Prediction method and system for prestress release loss value based on machine learning

Abstract

The invention belongs to the technical field of machine learning and pre-stress, and discloses a pre-stress release loss value prediction method and system based on machine learning, and the method comprises the steps: carrying out the multi-working-condition modeling and simulation of a pre-stress beam through finite element numerical software, and extracting the working parameters and design parameters of the pre-stress beam, carrying out data preprocessing, distribution check and feature importance analysis to obtain an initial data set; dividing the initial data set into an initial training data set and an initial test data set, and processing the initial training data set and the initial test data set to obtain a processed training data set and a processed test data set; constructing a full-connection multi-layer perceptron neural network model, defining training, verification and monitoring functions, training the full-connection multi-layer perceptron neural network model by using the processed training data set, and testing the trained model by using the processed test data set to obtain a prediction model; real parameters of the prestressed beam are obtained, the prediction model is used for predicting the prestress release loss value, and a prediction result is obtained.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of machine learning and prestress, and particularly relates to a prestress loss value prediction method and system based on machine learning. BACKGROUND

[0002] Prestress technology is a key technology in modern concrete structure engineering, which is widely used in important infrastructures such as long-span bridges, high-rise buildings, industrial plants, sleepers and nuclear power plant safety shells. The core principle of prestress technology is to apply a tensile force to the prestressed reinforcement (such as steel strand, steel wire or threaded steel bar) in advance, and then transfer the force to the concrete member after anchoring, so as to artificially establish pre-compressive stress before the structure bears external load. This pre-compressive stress can effectively offset the tensile stress caused by live load, wind load or earthquake action in the use stage, significantly improving the anti-cracking performance, stiffness and long-term durability of the structure. However, in the key link of prestress construction, the tension release process (i.e. when the tensioning equipment is released and the tensile force of the prestressed reinforcement is transferred to the concrete member), inevitable prestress loss occurs. This loss is mainly caused by the following factors: slip deformation and wedge back of anchor and backing plate; elastic compression deformation, creep and shrinkage effect of concrete; stress relaxation of prestressed reinforcement; temperature change, tension sequence deviation or friction loss, etc. These factors are coupled with each other, resulting in significant uncertainty of the final tension loss value (i.e. the difference between the initial tension and the actual effective prestress).

[0003] In engineering practice, the accurate prediction of the loss value of prestress tensioning is directly related to the safety and economy of the structure. If the predicted value is too small, it may lead to insufficient actual prestress, causing structural cracking or deformation beyond the limit; if the predicted value is too large, it will cause material waste or overload of the tensioning equipment. At present, the industry generally uses the theoretical formula provided in the specification to estimate, that is, based on elastic mechanics, the loss is divided into instantaneous loss (such as anchor deformation, elastic compression of concrete) and long-term loss (such as creep, shrinkage, relaxation), and the analytical solution is derived through simplifying assumptions (such as linear material, uniform stress distribution). However, the non-homogeneity of concrete, the spatial compression and time axis changes involved in environmental temperature and humidity, construction process fluctuations (such as uneven tensioning rate) and other factors in actual engineering will introduce nonlinear effects, making the prediction results of the specification formula deviate from the measured value by 15% to 30%. In view of this, some studies attempt to improve the prediction accuracy by introducing correction coefficients or combining finite element numerical simulation, but these methods are still limited by the empirical nature of parameter calibration or the complexity of calculation. In recent years, with the development of sensor technology and construction informatization, a large amount of multi-source data (such as tensioning force monitoring values, concrete strength time curve, environmental temperature and humidity records) in the prestress construction process has been accumulated, providing the possibility for new prediction methods such as random forest, support vector machine, and neural network driven by data. However, it is mainly aimed at a single engineering case, and lacks verification of the generalization ability across projects.

