Large-span cable-stayed bridge steel anchor beam cable force optimization method and system

Abstract

The invention relates to the technical field of cable-stayed bridge steel anchor beam cable force optimization, in particular to a large-span cable-stayed bridge steel anchor beam cable force optimization method and system. The method comprises the steps of obtaining original data of a large-span cable-stayed bridge steel anchor beam; performing data preprocessing on the obtained original data of the large-span cable-stayed bridge steel anchor beam; capturing a long-distance error dependency relationship by utilizing Transform and predicting an accumulative error of a future segment; using the convolutional LSTM network to generate an optimal adjustment amount based on the accumulative error; and solving an optimal cable force scheme by using a cable force optimization algorithm. According to the method, the precision limitation of a traditional installation process is thoroughly changed by constructing an intelligent prediction model and a multi-dimensional closed-loop control system. Based on the error prediction technology of multi-source data fusion, the nonlinear law of error transmission can be accurately captured, the error development trend is pre-judged in advance, active compensation is implemented, and the installation deviation of the steel anchor beam in the transverse bridge direction, the longitudinal bridge direction and the elevation direction is remarkably reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of cable force optimization for steel anchor beams of cable-stayed bridges, and in particular to a cable force optimization method and system for steel anchor beams of long-span cable-stayed bridges. Background Art

[0002] In the construction of long-span cable-stayed bridges, steel anchor beams serve as the key connecting components between the cables and the towers. Their installation accuracy and optimized cable tension directly impact the bridge's overall mechanical performance, service life, and safety. Traditional steel anchor beam installation and cable tension control technologies rely primarily on manual measurement, empirical judgment, and static analysis. As bridge spans continue to increase and construction precision requirements increase, existing technologies are gradually exposing significant limitations.

[0003] The existing technology mainly includes the following aspects:

[0004] Steel anchor beam installation technology: The traditional method uses manual measurement combined with simple machinery to locate and adjust the steel anchor beam. Coordinate data is obtained through equipment such as total stations, and the adjustment amount is determined based on the construction personnel's experience. This results in unstable installation accuracy and low efficiency.

[0005] Cable tension control technology: Static optimization methods are often used to determine cable tension based on finite element forward analysis. Dynamic factors and uncertainties in the construction process are not fully considered. Cable tension adjustment often requires multiple trial calculations, which is time-consuming and labor-intensive.

[0006] Error prediction and compensation technology: mainly relies on linear extrapolation or simple statistical models to predict errors, which makes it difficult to capture complex nonlinear error transmission laws and has poor cumulative error control effects.

[0007] Data processing and decision-making technology: Data collection is scattered and lacks systematic integration. The decision-making process relies on manual experience, making it difficult to achieve multi-objective optimization and real-time feedback control.

[0008] However, traditional installation technology relies on manual measurement and experience-based adjustments, which makes it difficult to overcome the influence of the high-altitude working environment (such as wind speed and temperature changes), resulting in unstable installation accuracy.

[0009] Accumulated measurement errors: Single-point measurements using total stations and other equipment lack multi-source data fusion, leading to cumulative errors that accumulate with each segment, ultimately exceeding the allowable range. For example, using traditional methods, the cumulative elevation error of six steel anchor beams can reach 8mm, far exceeding the 1.8mm target of this invention.

[0010] Low adjustment efficiency: Manual judgment of the adjustment amount lacks scientific model support and often requires repeated adjustments. The installation period of a single steel anchor beam is as long as 9 days, while the present invention can shorten it to 6.5 days.

[0011] Poor environmental adaptability: There is a lack of effective compensation measures for the influence of environmental factors such as temperature and wind speed. For example, when the temperature changes by 20°C, the displacement fluctuation of the tower in the traditional scheme reaches 4.1mm, which is 127.8% higher than that of the present invention.

[0012] In addition, the cable force optimization method lacks dynamic adaptability. It uses forward analysis to determine the cable force of the completed bridge, without considering the changes in structural stiffness and nonlinear effects during the construction phase, resulting in a large deviation between the actual cable force and the design value. The cable force deviation rate of the traditional trial method reaches 4.2%, while the present invention can control it within 1.8%. Cable force adjustment relies on manual trial calculations and finite element verification. A single cable adjustment takes up to 240 minutes, and has poor convergence, requiring 8-12 iterations. The multi-objective optimization of the present invention only requires 3-5 iterations and takes 45 minutes. It is difficult to take into account cable force accuracy, structural displacement and stress control at the same time, and it is often difficult to focus on one thing while neglecting the other. For example, the local stress concentration area of ​​the steel anchor beam in the traditional solution reaches 210MPa, which is close to the allowable stress of 235MPa. The present invention can control the maximum stress to 195MPa, a reduction of 7.0%.

