performance prediction model multiple times, and the outputs of some neurons in the concrete performance prediction model are randomly discarded each time, resulting in multiple predicted output values, including: After each hidden layer of the concrete performance prediction model, a dropout layer used during training is retained, and the dropout layer is kept active during the testing phase. The dropout layer only temporarily blocks the output of some neurons during forward propagation and does not change the network structure of the concrete performance prediction model. For the same candidate combination of construction parameters, the forward propagation calculation is repeated a preset number of times, and the discard layer randomly discards a different proportion of neurons during each forward propagation to obtain a preset number of predicted output values; The preset number of predicted output values ​​are used as multiple predicted output values.

6. The method for optimizing urban bridge construction parameters based on deep learning according to claim 1, characterized in that, Within the high uncertainty parameter region, new urban bridge construction parameters and their corresponding simulated concrete performance values ​​are generated based on the parameter update model, forming a synthetic sample set, including: A conditional generative adversarial network is constructed as a parameter generation model, wherein the conditional generative adversarial network includes a generator and a discriminator; Using the construction parameters and corresponding measured values ​​of concrete performance in the historical dataset as training samples, the parameter generation model is pre-trained to obtain the trained parameter generation model. The region identifier of the high uncertainty parameter region and the corresponding urban bridge construction feature code are input into the trained parameter generation model to generate multiple new construction parameter combinations. The urban bridge construction feature code includes at least the surface roughness level of old concrete, the type of interface agent, and the degree of spatial constraint of urban bridge construction. Multiple new construction parameter combinations are sequentially input into the concrete performance prediction model, and the corresponding concrete performance prediction values ​​are output. The multiple new construction parameter combinations are paired with the corresponding concrete performance prediction values ​​as concrete performance simulation values. A composite sample set is formed by combining all the new construction parameters and their corresponding simulated concrete performance values.

7. The method for optimizing construction parameters for urban bridge upgrades based on deep learning according to claim 1, characterized in that, Repeat steps S2 to S4 until the updated concrete performance prediction model converges across the entire construction parameter space, resulting in a converged concrete performance prediction model, including: After each completion of step S4, step S2 is re-executed within the construction parameter space to calculate the prediction uncertainty measure of the current updated concrete performance prediction model for all candidate construction parameter combinations in the entire construction parameter space. Calculate the maximum value of the prediction uncertainty measure for all candidate construction parameter combinations, and use the maximum value as the current maximum prediction uncertainty; The maximum prediction uncertainty is compared with a preset convergence threshold. When the maximum prediction uncertainty is less than the convergence threshold, it is determined that the updated concrete performance prediction model has converged, and the current updated concrete performance prediction model is used as the converged concrete performance prediction model. When the maximum prediction uncertainty is greater than or equal to the convergence threshold, return to step S3, continue to generate new synthetic samples and update the concrete performance prediction model again.

8. The method for optimizing construction parameters for urban bridge updates based on deep learning according to claim 1, characterized in that, Setting target performance requirements, and using the converged concrete performance prediction model as a fitness function evaluator, an optimization algorithm is used to search for urban bridge construction parameters that meet the target performance requirements, resulting in the optimal combination of construction parameters, including: Set target performance requirements, which include target values ​​or target ranges for various concrete performance indicators. Among them, the concrete performance indicators include at least 3-day compressive strength, 28-day compressive strength, bond strength between new and old concrete, slump, heat of hydration, internal and external temperature difference, and shrinkage strain. The converged concrete performance prediction model is used as a fitness function evaluator, and the combination of construction parameters is used as the input of the fitness function evaluator. The deviation between the predicted concrete performance value output by the converged concrete performance prediction model and the target performance requirement is used as the fitness value. With the minimum fitness value as the optimization objective, the particle swarm optimization algorithm is used to search for the optimal combination of construction parameters in the construction parameter space.

9. The method for optimizing construction parameters for urban bridge updates based on deep learning according to claim 8, characterized in that, Using the minimum fitness value as the optimization objective, the particle swarm optimization algorithm is employed to search for the optimal combination of construction parameters within the construction parameter space, including: Initialize a particle swarm, which contains multiple particles, each corresponding to a combination of construction parameters. The position of each particle is randomly initialized within the construction parameter space, and the velocity of each particle is randomly initialized within a preset velocity range. The construction parameters corresponding to the position of each particle are input into the converged concrete performance prediction model, the predicted concrete performance value of each particle is output, and the fitness value is calculated in combination with the target performance requirements. The current fitness value of each particle is compared with the historical best fitness value of that particle. If the current fitness value is less than the historical best fitness value, the current position of the particle is updated to the individual best position of that particle. The position corresponding to the smallest fitness value among all the individual optimal positions of all particles is taken as the global optimal position, and the fitness value corresponding to the global optimal position is taken as the global optimal fitness value. According to the preset velocity update rules and position update rules, the velocity and position of each particle are updated, and the particles are guided to gather towards the individual optimal position and the global optimal position until the change of the global optimal fitness value in multiple consecutive iterations is lower than the preset change threshold or the preset maximum number of iterations is reached. The combination of construction parameters corresponding to the final globally optimal position is output as the optimal combination of construction parameters.

10. A deep learning-based system for optimizing construction parameters for urban bridge updates, characterized in that: The system is used to implement the deep learning-based urban bridge renovation construction parameter optimization method according to any one of claims 1-9, the system comprising: The initial model training module is used to obtain multiple sets of urban bridge construction parameters and corresponding measured values ​​of concrete performance to form a historical dataset, and to train a deep neural network using the historical dataset to obtain a concrete performance prediction model. The uncertainty assessment module is used to calculate the prediction uncertainty of the concrete performance prediction model for each combination of construction parameters in the construction parameter space, and to screen out the parameter regions with high uncertainty. The synthetic sample generation module is used to generate new urban bridge construction parameters and their corresponding concrete performance simulation values ​​based on the parameter update model within the high uncertainty parameter region, thus forming a synthetic sample set. The incremental model training module is used to add the synthetic sample set to the historical dataset and retrain the concrete performance prediction model to obtain an updated concrete performance prediction model. The iterative convergence judgment module is used to repeatedly call the uncertainty assessment module, the synthetic sample generation module, and the model incremental training module until the updated concrete performance prediction model is verified to converge in the entire construction parameter space, and the converged concrete performance prediction model is obtained. The optimal parameter search module is used to set target performance requirements, use the converged concrete performance prediction model as a fitness function evaluator, and search for urban bridge construction parameters that meet the target performance requirements through optimization algorithms to obtain the optimal combination of construction parameters.

Images