[0004] Based on the above shortcomings, for the loss value of prestress tensioning in prestressed engineering, the present application proposes a hybrid prediction method combining finite element simulation and machine learning, which improves the prediction accuracy and model generalization ability by combining a large number of simulated samples and data-driven nonlinear compensation, and introduces relatively easy-to-measure feature engineering and model interpretation analysis to provide transparent support for engineering decision-making. SUMMARY

[0005] The present application aims to solve the problems of the prior art and provides the following solutions: A prediction method for the loss value of prestress tensioning based on machine learning, comprising the following steps: A finite element numerical software is used to model and simulate the prestressed beam under multiple working conditions, the working parameters and design parameters of the prestressed beam are extracted, and data preprocessing, distribution checking and feature importance analysis are performed to obtain an initial data set; The initial data set is divided into an initial training data set and an initial test data set and processed to obtain a processed training data set and a processed test data set; A fully connected multilayer perceptron neural network model is constructed, training, validation and monitoring functions are defined, the processed training data set is used to train the fully connected multilayer perceptron neural network model, and the processed test data set is used to test the trained model to obtain a prediction model. Real parameters of the prestressed beam are acquired, prestress loss value prediction is performed based on the real parameters by using the prediction model, and a prediction result is obtained.

[0006] Preferably, the method for obtaining the initial data set comprises: A finite element model is established, and the size and the prestressed tendon height parameter of the prestressed beam are traversed; The cross-sectional area, centroid, moment of inertia and deflection characteristics of the prestressed beam in different working conditions are extracted, and the extracted parameters are dimensionally eliminated to obtain preprocessed data; The preprocessed data is subjected to data distribution inspection and calculation of characteristic linear correlation, and n MLP feature importance analysis is performed by using a perturbation method to obtain the initial data set. repeat

[0007] Preferably, the method for obtaining the processed training data set and the processed test data set comprises: The initial data set is proportionally divided by using a function to obtain the initial training set and the initial test data set; train test split The initial training set and the initial test data set are subjected to standardization processing to obtain the standardized initial training set and the standardized initial test data set; The standardized initial training set and the standardized initial test data set are subjected to tensor conversion, and a function is introduced; TensorDataset and DataLoader The converted training data set is shuffled to obtain the processed training data set and the processed test data set.

[0008] Preferably, the method for obtaining the prediction model comprises: A fully connected multilayer perceptron neural network model is constructed, including a feature input layer, a hidden layer and an output layer; A double-coupling composite activation function combining LeakyReLU a function and SiLU a function is defined, and a forward propagation function is defined; A mean square error loss function MSELoss , an optimizer AdamW and a dynamic learning rate scheduling strategy ReduceLROnPlateau are introduced; The hidden layer combines batch normalization BatchNorm1d and hierarchical uneven neuron deactivation technology, and the fully connected multilayer perceptron neural network model is optimized. ​​​The processed training data set is used to train the optimized full-connection multi-layer perception neural network model, and the processed test data set is used to test the trained model, so as to obtain the prediction model.

[0009] Preferably, the LeakyReLU function is: , wherein, max represents the maximum value of the two, x represents the input value of the function, α represents the negative slope coefficient. The SiLU function is: , wherein, σ represents Sigmoid the function.

[0010] Preferably, the mean square error loss function MSELoss is: , wherein, n represents the total number of samples, i represents the i sample, represents the true value, represents the model prediction value. The optimizer AdamW is: , wherein, η represents the learning rate, represents the bias-corrected gradient first moment, represents the bias-corrected gradient second moment, λ represents the weight decay coefficient, ϵ represents the batch normalization numerical stability term. θ t represents the t step parameter. The dynamic adjustment learning rate scheduling strategy ReduceLROnPlateau is: , wherein, represents the scheduled new learning rate, represents the lower limit value of the learning rate, represents the current learning rate, factor represents the decay parameter.

[0011] Preferably, the batch normalization BatchNorm1d is: , , wherein, represents the l normalized value, represents the l input value, represents the current batch mean, σ batch represents the current batch variance, represents the final output value of batch normalization, γ represents the training learnable scaling parameter, β represents the training learnable shift parameter.

[0012] The application provides a prestress tension loss value prediction system based on machine learning, which applies the above prediction method and comprises a data acquisition module, a data division module, a model construction module and a prediction module. The data acquisition module uses finite element numerical software to model and simulate prestressed beams under multiple working conditions, extracts working parameters and design parameters of the prestressed beams, and performs data preprocessing, distribution checking and feature importance analysis to obtain an initial data set. The data division module is used to divide the initial data set into an initial training data set and an initial test data set and perform processing to obtain a processed training data set and a processed test data set. The model construction module is used to construct a fully connected multilayer perceptron neural network model, define training, verification and monitoring functions, train the fully connected multilayer perceptron neural network model using the processed training data set, test the trained model using the processed test data set, and obtain a prediction model. The prediction module is used to obtain real parameters of a prestressed beam, predict the prestress tension loss value based on the real parameters using the prediction model, and obtain a prediction result.