[0013] The use of linear models such as ARIMA to predict errors is unable to learn the nonlinear laws of error evolution, and the average absolute error reaches 1.2mm. However, the Transformer model of the present invention can reduce it to 0.45mm, an improvement of 62.5%. It can only perform post-correction and cannot achieve active compensation in advance. The traditional solution predicts a cumulative error deviation of 2.6mm for the six sections of the 15th steel anchor beam, while the prediction deviation of the present invention is only 0.3mm. After advance compensation, the final cumulative error is only 1.2mm. The factory pre-assembly, site pre-assembly and on-site monitoring data are not fully integrated, the error feature extraction is not comprehensive, and it is difficult to achieve accurate prediction and compensation. Summary of the Invention

[0014] In order to solve the above-mentioned problems, the present invention provides a method and system for optimizing the cable force of steel anchor beams of a long-span cable-stayed bridge.

[0015] In a first aspect, the present invention provides a method for optimizing the cable force of steel anchor beams in a long-span cable-stayed bridge, which adopts the following technical solutions:

[0016] A cable force optimization method for steel anchor beams of a long-span cable-stayed bridge, comprising:

[0017] Obtain original data of steel anchor beams of long-span cable-stayed bridges;

[0018] Perform data preprocessing on the original data of the steel anchor beam of the long-span cable-stayed bridge;

[0019] Use Transformer to capture long-range error dependencies and predict the cumulative error of future segments;

[0020] Use convolutional LSTM networks to generate optimal adjustments based on accumulated errors;

[0021] The optimal cable tension scheme is solved using the cable tension optimization algorithm.

[0022] Furthermore, the original data of the steel anchor beam of the large-span cable-stayed bridge is obtained, including obtaining the original accuracy data of the steel anchor beam during the processing stage, judging the processing accuracy by measuring the geometric dimensions of the loose parts, and quantifying the cumulative error effect by cumulatively calculating the assembly height difference of adjacent segments; obtaining the pre-assembly data on the construction site, including the linear deviation of the steel corbel embedded plate, the contact surface gap and the height difference of the four corner points; obtaining the real-time posture data of the steel anchor beam during installation, and providing real-time feedback for precise positioning and cable force optimization; and also including cable force monitoring data, specifically including the basic parameters of the inclined cable and real-time monitoring data.

[0023] Furthermore, the obtained original data of the steel anchor beam of the large-span cable-stayed bridge are preprocessed, including solving the rotation matrix and translation vector through the least squares method, establishing a coordinate transformation relationship, eliminating the difference between the tower column coordinate system and the local coordinate system of the steel anchor beam, and ensuring that the design value and the measured value are compared under the same benchmark; outliers are eliminated from the cable tension monitoring data, and the temperature correction coefficient is used to eliminate the influence of temperature changes on the cable tension measurement.

[0024] Furthermore, the Transformer is used to capture long-distance error dependencies and predict the cumulative errors of future segments, including using the Transformer's self-attention mechanism to capture long-distance error dependencies based on preprocessed data, thereby improving the prediction accuracy of multi-segment cumulative errors. The input layer contains the cumulative errors, temperature, and wind speed features of the first five segments. The correlation weights between the features are calculated through the multi-head attention mechanism to indicate the degree of influence of the processing error of a certain segment on the subsequent segments. The position encoding gives the sequence data time series information to avoid the model confusing the error features of different segments.

[0025] Furthermore, the use of Transformer to capture long-distance error dependencies and predict the cumulative errors of future segments also includes predicting the cumulative errors of future segments based on historical errors and environmental parameters. Here, through the encoder-decoder architecture, the input features are first mapped to a high-dimensional latent space, and then the predicted values ​​are output through the feedforward network. Residual connections are used to avoid the vanishing gradient of deep networks, and layer normalization is used to stabilize the training process.

