A method for predicting cable force loss of a cable net curtain wall based on a neural network

Abstract

The application provides a method for predicting cable force loss of a cable net curtain based on a neural network, data of existing cables is collected first, a data set of the existing cables is constructed, input features of the existing cables are extracted according to the data set of the existing cables, then a neural network is constructed, the neural network is trained with the data set of the existing cables after the construction of the neural network is completed, then the trained neural network is used to predict a cable force loss rate of a target cable, when the cable force loss rate of the target cable is predicted, data of the target cable is collected first and input features of the target cable are extracted, the input features of the target cable are input into the neural network, the cable force loss rate of the target cable is predicted through the neural network, then the cable force of the target cable at a future time is back calculated, so that the cable force of the target cable at the future time is accurately predicted, and the structural safety and stability of the cable net curtain are ensured.

Description

Technical Field

[0001] This invention relates to the field of cable tension prediction technology, and specifically to a method for predicting cable tension loss in cable-stayed curtain walls based on neural networks. Background Technology

[0002] As the core load-bearing component of cable-net curtain walls, the stability of the cable tension directly determines the safety and seismic performance of the curtain wall. During long-term service, curtain wall cables are susceptible to stress loss due to environmental loads such as temperature and wind, material aging, and fatigue effects. This stress can lead to deformation of the cable-net curtain wall, a decline in its wind and seismic resistance, and certain safety hazards. Therefore, accurately predicting the stress loss of curtain wall cables and establishing a damage model for cable-net curtain walls to predict curtain wall performance is of great significance for ensuring building structural safety and developing reasonable maintenance strategies.

[0003] Currently, methods for predicting cable stress loss mainly include theoretical analysis, numerical simulation, and experimental research. Theoretical analysis, based on the principles of materials mechanics and structural mechanics, derives the cable stress loss law by establishing mechanical models and material loss models; however, this method struggles to fully consider the complex nonlinear effects of environmental loads. Experimental research can study cable stress failure laws by setting various stress conditions, but experiments are costly in terms of both economy and time, and it is difficult to reproduce real stress conditions. Numerical simulation methods are highly dependent on parameter settings, and existing research is insufficient to provide enough key parameters. With the development of artificial intelligence technology, machine learning methods, especially neural networks, have been applied to cable stress prediction due to their powerful nonlinear fitting capabilities; however, there are still few methods specifically for predicting cable stress loss in cable-net curtain walls, failing to meet the requirements of wide applicability and high prediction accuracy. This invention provides a method for predicting cable stress loss in cable-net curtain walls based on neural networks to solve the above problems. Summary of the Invention

[0004] This invention provides a method for predicting the cable force loss of cable-stayed curtain walls based on neural networks. It utilizes measured data of the cable force of the curtain wall, takes the cable force loss rate as the prediction target, selects a reasonable feature extraction method for environmental load time series data, and leverages the efficient fitting and prediction capabilities of neural networks for complex nonlinear data to achieve efficient and accurate prediction of the cable force of cable-stayed curtain walls.