[0013] Compared with the prior art, the application has the following beneficial effects: The application provides a prestress tension loss value prediction method based on machine learning, which realizes high-precision prediction of prestress engineering tension loss values by combining finite element simulation and deep learning technology. In addition, the application provides comprehensive error evaluation indicators and intuitive visualization analysis tools to assist in judging the prediction reliability and facilitate users to evaluate the reliability of the prediction results, thereby providing a scientific basis for engineering decision-making. By optimizing tension design and controlling over-limit risks, the application reduces material waste and construction rework while ensuring structural safety, significantly improves the economic efficiency and construction efficiency of the project, and has a relatively wide application prospect. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a flowchart of data preprocessing in an embodiment of the present invention; Figure 3 This is a schematic diagram of the feature importance ranking results in an embodiment of the present invention; Figure 4 This is a schematic diagram of the training process monitoring in an embodiment of the present invention, wherein, Figure 4 (a) in the diagram is a schematic of the training / validation loss. Figure 4 (b) in the diagram is a schematic of learning rate scheduling. Figure 4 (c) in the diagram is a schematic diagram of the gradient norm change (first 5). Figure 4 (d) in the diagram represents the final gradient distribution (first 5). Figure 5 The determination coefficient R in the embodiments of the present invention 2 Fraction diagram; Figure 6 This is a schematic diagram of the model prediction result error in an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1: In this embodiment, as Figure 1 As shown, a method for predicting prestress release loss based on machine learning includes the following steps: S1. Finite element numerical software is used to model and simulate prestressed beams under multiple working conditions, extract the working parameters and design parameters of prestressed beams, and perform data preprocessing, distribution checking and feature importance analysis to obtain the initial dataset.

[0019] The method for obtaining the initial dataset includes: establishing a finite element model and iterating through the dimensions and prestressed tendon height parameters of the prestressed beam; extracting the cross-sectional area, centroid, moment of inertia, and deflection characteristics of the prestressed beam under different working conditions, and eliminating dimensional differences in the extracted parameters to obtain preprocessed data; performing data distribution checks and calculating the linear correlation of features on the preprocessed data, and then using... n _ repeat The perturbation method is used to perform feature importance analysis in MLP to obtain the initial dataset.

[0020] In this embodiment, considering the large amount of diverse data required for universality and generalization, a finite element model was established using the finite element numerical software ANSYS MECHANICAL ENTERPRISE 19.0 to simulate the prestressing tensioning and release process. For generalization modeling, the beam dimensional parameters, namely beam length... L and H Perform a traversal to determine the fixed position height of the prestressing tendons at the support section. The process involves iterating through the prestress loss values ​​after release according to the percentage of control stress. For formulas involving prestress friction loss, the coefficients are adjusted. μ With coefficient k Perform a traversal. Considering on-site measurements and design references, select to extract the mid-span cross-sectional area. Centroid of the mid-span section Moment of inertia at mid-span section The beam mid-span deflection under various working conditions after simulation calculation Δy .since Pandas The library introduces based on Numpy of DataFrame After creating a tabular data structure, corresponding regular expressions are written according to the data format of each result file. The paths to each result file are specified, and features are extracted row by row into a DataFrame for temporary storage. Corresponding data index labels are then created, and warnings for empty files / non-matching data are added, along with checks on the number of processed samples and data distribution. After extraction, feature engineering arrays and prediction label arrays are created according to the data index labels described above. Shape checks, variable statistical descriptions, and linear correlation analysis between features and labels, as well as feature MLP importance analysis, are added. The top 50% of features with the highest linear correlation are mid-span deflection. Δy Centroid of the mid-span section Moment of inertia at mid-span section Prestressing tendon height The specific data preprocessing flowchart is as follows: Figure 2As shown; the specific feature importance ranking results heatmap is as follows. Figure 3 As shown.