[0026] Furthermore, the use of a convolutional LSTM network to generate the optimal adjustment amount based on the cumulative error includes using the convolutional LSTM network to fuse spatial feature extraction and time series analysis to achieve accurate calculation of the adjustment amount of the steel anchor beam posture, wherein the convolution layer is used to extract the spatial features of the steel anchor beam posture, the LSTM layer is used to capture the time dependency of historical adjustments, and the gating mechanism is used to selectively retain key information. By learning the adjustment priority experience of the transverse bridge direction and the longitudinal bridge direction elevation, when the transverse bridge direction deviation is large, the adjustment amount in this direction is automatically calculated first.

[0027] Furthermore, the use of a convolutional LSTM network to generate an optimal adjustment amount based on cumulative errors also includes generating an optimal adjustment amount based on design deviations, prediction errors, and historical records, wherein a deep neural network DNN is used to map the multi-dimensional input design deviations, prediction errors, and historical adjustments to three-dimensional adjustment amounts, and the weights are optimized through back propagation to minimize the final deviation of the output adjustment amount, and the nonlinear activation function of the hidden layer is used to learn complex adjustment rules, including nonlinear compensation for cumulative errors.

[0028] Furthermore, the method of using the cable force optimization algorithm to solve the optimal cable force scheme includes establishing a quantitative relationship between cable force and structural displacement to provide physical constraints for cable force optimization. A simplified finite element model based on the principles of structural mechanics is constructed, and the mapping of cable force to displacement is described by the stiffness matrix. A neural network is used to approximate the nonlinear part of the finite element model, and a feasible optimization scheme is generated by seeking a balance between cable force accuracy, displacement control and structural safety. Finally, the optimal cable force scheme that meets multiple objective constraints is found through iterative calculation.

[0029] Furthermore, a feasible optimization scheme is generated by seeking a balance between cable tension accuracy, displacement control and structural safety, including constructing an objective function including cable tension deviation, displacement deviation and maximum stress, and reflecting the engineering priority through weight coefficients; finding the optimal cable tension scheme that meets multiple objective constraints through iterative calculation, wherein the Adam algorithm is combined with momentum and adaptive learning rate to efficiently search the optimal solution space, continuously adjust the cable tension vector during the iteration process to minimize the objective function, and set a convergence threshold to ensure the accuracy of the solution and avoid computational redundancy.

[0030] In the second aspect, a cable force optimization system for steel anchor beams of a long-span cable-stayed bridge is provided, comprising:

[0031] The data acquisition module is configured to acquire original data of the steel anchor beam of the long-span cable-stayed bridge;

[0032] The preprocessing module is configured to perform data preprocessing on the acquired raw data of the steel anchor beam of the long-span cable-stayed bridge;

[0033] The error module is configured to use Transformer to capture long-range error dependencies and predict the cumulative errors of future segments;

[0034] An adjustment module is configured to generate an optimal adjustment amount based on the accumulated error using a convolutional LSTM network;

[0035] The optimization module is configured to solve the optimal cable tension scheme using a cable tension optimization algorithm.

[0036] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for the method for optimizing the cable force of steel anchor beams of a large-span cable-stayed bridge.

[0037] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor as a method for optimizing the cable force of steel anchor beams of a large-span cable-stayed bridge.

[0038] In summary, the present invention has the following beneficial technical effects:

[0039] This invention radically changes the precision limitations of traditional installation processes by constructing an intelligent prediction model and a multi-dimensional closed-loop control system. Error prediction technology, based on multi-source data fusion, can accurately capture the nonlinear laws of error transmission, predict error development trends in advance, and implement active compensation, significantly reducing the installation deviation of steel anchor beams in the transverse, longitudinal, and elevation directions. Through dynamic adjustment strategies, it effectively overcomes interference from factors such as temperature and wind speed in high-altitude working environments, ensuring the consistency of installation accuracy across each segment, establishing a high-precision benchmark for subsequent segment construction, and fundamentally resolving the problem of uncontrolled error accumulation in traditional processes.

[0040] The cable tension collaborative optimization algorithm proposed in this invention breaks through the limitations of traditional static optimization methods. Through the deep integration of physical models and data-driven models, it accurately depicts the complex mapping relationship between cable tension and structural displacement and stress. The multi-objective optimization strategy can simultaneously take into account cable tension accuracy, structural displacement control and stress distribution of steel anchor beams, while ensuring cable tension uniformity and avoiding local stress concentration problems. The real-time feedback correction mechanism can quickly respond to sudden disturbances during construction, significantly shorten the cable tension adjustment cycle, and achieve a technological leap from "post-correction" to "real-time regulation", significantly improving the controllability of the construction process.