[0005] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows: A method for predicting cable stress loss in cable-stayed curtain walls based on neural networks includes the following steps: S1: Collect data, extract input features, and construct the dataset; Collect data on cable-net curtain wall testing projects, including: curtain wall design data, cable stress testing data, curtain wall type, main structure type, and environmental impact data; S11, Obtain the raw data for each original cable; The raw data includes cable parameters, time parameters, environmental action history data, initial cable force, and measured cable force; The cable's parameters include its diameter d. i Elastic modulus E i Length L i and cable specification representative coefficient γ i ; The time parameter is the service interval Δt. i =t i1 -t i0 Unit: days, t i0 For the design cable force or construction completion time, t i1 This is the deadline for the testing project; Environmental action history data include Δt i Temperature history t during the period ti Unit: °C, and wind history f fi Unit: m / s; Initial cable force l i0 For t i0 The moment-in-time cable force, in kN, is the design cable force or the actual measured cable force at the moment of construction completion, obtained from the curtain wall design data or the completion acceptance data. Detecting cable tension i1 For t i1 The service cable force at the moment of inspection, in kN, is the cable force measured at the end of the inspection project. S12, Calculate the cable force loss rate η of the original cable. i ; The original cable tension loss rate η i As a relative value index, the calculation formula is: η i =(l i0 -l i1 ) / l i0 ×100%, Where η i Let be the cable force loss rate of the i-th original cable; S13, Statistical data are extracted from environmental action history data to obtain temperature statistical characteristics and wind statistical characteristics; Temperature statistical characteristics include t i0 Temperature T at time i0 t i1 Temperature at time T i1 Temperature difference T is used to account for the influence of temperature changes on cable tension test results. diff =Ti1 -T i0 Maximum temperature T MAX The average temperature T reflects the average temperature level. avg Temperature range T, reflecting the range of temperature fluctuations range Temperature standard deviation T, which reflects the degree of temperature fluctuation. std and the maximum daily temperature difference T that reflects the degree of temperature change. maxdiff Temperature range T range Determined by subtracting the lowest temperature from the highest temperature; Wind statistical characteristics include the mean wind force V, which reflects the average intensity of wind load. avg Wind range V, which reflects the range of wind load fluctuations. range The standard deviation of wind force (V) reflects the degree of fluctuation in wind load. std and the maximum wind speed V, which reflects the intensity of extreme wind loads. max Extremely poor wind force V range It is obtained by subtracting the minimum wind speed from the maximum wind speed; Temperature history t ti Wind power history f fi Obtained from the meteorological information center; S14 integrates the cable's own parameters, time parameters, temperature statistical characteristics, and wind statistical characteristics to form the input features X of the i-th single cable. i :X i =[l i0 , Δt i T i0 T avg T diff T MAX T range T std T maxdiff V avg V range V std V max d i E i L i γ i ]; Input features X of the original cable i Adjustments can be made based on correlation analysis and neural network training optimization; Input features X of the original cable i and the corresponding original cable force loss rate η i A single sample is formed, and all the original samples of the LAC form a dataset; S2, preprocessing the dataset, including... S21, Outlier handling; S22, Dataset partitioning; S23, normalization processing; S3, Design the neural network model structure, train the neural network model and suppress overfitting, and verify the performance of the neural network model; S4. Based on the trained neural network model, predict the cable force loss rate of the target cable and calculate the cable force of the target cable at future moments; S41, Collect relevant data on the target cable and extract the input features of the target cable; The relevant data for the target cable include the initial cable force l aim,0 The time interval Δt of the prediction period aim Cable diameter d aim Elastic modulus E aim Length L aim , as well as the temperature and wind history during the predicted period; Extract statistical features of temperature and wind force, and integrate them to form the input feature X of the target cable. aim ; S42, using the maximum value X of the original cable's input features saved in step S23. t,max and minimum value X t,min The input feature X of the target cable aim Normalization is performed to obtain the normalized input feature X' of the target cable. aim The normalized input feature X' aim The input is fed into the trained optimal neural network model to obtain the normalized predicted cable force loss rate η' of the target cable. r,aim ; The maximum value X of the input feature of the original cable t,max and minimum value X t,min Obtained during dataset preprocessing; Using the maximum value η of the original cable force loss rate saved in step S23 t,max and minimum value η t,min The predicted value of the cable force loss rate η' for the target cable. r,aim Inverse normalization yields the actual cable force loss rate η of the target cable. r,aim , The inverse normalization formula is η r,aim =η' r,aim ×(η t,max -η t,min )+η t,min ; The maximum cable force loss rate η of the original cable t,max and minimum value η t,min Obtained during dataset preprocessing; S43, using the actual cable force loss rate η of the target cable r,aimDeducing the cable force at future moments of the target cable; The formula for calculating the inverse force of the target cable at future moments is: l aim，f =l aim，0 ×(1-η r，aim ), Among them l aim，f For the target cable future moment t f The cable force, l aim，0 The initial cable force of the target cable is obtained from the curtain wall design data or measurements, η. r，aim This is the actual cable force loss rate of the target cable calculated in step S42.

[0006] Furthermore, in step S21, the specific operation for outlier handling is as follows: using box plot method or 3σ criterion to identify outlier samples in the dataset, removing the identified outlier samples, and then filling the small number of missing values ​​with the mean. Abnormal samples include cable loss rate η i Samples with >30% or negative values, temperature outlier samples, and wind outlier samples; The negative samples are those where the cable force value has increased compared to the design value or the value at completion.