[0021] Before performing feature importance analysis in MLP, the original features themselves have different units and dimensions, with significant differences in numerical range. If training is performed directly on them using importance analysis at different scales, the model may incorrectly assign higher weights to certain features simply due to differences in numerical magnitude. Furthermore, large-scale differences in data will lead to significant differences in gradient update magnitudes between different features, causing the optimizer to adjust its strategy in multiple directions, ultimately resulting in difficulty in model convergence and predictions that are purely numerical and detached from reality. To enable data-driven model learning that is independent of data collection units and only reflects the physical influence of parameters, and to eliminate the interference of dimensional differences on the model itself and evaluation metrics to ensure that importance scores objectively reflect the actual physical influence of features, the aforementioned feature array is modified... Z - Score Standardization. Adoption. n _ repeat The perturbation method establishes an MLP regression model consistent with subsequent training, specifying the activation function basis as... relu Classes, optimizers adam Class, initial learning rate 0.001, maximum iterations 200 n _ repeat With a repeatability of 10 and a random seed of 42, the characteristic of the number of repeatability counts after randomization is calculated—the mean of model importance—and a heatmap is plotted for visualization.

[0022] The standardized functions involved are as follows: , in, Z This represents the standardized value. x 0 represents the original feature value. Represents the characteristic mean. s It represents the characteristic standard deviation.

[0023] The activation function is as follows: , in, x This represents the input value of the function.

[0024] The optimizer is as follows: gradient: ; First-order moment estimation: ; Second-order moment estimation: ; Deviation correction: , ; Update parameters: ; where, θ t denotes the t step parameter, denotes the loss function, denotes the t step first moment estimate (mean), denotes the first decay rate threshold (usually 0.9), denotes the t step second moment estimate (uncentered variance), denotes the second decay rate threshold (usually 0.999), denotes element-wise squaring, denotes the bias-corrected gradient first moment, denotes the bias-corrected gradient second moment, η denotes the learning rate, ϵ denotes the batch normalization numerical stability term (a small constant to prevent division by zero).

[0025] The importance is calculated as follows: , where, denotes the j th feature X j importance to the model prediction, denotes the performance value of the model on the original data (e.g., R 2 squared or MSE index, etc.), denotes the j th feature X j perturbed dataset generated by shuffling the values (i.e., breaking their relationship with the target) the perturbed performance value.

[0026] For example, given an existing baseline prediction , perturb the x 1st or x 2nd feature n _ repeat times with random perturbations such as random sampling replacement or adding different types of noise, etc., while keeping the rest of the features unchanged to obtain the perturbed input ; calculate the new prediction value on the perturbed sample y i the degree of change in the prediction value, i.e., the importance of the corresponding perturbed feature . The importance of the remaining features is calculated in the same way, i.e., by calculating the degree of change in the prediction after multiple random perturbations of a certain feature to measure the importance of the feature to the prediction.

[0027] S2. Divide the initial dataset into an initial training dataset and an initial test dataset and process them to obtain a processed training dataset and a processed test dataset.

[0028] Methods for obtaining processed training and test datasets include: utilizing train _ test _ split The function proportionally partitions the initial dataset to obtain an initial training set and an initial test set; it then standardizes the initial training set and the initial test set to obtain standardized initial training set and standardized initial test set; finally, it performs tensor transformation on the standardized initial training set and the standardized initial test set and introduces... TensorDataset and DataLoader The transformed training dataset is shuffled to obtain the processed training dataset and the processed test dataset.

[0029] In this embodiment, (1) data partitioning: to facilitate the subsequent use of a dataset independent of the training process to avoid overfitting verification, assist in hyperparameter tuning, and avoid data leakage, etc., self sklearn In the library model _ selection Introduced in China train _ test _ split Dataset partitioning function, defining the proportion of the test set. test _ size =0.2, random seed 42, proportionally divide the original dataset. (2) Data standardization: Use Z-Score standardization in step S1 to avoid possible slow convergence, gradient vanishing or gradient explosion, model over-dependence, training overfitting and training over-oscillation. Based on the above, standardize the features and labels respectively (the training set fits the scaler, and the test set only applies the same scaling parameter), and output them together with the standardizer object for subsequent repetitive calls. (3) PyTorch tensor transformation: In order to fit the MLP neural network architecture built later and unify the computation framework, optimize the batch computation, storage and training performance of the model itself, integrate the built-in functions of torch and the basic configuration of the network layer, apply mathematical support such as differential compatibility and automatic gradient calculation, obtain the standardized data in the previous text and convert it into PyTorch tensors, and flatten the label data according to the output requirements of the regression task. (4) Data loader: In order to accelerate batch loading, out-of-order and multi-process to improve training efficiency, code conciseness and compatibility and functional extensibility within possible complex requirements, the data loader is automatically loaded. torch Inside the library utils Below data Library import TensorDataset and DataLoaderTwo functions, the former is used to combine the processed tensor array and make it correct, and the latter sets the batch size, supports random shuffling of the training set, and keeps the order of the test set to ensure that the model learning result is influenced by pure data characteristics, physical meaning and distribution law, etc. and discards the numerical value significance of the data itself.