[0041] This invention uses digital technology throughout the entire process of steel anchor beam prefabrication, pre-assembly, and installation, forming an intelligent construction model with "data-driven decision-making." Dynamic matching technology is used in the pre-assembly stage to improve the on-site matching accuracy of components and reduce rework and adjustments during the installation process. During the installation stage, intelligent prediction and automated control are used to shorten the construction period of a single section and achieve efficient utilization of construction resources. This technical system not only significantly reduces labor dependence and machinery rental costs, but also improves the fatigue life and structural durability of steel anchor beams by optimizing the structural stress state, creating long-term value for the management of the entire bridge life cycle and promoting the development of large-span cable-stayed bridge construction in an intelligent and intensive direction. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1Schematic diagram of a method for optimizing cable forces in steel anchor beams of a large-span cable-stayed bridge according to Example 1 of the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described in detail below with reference to the accompanying drawings.

[0044] Example 1

[0045] Reference Figure 1 A method for optimizing cable forces in steel anchor beams of a long-span cable-stayed bridge according to this embodiment includes:

[0046] 1. Data Acquisition Phase

[0047] 1.1 Basic Data Collection Principles

[0048] 1.1.1 Principle of factory pre-assembled data collection

[0049] Core purpose: Obtain original accuracy data during the steel anchor beam processing stage to provide a benchmark for subsequent installation and identify processing error trends in advance.

[0050] The processing accuracy is judged by measuring the geometric dimensions (length / width / height) of the parts. The cumulative calculation of the height difference of adjacent segments can quantify the cumulative effect of errors. The flatness index is used to control the surface accuracy of the components to avoid stress concentration on the contact surface during installation.

[0051] Factory pre-assembly data is the basis for subsequent error prediction, and its accuracy directly affects the adjustment strategy for on-site installation. When the processing height error of a certain segment exceeds the limit, it can be corrected in advance during the on-site pre-assembly stage.

[0052] 1.1.2 Principles of Pre-Assembly Data Collection at Construction Sites

[0053] Verify whether the components are deformed during transportation and correct the matching errors caused by off-site processing.

[0054] The linear deviation of the steel corbel embedded plate reflects the overall linear consistency after assembly, the contact surface gap controls the tightness of the connection, and the height difference of the four corner points is used to assess the horizontality between segments. Through secondary pre-assembly (pre-assembling three groups each time, retaining the last group as the parent), continuous matching error control can be achieved.

[0055] Pre-assembly on site is a key step in eliminating transportation deformation. For example, when it is found that the height difference of the four corner points of a certain section of steel corbel exceeds the limit, compensation can be made by locally adjusting the elevation of the pedestal to avoid bringing errors into the installation stage.

[0056] 1.1.3 Principle of on-site installation data collection

[0057] Obtain real-time posture data during steel anchor beam installation to provide real-time feedback for precise positioning and cable force optimization.

[0058] The three-dimensional coordinates of the characteristic points of the steel anchor beam are collected using a total station, and the deviation vector is generated by comparing it with the design coordinates. The measurement of the cable guide outlet coordinates ensures the accuracy of the cable axis (deviation ≤ 3mm).

[0059] Real-time data is the "nerve endings" of closed-loop control. For example, when the measured elevation deviation of the steel anchor beam exceeds 2mm, the system can immediately trigger the jack adjustment command to avoid error accumulation.

[0060] 1.2 Principle of cable force related data collection

[0061] 1.2.1 Principle of collecting basic parameters of stay cables

[0062] Establish a database of physical properties of inclined cables to provide basic parameters for cable force calculation and optimization.

[0063] The design cable force (Fd) is the target value for optimization, the elastic modulus (Ec) and cross-sectional area (Ac) are used to calculate the cable deformation, and the linear expansion coefficient (\alphac) is used to compensate for temperature effects.

[0064] The accuracy of basic parameters directly affects the cable force optimization results. For example, a 1% deviation in the elastic modulus will result in a cable force calculation error of approximately 0.8%. Laboratory calibration is required to ensure parameter accuracy.