[0007] Furthermore, in step S22, when dividing the dataset, the preprocessed samples are randomly divided into training set, validation set and test set in a ratio of 6:2:2. The training set accounts for 60% and is used for iterative learning of the model weight parameters; The validation set comprises 20% and is used to monitor overfitting during training and determine the optimal training epochs; it does not participate in training. The test set accounts for 20% and is used to evaluate the model's generalization ability; it is not used for training.

[0008] Furthermore, in step S23, the specific operation of normalization is as follows: The input feature X of each original cable in the dataset is... i and the corresponding original cable force loss rate η i Synchronous mapping to the [0,1] interval, the normalization formula is: X' i =(X i -X t,min ) / (X t,max -X t,min ), η' i =(η i -η t,min ) / (η t,max -η t,min ); Where X' i X represents the input features of the original cable after normalization. iX is the input feature of the original cable. t,max and X t,min η' represents the maximum and minimum values ​​of the input features of the original cable, respectively. i η is the normalized original cable tension loss rate. t,max and η t,min These are the maximum and minimum values ​​of the original cable force loss rate, respectively, used for inverse normalization of subsequent prediction results.

[0009] Furthermore, in step S3, the neural network model structure includes an input layer, a hidden layer, and an output layer; The neural network model structure is preferably designed with 3 layers, namely 1 input layer + 1 hidden layer + 1 output layer, to avoid the gradient vanishing problem in multi-layer networks; The input layer has one layer with no activation function, directly passing the preprocessed input features. The number of neurons in the input layer is equal to the vector dimension of the input features. The original input features X of the Lasso are... i The vector has 17 dimensions, corresponding to 17 neurons; The hidden layer has 1 layer and n neurons. h Using the empirical formula n h =(n i +n o ) 0.5 +a is determined, where n i n is the number of neurons in the input layer. i =17, n o n is the number of neurons in the output layer. o =1, where a is a constant from 1 to 10, calculate the number of neurons n in the hidden layer. h (17+1) 0.5 +5, n h The value is 10; the ReLU function is used as the activation function for the hidden layer. Number of neurons in the output layer n o =1 corresponds to the normalized cable force loss rate of the single-value output; the output layer activation function is a linear function.

[0010] Furthermore, in step S3, the neural network model is trained and overfitting is suppressed to ensure the model's generalization ability. The specific operations are as follows: S321, Configure training parameters: For the regression prediction task of cable loss rate, the mean squared error (MSE) is used as the loss function, with the formula Loss=(1 / N)∑ N i=1 (η' r,i -η' i ) 2 , Where N is the sample size, η' r,iη' is the normalized predicted value of the original cable tension loss rate. i The normalized original cable tension loss rate; Adam was used as the optimizer; The initial learning rate was set to 0.001-0.01, and the maximum number of iterations was set to 200-1000. S322, Calculate the predicted value through forward propagation: The input features X of the original cable in the training set... i After normalization, the normalized original cable input features X' are obtained. i The normalized original cable input feature X' i The data is fed into a neural network, and the normalized predicted value η' of the original cable force loss rate is calculated layer by layer in the output layer. r,i ; Backpropagation updates model weights and biases: Calculate the loss function values ​​of the predicted and true values, and update the weights and bias parameters of the hidden and input layers in reverse using gradient descent, which is implemented by the Adam optimizer; Validation set monitoring: After each training round, the loss value is calculated using the validation set; S33 combines early stopping and L2 regularization to suppress overfitting; The specific implementation of early stopping is as follows: after each round of training, the input feature vector of the validation set is input into the model. If the loss value of the validation set does not decrease for 10-20 consecutive rounds, training is stopped immediately and the current optimal model parameters are saved. The specific implementation of L2 regularization is as follows: a weight penalty term is added to the loss function, and the regularized loss function is: Loss reg =Loss+λ∑ w w 2 λ is the regularization coefficient, ranging from 0.0001 to 0.001, and w is the model weight parameter, which restricts the weight from being too large. If the sample size is insufficient, the temperature and wind statistical features are randomly perturbed within ±5% to generate virtual samples for data augmentation.