[0030] S3. Construct a fully connected multilayer perceptron neural network model, define training, verification and monitoring functions, train the fully connected multilayer perceptron neural network model using the processed training data set, and test the trained model using the processed test data set to obtain a prediction model.

[0031] The method for obtaining the prediction model comprises: constructing a fully connected multilayer perceptron neural network model, including: a feature input layer, a hidden layer and an output layer; defining a double-coupling composite activation function combining a LeakyReLU function and a SiLU forward propagation function; introducing a mean square error loss function MSELoss , an optimizer AdamW and a dynamic learning rate scheduling strategy ReduceLROnPlateau ; combining batch normalization BatchNorm1d and hierarchical uneven neuron deactivation technology in the hidden layer to optimize the fully connected multilayer perceptron neural network model; training the optimized fully connected multilayer perceptron neural network model using the processed training data set, and testing the trained model using the processed test data set to obtain a prediction model.

[0032] In this embodiment, an MLP model architecture is established. With the help of contents including but not limited to fully connected layer structure, batch normalization BatchNorm1d , double-coupling activation function combining SiLU and LeakyReLU , neuron deactivation layer Dropout decreasing layer by layer as the dimension increases, etc., nonlinear mapping from feature input to prediction output is realized, and then a prediction continuous value is output for the regression task. Among them, for the deactivation function, uses LeakyReLU to ensure to avoid neuron death and alleviate negative zone gradient disappearance while providing simple and efficient nonlinearity to reduce computational cost and improve convergence speed, uses SiLU a smooth non-monotonic activation function near the high-dimensional attention mechanism and the output prediction layer, which uses its own self-gating characteristics and more stable gradient to learn more complex patterns or rules. After the alternation of the former LeakyReLU and the latter LeakyReLU , the initial hard activation of SiLU and the high-dimensional soft activation of BatchNorm1d balance the model calculation cost and learning performance. Batch normalization DropoutTo enhance the stability of training while avoiding overfitting / underfitting phenomenon. Considering the regression task and learning efficiency, combined use AdamW Optimizer, MSELoss Mean square error loss function and specific loss index monitoring learning rate scheduling strategy ReduceLROnPlateau . Involved between layers Linear Linear transformation as follows: , Where, y The output vector is represented by, v The input vector is represented by, The weight matrix is represented by, b The bias vector is represented by.

[0033] Batch normalization BatchNorm1d For: , , Where, The normalized value is represented by, l The input value is represented by, The current batch mean is represented by, l batch The current batch variance is represented by, The final output value of batch normalization is represented by, σ The training learnable scaling parameter is represented by, The training learnable shift parameter is represented by. γ β

[0034] LeakyReLU The function is: , Where, max The maximum of the two is represented by, x The input value of the function is represented by, α The negative slope coefficient (usually 0.01) is represented by; SiLU The function is: , Where, σ The Sigmoid Function is represented by.

[0035] Dropout As follows (only effective during training, not effective during validation and testing): , Where, m Bernoulli mask (0 or 1); p The drop probability is represented by, that is​​Dropout ( p The value varies from 0.1 to 0.5 depending on the specific requirements.

[0036] Mean squared error loss function MSELoss for: , in, n This represents the total number of samples. i Indicates the first i One sample, Represents the actual value. This represents the model's predicted value.

[0037] Optimizer AdamW for: , in, λ This represents the weight decay coefficient; Dynamically adjust learning rate scheduling strategy ReduceLROnPlateau for: , in, This represents the new learning rate after scheduling. This represents the lower limit of the learning rate. This represents the current learning rate. factor This represents the attenuation parameter.