[0065] 1.2.2 Principles of real-time monitoring data collection

[0066] Capture the dynamic changes of cable forces and structural responses to achieve real-time feedback of cable force optimization.

[0067] Vibrating-wire pressure sensors measure the cable's vibration frequency to infer the cable force (Fm). Fiber Bragg grating sensors monitor the strain of the steel anchor beam (\varepsilonbeam) to assess local forces. The tower displacement (dtower) reflects overall structural deformation. Environmental parameters (temperature and wind speed) are used to correct measurement errors and predict structural response.

[0068] Real-time monitoring data constitutes a closed-loop feedback loop for cable force optimization. For example, when temperature changes cause fluctuations in the measured cable force values, the system can automatically activate the temperature compensation model (F\textcorr) to avoid misadjustments.

[0069] 2. Principle Description of Data Preprocessing

[0070] 2.1 Principle of Coordinate System Unification

[0071] 2.1.1 Principle of coordinate transformation model

[0072] Eliminate the differences between the tower column coordinate system and the local coordinate system of the steel anchor beam to ensure that the design value and the measured value are compared under the same benchmark.

[0073] The least squares method is used to solve the rotation matrix (\mathbfR) and the translation vector (\vect), establishing the coordinate transformation relationship (\vecP\texttower = \mathbfR\cdot\vecP\textbeam + \vect). This process is similar to "spatial calibration," ensuring consistency in data acquired by different measurement devices.

[0074] If the coordinate system is inaccurate, the positioning deviation of the steel anchor beam may be misjudged. For example, when the rotation matrix error is 1°, the position deviation of a point 10m away can reach 174mm. Therefore, it is necessary to calibrate multiple feature points (≥3 points) to improve the conversion accuracy.

[0075] 2.2 Error feature extraction principle

[0076] 2.2.1 Principle of geometric error calculation

[0077] Extract key error indicators from the original measurement data to determine whether the component meets the installation accuracy requirements.

[0078] Flatness reflects the surface roughness of a component, perpendicularity controls vertical deviation during installation, and the misalignment between adjacent segments affects connection stiffness. These metrics quantify three-dimensional spatial errors, transforming complex geometric problems into comparable one-dimensional values.

[0079] When the flatness of a segment is 1 / 1000, it indicates that its surface undulation exceeds the specification requirement (1 / 2000), and it needs to be adjusted by grinding or shimming in the pre-assembly stage to avoid additional bending moment after installation.

[0080] 2.2.2 Principles of Data Standardization

[0081] Eliminate the dimensionality differences of different parameters to make the data suitable for machine learning model training.

[0082] The original data is converted into a standard normal distribution through the mean and standard deviation to ensure that the weight distribution of each input feature of the neural network is reasonable.

[0083] If the elevation deviation (mm level) and temperature (℃ level) are not standardized, the model may focus too much on temperature changes and ignore the key posture deviation. After standardization, the influence weight of each feature can be balanced.

[0084] 2.3 Principle of cable force data preprocessing

[0085] 2.3.1 Principle of Outlier Removal

[0086] Filter sudden interference data to avoid noise affecting the accuracy of cable force optimization.

[0087] Based on the statistical 3σ principle, when the measured cable force value (Fm) deviates from the mean (\muF) by more than three standard deviations (\sigmaF), it is identified as an outlier and removed. This method assumes that normal data follows a normal distribution, and sudden disturbances (such as sensor failure) can produce extreme values.

[0088] At a certain moment, the measured value of the cable tension suddenly increased by 20%, far exceeding the historical fluctuation range. This may be caused by poor contact of the sensor cable. Eliminating it can prevent the system from misjudging it as a real cable tension change.

[0089] 2.3.2 Principle of temperature compensation model

[0090] Eliminate the influence of temperature change on cable tension measurement and obtain the real stress state.

[0091] Stay cables expand and contract due to temperature changes, causing cable tension fluctuations. The temperature correction coefficient comprehensively considers the cable material's thermal expansion and the temperature-dependent effects of its elastic modulus.

[0092] In areas with large temperature differences between day and night (e.g., 20°C), the uncompensated cable force error can reach 1.5% Fd, which may lead to deviations in the calculated force on the steel anchor beam. After temperature compensation, the error can be controlled within 0.5%.