[0011] Furthermore, in step S3, the performance of the trained neural network model is evaluated using a test set. The evaluation metrics include root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 ; , , , Where η' r,i η' is the normalized predicted value of the original cable tension loss rate. i η' is the normalized original cable tension loss rate.m,i This represents the average cable force loss rate of the original cables after normalization. The performance evaluation criteria for neural network models are: R 2 A model that meets the criteria of having an accuracy of ≥0.85 and an RMSE of ≤1.5% can be used for prediction in practical engineering projects.

[0012] Furthermore, in step S3, if the performance of the neural network model does not meet the standard, adjustments are made by modifying the structure of the neural network model or by adjusting the input feature vectors to make the performance of the neural network model meet the standard. During the adjustment, the first step is to adjust the structure of the neural network model. If the performance still fails to meet the standard after the adjustment, then the input feature vector is adjusted.

[0013] Furthermore, the specific operation for adjusting the structure of the neural network model is as follows: first determine the current number of layers in the neural network model. If it is a 3-layer structure, then adjust it to a 4-layer structure. If it is already a 4-layer structure, then optimize the hidden layer parameters, including adjusting the number of hidden layer neurons or changing the hidden layer activation function. After adjustment, retrain and validate the neural network model. When adjusting the number of neurons in the hidden layer, the empirical formula n h =(n i +n o ) 0.5 Adjust the value of 'a' based on +a, with an adjustment range of 1-15; When changing the activation function of the hidden layer, use compatible variants of the ReLU function, including Leaky ReLU, PreLU, and ELU functions.

[0014] Furthermore, when the number of layers in the neural network model is adjusted from 3 to 4, the structure of 1 input layer + 1 hidden layer + 1 output layer is adjusted to a structure of 1 input layer + 2 hidden layers + 1 output layer.

[0015] The beneficial effects of this invention are as follows: Using the relative cable force loss rate as the prediction target and comprehensively considering various characteristics of the cable itself as input features, such as elastic modulus, cable specifications, and cross-section, it can be applied to the prediction of various types of cables and has a wide range of applications. Statistical feature extraction is used to extract environmental load history data such as temperature and wind force time series data, transforming high-dimensional time series data into low-dimensional fixed-dimensional statistical features. This avoids the dimensionality explosion problem caused by directly inputting the original time series data, while retaining the core impact information of environmental load. Combined with the nonlinear fitting capability of neural networks, it can accurately capture the complex mapping relationship between cable parameters, environmental load and cable force loss rate, improve prediction accuracy, and achieve high-precision accurate prediction. The input features are parameters that are easy to collect in daily maintenance operations, such as initial cable force, dimensional parameters, temperature, and wind force. This effectively reduces the difficulty of data acquisition and increases the amount of data. Furthermore, the model can quickly and accurately predict the cable force loss rate and back-calculate the future cable force, ensuring the safety of the building structure. By employing multiple methods such as outlier handling, normalization, early stopping, and L2 regularization, the training stability of the model was effectively improved, the overfitting problem was suppressed, and the model was ensured to have good generalization ability on the new specification cable, thus having better applicability. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram illustrating the model performance evaluation results of a specific embodiment of the present invention. Detailed Implementation

[0017] The technical solutions in 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.

[0018] In the description of this invention, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0019] The specific principle of this invention is as follows: First, data on existing cables is collected to construct a dataset of existing cables. Input features of existing cables are extracted from the dataset. Then, a neural network is constructed. After the neural network is constructed, it is trained using the dataset of existing cables. The trained neural network can fit the input features of the cables with the cable force loss rate. Then, the trained neural network is used to predict the cable force loss rate of the target cable. When predicting the cable force loss rate of the target cable, data on the target cable is first collected and the input features of the target cable are extracted. The input features of the target cable are input into the neural network. The cable force loss rate of the target cable is predicted by the neural network. Then, the cable force of the target cable at future moments is deduced. This allows for accurate prediction of the cable force of the target cable at future moments, ensuring the structural safety and stability of the cable net curtain wall. When constructing a neural network, the neural network model structure is designed first. Then, the neural network model is trained using a dataset, and overfitting is suppressed. Finally, the performance of the neural network model is validated using the dataset. If the performance of the trained neural network model fails to meet the judgment criteria, the neural network structure needs to be adjusted. If the performance of the adjusted neural network meets the judgment criteria, it is used to predict the cable force loss rate of the target cable. If the performance of the adjusted neural network still fails to meet the judgment criteria, the input feature vector is adjusted to make the neural network usable. When predicting the cable force of the target cable in the future, the predicted value of the cable force loss rate of the target cable is first obtained through the optimal neural network model. Then, the predicted value of the cable force loss rate of the target cable is inversely normalized to obtain the actual cable force loss rate of the target cable. Finally, the actual cable force loss rate of the target cable is used to deduce the cable force of the target cable in the future.