[0038] above, AdamW The optimizer uses L2 regularization, which is independent of gradient updates and is tuned by a single parameter, thus avoiding the limitations of traditional optimization methods. Adam The coupling problem between weight decay and adaptive learning rate in the optimizer; a unique approach is used for regression problems. MSELoss The mean squared error loss function penalizes large errors more strongly, making the model pay more attention to significant biases, while its continuous differentiability facilitates gradient calculation at a point and makes it consistent with the mean squared error. AdamW The optimization process has better compatibility; ReduceLROnPlateau The learning rate scheduling strategy automatically reduces the learning rate when validation loss stagnates, rather than manually adjusting hyperparameters. This phased reduction of the learning rate helps the model fine-tune hyperparameters and converge towards a better solution when it gets stuck in a potentially local optimum. Combining these strategies can reduce training costs, improve model generalization ability, efficiently optimize the model's learning structure, and enhance robustness. Furthermore, the output results are processed after the model outputs. flatten The flattening operation is designed to suit the needs of gradient calculation and autonomous differentiation during the training process.

[0039] With the constructed MLP architecture, forward propagation, loss function, optimizer and learning rate scheduling strategy, define the training loop function to realize batch gradient descent, define the validation loop function to evaluate the performance of the model on the test set and close the gradient calculation to save resources, define the training monitoring function to record the training loss, validation loss, learning rate change, gradient norm and distribution and other indicators during the training process, define the inverse normalization function to restore the prediction results to the original dimension in subsequent prediction. Based on the loss value loss training, save the best training model, data standardizer, call the monitoring visualization part after the end of the training process to perform result chart visualization processing and save. The specific training process indicator monitoring diagram is shown in Figure 4 .

[0040] For the training loop function, specify the network mode as the training mode, and initialize the training loss value and process monitoring history label value to 0 before training, clear the optimizer model parameter gradient to prevent gradient accumulation in back propagation and ensure the parameter transmission direction, according to the forward propagation, training loss calculation, loss back propagation, dynamic gradient clipping, optimizer parameter update, monitoring label recording, loss accumulation calculation, this training process is cycled. The formulas related to back propagation and dynamic gradient clipping not mentioned in the foregoing are as follows: Back propagation is as follows: Chain rule: ; Output layer ( MSE ): ; Wherein, θ represents the model parameter, N represents the data batch size, represents the prediction output, represents the true value corresponding to the prediction output.

[0041] Dynamic gradient clipping is as follows: , , Wherein, represents the current gradient, represents the manually set dynamic clipping hyperparameter threshold, which is assigned to min (3.0, 1.0+0.05t) in this embodiment, t represents the training round, represents the L2 norm of the gradient, represents the clipped gradient.

[0042] The complete training process is obtained by initialization (zeroing) - training for each batch iteration - gradient zeroing - forward propagation - back propagation - dynamic gradient clipping - parameter update - convergence or early stopping. For the early stopping mechanism, the given improvement patience counter is 10, i.e. the validation loss of the model is not improved within 10 iteration rounds (the current loss is greater than the best loss each time, the counter is incremented by 1 until the upper limit), the best validation loss is fixed as the model performance evaluation and the loop is exited. The training progress is printed once every 20 rounds to stabilize the update log output, and additional prompts are given when the learning rate changes.

[0043] For the validation function, the network mode is specified as the validation mode, the model gradient calculation and parameter update are disabled after the validation loss is initialized, the model is called to calculate the validation loss value using the previous loss function, and the average validation loss is returned.

[0044] For the training loop, the total number of training rounds is given as 200, the best loss, early stopping counter, training loss and validation loss are initialized, the learning rate is initialized, the monitoring function and monitoring indicators are initialized. Traverse all training rounds, call the training function module and validation function module in turn and give the hyperparameters to calculate the training loss of the current round and the validation loss of the current round. Update the monitoring indicator record, update the learning rate according to the validation loss. After returning the validation loss, i.e. the learning rate calls the validation loss dynamic adjustment, return the current learning rate and perform monitoring comparison to determine whether to update and print prompts, enable early stopping judgment for the validation loss and update the patience counter simultaneously. After the early stopping judgment passes or the training is completed, call visualize to visualize the changes in the training process monitoring indicators.

[0045] For the inverse normalization function, users or downstream systems need the results in the original scale to make more intuitive judgments, and evaluation indicators such as MAE also need to be calculated in the original scale to have meaning. The normalized model output prediction results are inverse normalized and inverse transformed to restore the prediction values to the scale of the original data. The inverse normalization is as follows: , where, denotes the original value recovered by inverse normalization, denotes the normalized value, denotes the standard deviation of the original data, denotes the mean of the original data.