[0093] 3. Model Construction and Algorithm Design Principles

[0094] 3.1 Principle of Transformer Error Prediction Model

[0095] 3.1.1 Network Architecture Design Principles

[0096] The Transformer's self-attention mechanism is used to capture long-distance error dependencies and improve the accuracy of multi-segment cumulative error prediction.

[0097] The input layer contains features such as the cumulative error of the first five sections, temperature, and wind speed. A multi-head attention mechanism calculates the correlation weights between these features, such as the degree to which the machining error of a particular section affects subsequent sections. Positional encoding imparts temporal information to the sequence data, preventing the model from confusing error characteristics across different sections.

[0098] Traditional ARIMA models can only capture linear time dependencies, while Transformer can learn the nonlinear laws of error evolution, such as the sudden impact of sudden temperature changes on errors, and the prediction accuracy has increased from 78% to 92%.

[0099] 3.1.2 Error prediction logic principle

[0100] Based on historical errors and environmental parameters, the cumulative errors of future segments are predicted to provide a basis for early compensation.

[0101] Through an encoder-decoder architecture, input features are first mapped to a high-dimensional latent space, and then a feedforward network outputs predicted values. Residual connections prevent vanishing gradients in deep networks, and layer normalization (\textLayerNorm) stabilizes the training process.

[0102] When it is predicted that the cumulative error in the next 6 sections will reach +4.8mm, the system can compensate 0.8mm / section in advance in the current section and control the final error within +1.2mm (such as the Wuhai Bridge case).

[0103] 3.1.3 Parameter Configuration Principle

[0104] Balance model complexity and computational efficiency to optimize prediction accuracy.

[0105] The number of attention heads (h = 4) determines the model's ability to capture multi-dimensional feature correlations, the hidden layer dimension (dh = 256) affects feature representation, and the learning rate (λ = 5e4) controls the parameter update step size. Dropout (p = 0.15) enhances the model's generalization ability and prevents overfitting by randomly deactivating neurons.

[0106] Parameter configuration needs to be optimized through cross-validation. For example, too many heads will increase the amount of calculation but the accuracy improvement will be limited. Reasonable configuration can enable the model to achieve real-time prediction on the GPU (delay ≤ 50ms).

[0107] 3.2 Principle of Neural Network Posture Adjustment Model

[0108] 3.2.1 Principle of Convolutional LSTM Network

[0109] By integrating spatial feature extraction and time series analysis, accurate calculation of the position adjustment of the steel anchor beam can be achieved.

[0110] The convolutional layer extracts spatial features of the anchor beam's posture (such as the deviation distribution of each feature point), while the LSTM layer captures the temporal dependencies of historical adjustments (such as the impact of previous segment adjustments on the current one). The gating mechanism (input gate / forget gate / output gate) selectively retains key information, such as forgetting irrelevant early adjustment records while retaining recent error trends.

[0111] The model can learn the adjustment priority experience of "transverse bridge direction → longitudinal bridge direction → elevation". For example, when the deviation in the transverse bridge direction is large, it will automatically prioritize the calculation of the adjustment amount in this direction, which is 30% more efficient than manual adjustment.

[0112] 3.2.2 Logic principle of adjustment calculation

[0113] Generate optimal adjustments based on design deviations, forecast errors, and historical records.

[0114] A deep neural network (DNN) maps multidimensional inputs (design bias, forecast error, and historical adjustments) into three-dimensional adjustments ([\Deltax,\Deltay,\Deltaz]). Backpropagation optimizes the weights so that the output adjustment minimizes the final bias. The nonlinear activation function (ReLU) in the hidden layer can learn complex adjustment rules, such as nonlinear compensation for accumulated errors.

[0115] When the design elevation deviation is +5mm and the predicted subsequent error is +2mm, the model can calculate an adjustment of +7.2mm (including compensation), which is more consistent with the nonlinear effects in actual construction than simple superposition calculation.

[0116] 3.3 Principle of cable force optimization algorithm

[0117] 3.3.1 Principle of cable force-displacement mapping model

[0118] A quantitative relationship between cable force and structural displacement is established to provide physical constraints for cable force optimization.

[0119] The simplified finite element model is based on the principles of structural mechanics, and the stiffness matrix describes the mapping from cable force to displacement; the neural network is used to approximate the nonlinear part of the finite element model, such as stiffness degradation under large deformation.