[0020] like Figure 1 , 2 As shown in the figure, a specific embodiment of the present invention takes the cable net glass curtain wall cable of a commercial complex as the research object and uses the prediction method disclosed in the present invention to predict the cable force loss.

[0021] S1: Collect inspection data for the curtain wall of the commercial complex, including: curtain wall design data, cable stress test data, curtain wall type, main structure type and environmental impact data. Among them, the cable stress test data comes from 150 sets of cable test data or artificially extended data of the curtain wall. Build the dataset and extract input features: S11: Obtain the raw data for each existing cable: Existing cable parameters include diameter d i 30-36 mm, elastic modulus E i 196 GPa, length L i It is 10-15 m; Time parameter Δt i It lasts for 1095 days, for a total of 3 years; Environmental load history data are temperature history data over 3 years. ti Temperature range of -5 to 35°C and wind speed history f fi The speed is 1-10.8 m / s; Initial cable force l i0 ; Detecting cable tension i1 All cables are high-vanadium cables, and the cable specification representative coefficient γ i Take 1; S12: Calculate the cable loss rate η i ; Among them, l i0 The value range is 32-155 kN, l i1 The value ranges from 21 to 101 kN, and the calculated cable force loss rate η i The value range is 1.0%-36.0%; S13: Extracting statistical features: Since only the detection data of a single building was selected and the cable force detection time was the same, the statistical features of environmental effects are the same for all cables; S14: Forming Input Features: X i =[l i0 ,1095,d i ,l i1 ,196,L i ,1]; S2: Dataset preprocessing: S21: Outlier Handling: The 3σ criterion is used to identify outlier samples, and three η values ​​are found. i Of the samples with >30% or negative values, two temperature outlier samples were removed, leaving 145 valid samples. S22: Dataset partitioning: 145 samples were divided into a training set of 87 samples, a validation set of 29 samples, and a test set of 29 samples in a 6:2:2 ratio; S23: Normalization process: Normalize the input feature X i and cable force loss rate η i Mapped to the [0,1] interval, X values ​​of each feature are stored. t,max X t,min and η t,max =32%, η t,min =1.0%; S3: Design the neural network model structure, train the neural network model and suppress overfitting, and verify the performance of the neural network model; Design the neural network model structure: 7 neurons in the input layer, and the number of neurons in the hidden layer is determined by the empirical formula (7+1). 0.5+5=8 is determined, the number of neurons in the output layer is 1; the activation function of the hidden layer is ReLU, and the output layer is a linear function; the neural network model structure is: input layer (7) → hidden layer (8) → output layer (1); S32: Model Training and Overfitting Suppression S321: Configure training parameters: loss function is MSE, optimizer is Adam, initial learning rate is 0.005, maximum number of iterations is 500; S322: Training process: Forward propagation calculates the predicted value, back propagation updates the weights and biases, and the loss is monitored using the validation set after each training round. S33: Overfitting suppression: Early stopping method is used: if the loss does not decrease after 15 consecutive rounds of validation set training, training is stopped and L2 regularization is used: regularization coefficient λ=0.0005; When the validation set loss reaches its minimum after 180 training iterations, training is stopped and the model parameters are saved. S34: Model Performance Validation: Evaluate model performance using 29 samples in the test set, such as... Figure 2 As shown, RMSE=0.85%, MAE=0.62%, and R0.05 were obtained. 2 =0.91, satisfying R 2 Engineering-practical standards with a strength ≥0.85 and RMSE ≤1.5%; S4: Target Cable Prediction: Three new cables in the curtain wall of this commercial complex are being replaced. The cable material is the same as the original cables, but the elastic modulus and cross-section differ. Using the neural network model trained in step S3, the cable stress loss after one year is predicted. After temperature correction, the stress loss rates are obtained as 1.5%, -4.3%, and 5.1%, respectively. Based on the above data, the curtain wall is modeled and analyzed to examine its stress and deformation under accidental strong winds or earthquakes.