[0046] In summary, the training of the MLP deep learning model including data standardization, model training, validation, monitoring and early stopping mechanism is completed. It ensures the stability of the training and the optimization of the training efficiency while providing monitoring tools to help debug the model performance and assist in model optimization, path-based saving of reusable models and standardizers facilitates deployment and inference, and ensures consistency between production environment and test environment.

[0047] The test data set in the loading step S2 is loaded and prediction is performed using the trained model. Pathing calls the standardizer consistent with the training process to perform prediction after standardization and returns the prediction result. The prediction result is converted to the original dimension scale by inverse standardization, and then the multiple evaluation indexes are calculated, including the determination coefficient R 2 The three indexes of score, mean absolute error MAE index and root mean square error RMSE index are used to quantitatively evaluate the prediction performance (prediction bias), and at the same time, the ideal line is superimposed on the real value and the prediction value to quickly judge the model system bias and generate the prediction error distribution histogram for intuitive display of the prediction effect. Based on the above definition of the prediction function and encapsulation, the user is prompted to input different features in order, and the encapsulated prediction function is called to realize the interactive deployment of the prediction returning the prediction value. The specific real value / prediction value ideal fitting scatter plot is as shown in Figure 5 The specific prediction error distribution histogram is as shown in Figure 6 The determination coefficient R 2 score is as follows: , The mean absolute error MAE index is as follows: , The root mean square error RMSE index is as follows: , Wherein, represents the mean value of the real value.

[0048] The above, after importing the necessary dependent library, the original data set is loaded and the standardizer is called for standardization, the prediction is performed on the test set under the premise of disabling the gradient, and the inverse standardization is performed on the prediction value and the real value respectively, the return result is obtained, and the different prediction evaluation indexes are calculated, and then the relevant visualization is performed to show the distribution. At the end of the above, the interactive deployment is added, and the whole is packaged and encapsulated as a prediction function to provide a convenient, reusable interface for the user to input parameters for process simplification prediction.

[0049] S4. Obtain the real parameters of the prestressed beam, and based on the real parameters, use the prediction model to predict the prestressed tension loss value, and obtain the prediction result.

[0050] In the embodiment, the machine learning framework dependency library, numerical calculation library, file loading library and the previously defined related modules are imported, the above are defined as resource loading functions and the evaluation mode is enabled. After calling the loading function in the integration module, the output is integrated and interactively deployed, and the end-to-end prediction process system is encapsulated. Compared with repetitive training, the dependency file model weight.pth and the standardizer.pkl after single training ensure the reproducibility of automatic and fast prediction. The preprocessing part corresponding to model loading and data processing is hidden to reduce the weight of deployment while ensuring portability and robustness, reducing learning memory occupation and computing cost. To a certain extent, it reduces the prediction technology requirements and use difficulty, that is, user-friendly. Then, the real parameters of the prestressed beam are obtained, the prestressed loss value prediction is performed by using the prediction model, and the prediction result is obtained.

[0051] Embodiment two: In the embodiment, a prestressed loss value prediction system based on machine learning includes a data acquisition module, a data division module, a model construction module, and a prediction module.

[0052] The data acquisition module uses finite element numerical software to model and simulate the prestressed beam under multiple working conditions, extracts the working parameters and design parameters of the prestressed beam, and performs data preprocessing, distribution inspection and feature importance analysis to obtain an initial data set. The data division module is used to divide the initial data set into an initial training data set and an initial test data set and perform processing to obtain a processed training data set and a processed test data set. The model construction module is used to construct a fully connected multilayer perceptron neural network model, define training, validation and monitoring functions, train the fully connected multilayer perceptron neural network model using the processed training data set, and test the trained model using the processed test data set to obtain a prediction model. The prediction module is used to obtain real parameters of the prestressed beam, based on the real parameters, to perform prestressed loss value prediction using the prediction model, and to obtain a prediction result.