[0120] The model can predict the impact of a cable tension adjustment on the displacement of the entire bridge. For example, if the tension of a certain cable is adjusted by 10%, the displacement of the tower can be estimated to change by about 0.5 mm, thus avoiding the adjustment causing the adjacent components to exceed the limit.

[0121] 3.3.2 Principle of multi-objective optimization function

[0122] Find a balance between cable tension accuracy, displacement control, and structural safety to generate a feasible optimization solution.

[0123] The objective function (J) includes cable tension deviation, displacement deviation, and maximum stress. The weight coefficients (w1 = 0.5, w2 = 0.3, and w3 = 0.2) reflect the project priority. Constraints ensure that cable tension is within a safe range, structural stress does not exceed the limit, and single-step adjustment amounts are controllable, thus avoiding construction risks.

[0124] When the construction period is tight, the displacement control weight (w2) can be dynamically increased to prioritize ensuring that the structural displacement meets the standards within the allowable range of cable tension deviation, reflecting the flexibility of optimization.

[0125] 3.3.3 Principles of cable force optimization solution

[0126] The optimal cable tensioning solution that meets multiple objective constraints is found through iterative calculation.

[0127] The Adam algorithm combines momentum with an adaptive learning rate to efficiently search for the optimal solution space. It iteratively adjusts the cable force vector (\mathbfF) to minimize the objective function. A convergence threshold (1e4) ensures solution accuracy and avoids computational redundancy.

[0128] This algorithm can complete the cable force optimization of the entire bridge's 28 inclined cables within 5 minutes, significantly improving efficiency compared to traditional trial calculation methods (which take several hours), and the feasibility of the solution reaches 100%.

[0129] 3.4 Gradient boosting tree cable force prediction principle

[0130] 3.4.1 Principles of Ensemble Learning Model

[0131] By integrating multiple decision trees, the robustness and accuracy of cable force prediction are improved.

[0132] Gradient boosted trees (GBT) iteratively add decision trees (ht(x)). Each tree fits the residuals of the previous model, and the final prediction is the weighted sum of all trees. Regularization prevents overfitting and improves generalization.

[0133] This model can capture the nonlinear relationship between cable tension and factors such as temperature and load. For example, for every 10°C increase in temperature, the cable tension may decrease by 0.8%. However, this relationship is not strictly linear, and GBT can accurately learn through decision tree branches.

[0134] 3.4.2 Principle of cable force and temperature coupling prediction

[0135] Considering the combined effects of temperature and load, the real-time performance and accuracy of cable force prediction are improved.

[0136] Taking temperature (T), load (\textLoad), and construction stage (\textStage) as input features, each decision tree learns the cable tension variation pattern under specific conditions. For example, the temperature influence coefficient in construction stage 1 is different from that in stage 2.

[0137] During the construction stage when concrete shrinkage and creep are significant, the model can accurately predict the fluctuations in cable tension caused by changes in structural stiffness, avoiding misjudgment caused by confusion between temperature effects and structural effects.

[0138] 4. Principles of the Rope Force Optimization Execution Process

[0139] 4.1 Principle of steel anchor beam posture optimization

[0140] 4.1.1 Error Prediction Process Principle

[0141] Data preprocessing logic: Historical error standardization eliminates dimensionality effects, and temperature / wind speed normalization ensures consistency in the model input range to avoid prediction bias caused by numerical differences.

[0142] Transformer prediction logic: The self-attention mechanism automatically assigns weights to each historical error. For example, recent errors are weighted higher than long-term errors, while capturing the nonlinear effects of temperature and wind speed on errors.

[0143] Through this process, the cumulative error can be predicted 6 sections in advance, providing a basis for construction planning. For example, when the predicted error exceeds the limit, additional compensatory materials can be prepared in advance.

[0144] 4.1.2 Principle of Adjustment Calculation

[0145] Design deviation provides the basic adjustment target, prediction error supplements future compensation, and historical adjustment records avoid repeated errors. The three are integrated through the LSTMCNN model to generate an adjustment plan that includes spatial and temporal dimensions.

[0146] The model learns the nonlinear relationship of "bias adjustment amount" through training. For example, when the deviation is large, the adjustment amount needs to be appropriately amplified (to avoid under-adjustment), and when the deviation is small, the adjustment is fine-tuned (to avoid over-adjustment).