[0022] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for predicting cable force loss in cable-stayed curtain walls based on neural networks, characterized in that, Includes the following steps, S1: Collect data, extract input features, and construct the dataset; S11, Obtain the raw data for each original cable; The raw data includes cable parameters, time parameters, environmental action history data, initial cable force, and measured cable force; The cable's parameters include its diameter d. i Elastic modulus E i Length L i and cable specification representative coefficient γ i ; The time parameter is the service interval Δt. i =t i1 -t i0 , t i0 For the design cable force or construction completion time, t i1 This is the deadline for the testing project; Environmental action history data include Δt i Temperature history t during the period ti Wind power history f fi ; Initial cable force l i0 For t i0 Constant tension; Detecting cable tension i1 For t i1 The service load at all times; S12, calculate the cable force loss rate of the original cable; The formula for calculating the cable force loss rate of the original cable is: η i =(l i0 -l i1 ) / l i0 ×100%, Where η i Let be the cable force loss rate of the i-th original cable; S13, Statistical data are extracted from environmental action history data to obtain temperature statistical characteristics and wind statistical characteristics; Temperature statistical characteristics include t i0 Temperature at time T i0 t i1 Temperature at time T i1 Temperature difference T diff =T i1 -T i0 Maximum temperature T MAX Average temperature T avg Temperature range T range Temperature standard deviation T std and maximum daily temperature difference T maxdiff ; Wind statistical characteristics include mean wind speed V avg Extremely poor wind force V range Wind standard deviation V std and maximum wind speed V max ; S14 integrates the cable's own parameters, time parameters, temperature statistical characteristics, and wind statistical characteristics to form the input feature X of a single original cable. i :X i =[l i0 , Δt i T i0 T avg T diff T MAX T range T std T maxdiff V avg V range V std V max d i E i L i γ i ]; The dataset consists of all the input features of the original cables and the corresponding cable force loss rates of the original cables; S2, preprocessing the dataset, including: S21, Outlier handling; S22, Dataset partitioning; S23, normalization processing; S3, Design the neural network model structure, train the neural network model and suppress overfitting, and verify the performance of the neural network model; S4. Based on the trained neural network model, predict the cable force loss rate of the target cable and calculate the cable force of the target cable at future moments; S41, Collect relevant data on the target cable and extract the input features of the target cable; S42, after normalizing the input features of the target cable, input them into the trained neural network model to obtain the normalized predicted cable force loss rate η'. r,aim And the predicted value of cable loss rate η' r,aim Inverse normalization yields the actual cable force loss rate η of the target cable. r，aim ; S43, use the actual cable force loss rate of the target cable to deduce the cable force of the target cable at future moments; The formula for calculating the inverse force of the target cable at future moments is: l aim，f =l aim，0 ×(1-η r，aim ), Among them l aim，f For the target cable future moment t f The tension at that time, l aim，0 Let η be the initial cable force of the target cable. r，aim The actual cable force loss rate of the target cable.

2. The method for predicting cable force loss of cable-stayed curtain walls based on neural networks according to claim 1, characterized in that, In step S21, the specific operation for outlier handling is as follows: identify outlier samples in the dataset, remove the identified outlier samples, and then fill the missing values ​​with the mean. Abnormal samples include cable loss rate η i Samples with >30% or negative values, temperature outlier samples, and wind outlier samples.

3. The method for predicting cable force loss of cable-stayed curtain walls based on neural networks according to claim 1, characterized in that, In step S22, when splitting the dataset, the preprocessed samples are randomly divided into training set, validation set and test set in a ratio of 6:2:

2. The training set accounts for 60% and is used for iterative learning of the model weight parameters; The validation set accounts for 20% and is used to monitor overfitting during the training process and determine the optimal number of training rounds. The test set accounts for 20% and is used to evaluate the model's generalization ability.