[0053] The above-described embodiments are only descriptions of the preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A method for predicting prestress loss value based on machine learning, characterized in that, The method comprises the following steps: A finite element numerical software is used to model and simulate the prestressed beam under multiple working conditions, working parameters and design parameters of the prestressed beam are extracted, data preprocessing, distribution checking and feature importance analysis are performed, and an initial data set is obtained; The initial data set is divided into an initial training data set and an initial test data set and processed to obtain a processed training data set and a processed test data set; A fully connected multilayer perceptron neural network model is constructed, training, validation and monitoring functions are defined, the processed training data set is used to train the fully connected multilayer perceptron neural network model, and the processed test data set is used to test the trained model to obtain a prediction model; Real parameters of the prestressed beam are obtained, and the prediction model is used to predict the prestressed tension loss value based on the real parameters to obtain a prediction result. 2.The method of claim 1, wherein, The method for obtaining the initial data set comprises: A finite element model is established, and the size and prestressed tendon height parameters of the prestressed beam are traversed; The cross-sectional area, centroid, moment of inertia and deflection features of the prestressed beam in different working conditions are extracted, and the extracted parameters are dimensionless to obtain preprocessed data; performing data distribution check, calculating feature linear correlation on the pre-processed data, and performing MLP feature importance analysis by using perturbation method to obtain the initial data set. n repeat performing data distribution check, calculating feature linear correlation on the pre-processed data, and performing MLP feature importance analysis by using perturbation method to obtain the initial data set.​ 3.The method of claim 1, wherein, The method for obtaining the processed training data set and the processed test data set comprises: Utilizing train test split a function divides the initial data set into proportions to obtain the initial training set and the initial test data set;​​ The initial training set and the initial test data set are standardized to obtain a standardized initial training set and a standardized initial test set; tensor transforming the standardized initial training set and the standardized initial test set, and introducing TensorDataset with DataLoader shuffling the transformed training data set to obtain the processed training data set and the processed test data set. 4.The method of claim 1, wherein, The method for obtaining the prediction model comprises: A fully connected multilayer perceptron neural network model is constructed, including a feature input layer, a hidden layer and an output layer; Definitions LeakyReLU Functions and SiLU Dual-coupled composite activation functions, and forward propagation functions; Introducing mean squared error loss function MSELoss , optimizer AdamW and dynamic adjustment of learning rate scheduling strategy ReduceLROnPlateau ; The hidden layer incorporates batch normalization BatchNorm1d The fully connected multilayer perceptron neural network model is optimized with a hierarchical heterogeneous neuron inactivation technique; The processed training data set is used to train the optimized fully connected multilayer perceptron neural network model, and the processed test data set is used to test the trained model to obtain the prediction model. 5.The method of claim 4, wherein, The LeakyReLU The function is: ， wherein max denotes the maximum of both, x denotes the input value of the function, α denotes the negative slope coefficient; The SiLU The function is: ， wherein σ represents Sigmoid function. 6.The method of claim 4, wherein, The mean squared error loss function MSELoss is: ， wherein, n denotes the total number of samples, i denotes the i-th sample, i denotes the i-th sample, denotes the true value, denotes the model predicted value; The optimizer AdamW Is: ， wherein, η denotes the learning rate, denotes the bias-corrected first moment of the gradient, denotes the bias-corrected second moment of the gradient, λ denotes the weight decay coefficient, ϵ denotes the batch normalization numerical stability term, θ t denotes the parameters of the t step. The dynamic adjustment learning rate scheduling strategy ReduceLROnPlateau is: ， wherein, denotes the scheduled new learning rate, denotes the learning rate lower bound value, denotes the current learning rate, factor denotes the decay parameter. 7.The method of claim 6, wherein, The batch normalization BatchNorm1d Is: ， ， wherein, represents the normalized value of the first input value, l represents the normalized value of the second input value, l represents the current batch mean value, σ batch represents the current batch variance value, represents the final output value of the batch normalization, γ represents a trainable learnable scale parameter, β represents a trainable learnable shift parameter.​​ 8. A machine learning-based prediction system of prestress loss value, the prediction system applying the prediction method of any one of claims 1-7, characterized in that, It comprises: A data acquisition module, a data division module, a model construction module and a prediction module; The data acquisition module uses a finite element numerical software to model and simulate the prestressed beam under multiple working conditions, extracts working parameters and design parameters of the prestressed beam, and performs data preprocessing, distribution checking and feature importance analysis to obtain an initial data set; The data division module is used to divide the initial data set into an initial training data set and an initial test data set and process them to obtain a processed training data set and a processed test data set; The model construction module is configured to construct a fully connected multilayer perceptron neural network model, define training, verification and monitoring functions, train the fully connected multilayer perceptron neural network model by using the processed training data set, test the trained model by using the processed test data set, and obtain a prediction model; The prediction module is configured to obtain real parameters of the prestressed beam, perform prestress loss value prediction by using the prediction model based on the real parameters, and obtain a prediction result.

Images