4. The method for predicting cable force loss of cable-stayed curtain walls based on neural networks according to claim 1, characterized in that, In step S23, the specific operation of normalization is as follows: the input features X of the original cable are normalized. i The cable force loss rate η of the original cable i Synchronous mapping to the [0,1] interval, the normalization formula is: X' i =(X i -X t,min ) / (X t,max -X t,min ), η' i =(η i -η t,min ) / (η t,max -η t,min ); Where X' i X represents the input features of the original cable after normalization. i X is the input feature of the original cable. t,max and X t,min These are the maximum and minimum values ​​of the input features of the original cable, respectively; η' i η is the normalized original cable tension loss rate. t,max and η t,min These represent the maximum and minimum values ​​of the original cable force loss rate, respectively.

5. The method for predicting cable force loss of cable-stayed curtain walls based on neural networks according to claim 1, characterized in that, In step S3, the neural network model structure includes an input layer, a hidden layer, and an output layer; The number of neurons in the input layer is equal to the vector dimension of the input features; The hidden layers can be set to 1-2 layers, and the number of neurons can be determined by the empirical formula n. h =(n i +n o ) 0.5 +a is determined, where n i n is the number of neurons in the input layer. o n is the number of neurons in the output layer. h denoted as the number of neurons in the hidden layer, where a is a constant from 1 to 10; The output layer has 1 neuron, corresponding to the cable loss rate of a single-value output; The hidden layer activation function uses the ReLU function, and the output layer activation function uses a linear function.

6. The method for predicting cable force loss of cable-stayed curtain walls based on neural networks according to claim 1, characterized in that, In step S3, the specific operations for training the neural network model and suppressing overfitting are as follows: Configure training parameters, using mean squared error as the loss function, Adam as the optimizer, initial learning rate set to 0.001-0.01, and maximum number of iterations set to 200-1000; The normalized original cable input feature X' i Input the neural network model, calculate the predicted value through forward propagation, calculate the loss function value, update the model weights and biases through backpropagation, and combine early stopping and L2 regularization to suppress overfitting; The specific implementation of early stopping is as follows: if the validation set loss does not decrease for 10-20 consecutive rounds, training is stopped immediately and the current optimal model parameters are saved; The specific implementation of L2 regularization is as follows: a weight penalty term is added to the loss function, and the regularized loss function is: Loss reg =Loss+λ∑ w w 2 , where λ is the regularization coefficient, ranging from 0.0001 to 0.001, and w is the model weight parameter.

7. The method for predicting cable force loss of cable-stayed curtain walls based on neural networks according to claim 1, characterized in that, In step S3, the performance of the trained neural network model is evaluated using a test set. The evaluation metrics include root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 ; The performance evaluation criteria for neural network models are: R 2 ≥0.85 and RMSE ≤1.5%.

8. The method for predicting cable force loss of cable-stayed curtain walls based on neural networks according to claim 7, characterized in that, In step S3, if the performance of the neural network model does not meet the standard, adjustments are made by adjusting the structure of the neural network model or by adjusting the vectors of the input features. During the adjustment, the first step is to adjust the structure of the neural network model. If the performance still fails to meet the standard after adjustment, then the input feature vector is adjusted.

9. The method for predicting cable force loss of cable-stayed curtain walls based on neural networks according to claim 8, characterized in that, The specific steps for adjusting the structure of a neural network model are as follows: First, determine the current number of layers in the neural network model. If it is a 3-layer structure, adjust it to a 4-layer structure. If it is already a 4-layer structure, optimize the hidden layer parameters, including adjusting the number of hidden layer neurons or changing the hidden layer activation function. After adjustment, retrain and validate the neural network model.

10. The method for predicting cable force loss of cable-stayed curtain walls based on neural networks according to claim 9, characterized in that, When the number of layers in a neural network model is adjusted from 3 to 4, the structure of 1 input layer + 1 hidden layer + 1 output layer is changed to 1 input layer + 2 hidden layers + 1 output layer.

Images