[0147] This calculation process can generate an adjustment plan that conforms to the construction process. For example, the elevation adjustment amount is converted by the support jack stroke (Lz=\frac\Deltaz\cdotL\textsupportH\textsupport) to ensure the feasibility of the adjustment.

[0148] 4.2 Principle of cable-force collaborative optimization

[0149] 4.2.1 Principle of determining target cable force

[0150] Finite element mapping ensures the physical consistency between the target cable force and the designed displacement, avoiding the infeasibility of solutions caused by pure data-driven methods; neural networks approximate complex nonlinear relationships and improve the adaptability of the model.

[0151] By allocating weight coefficients, a balance is found between cable tension accuracy, displacement control, and structural safety. For example, priority is given to ensuring that displacement meets the standard (w2 increases) during the bridge completion stage, and priority is given to ensuring that it meets the standard during the construction stage.

[0152] Specifically:

[0153] 1. Data Acquisition Phase

[0154] 1.1 Basic Data Collection

[0155] 1.1.1 Factory pre-assembled data

[0156] The geometric dimensions of the steel anchor beam parts: length L, width W, height H, the allowable deviation is ±2mm;

[0157] Height difference between adjacent segments: δh = Hi-Hi-1, cumulative height error:

[0158] Used to evaluate the cumulative effect of machining accuracy;

[0159] Component flatness: The allowed value is ≤1 / 2000.

[0160] 1.1.2 Pre-assembly data at the construction site

[0161] Linear deviation of steel corbel embedded plate: The allowed value is ≤1 / 1500, reflecting the consistency of assembly line type;

[0162] Contact surface gap: δgap=max(g1,g2,g3,g4), the allowable value is ≤0.2mm to avoid stress concentration;

[0163] Height difference of the four corner points of the steel corbel: Δh=h max -h min , the allowable value is ≤±1×nmm (n is the number of segments), which controls the cumulative elevation error.

[0164] 1.1.3 On-site installation data

[0165] 3D pose data:

[0166] The coordinates of the characteristic points of the steel anchor beam (xc, yc, zc) were collected using the TS09plus total station with an accuracy of ±1 mm;

[0167] The deviation of the cable duct outlet coordinates (xs,c,ys,c,zs,c) from the design value is ≤3mm.

[0168] 1.2 Cable force related data collection

[0169] 1.2.1 Basic parameters of the stay cable

[0170] Design cable force Fd, elastic modulus Ec = 1.95 × 105 MPa, cross-sectional area Ac;

[0171] Cable length Lc, linear expansion coefficient αc = 1.2×10-5 / ℃.

[0172] 1.2.2 Real-time monitoring data

[0173] Cable force measured value Fm: collected by vibrating wire pressure sensor, accuracy ±1% FS;

[0174] Steel anchor beam strain εbeam: using fiber Bragg grating sensor, resolution 1με;

[0175] Tower displacement dtower: real-time monitoring by total station, accuracy ±2mm;

[0176] Environmental parameters: temperature T (accuracy ±0.5°C), wind speed v (accuracy ±0.1m / s).

[0177] 2. Data Preprocessing Stage

[0178] 2.1 Coordinate system standardization

[0179] 2.1.1 Coordinate transformation model

[0180]

[0181] in, is the coordinate in the tower column coordinate system, is the local coordinate of the steel anchor beam; the rotation matrix R is solved by singular value decomposition (SVD), and the translation vector is the offset of the coordinate origin; conversion accuracy: plane position ≤1mm, elevation ≤0.5mm.

[0182] 2.2 Error feature extraction

[0183] 2.2.1 Calculation of geometric error

[0184] Flatness: Unit: mm / m, allowable value ≤ 0.5mm / m;

[0185] Verticality: Unit: mm / m, single section embedded steel plate ≤ 1mm / m;

[0186] Misalignment between adjacent segments: δ offset =|z i -z i-1 |, the allowed value is ≤0.5mm.

[0187] 2.2.2 Data Standardization

[0188]

[0189] Eliminate the dimension effect and make the input data obey the standard normal distribution N(0,1).

[0190] 2.3 Cable force data preprocessing

[0191] 2.3.1 Outlier Removal

[0192] |Fm-μF|>3σF, considered as an outlier;

[0193] The 3σ principle is used to filter the sudden interference data, where μF is the mean of the cable force and σF is the standard deviation.

[0194] 2.3.2 Temperature compensation model

[0195] F corr =F m·[1+α T (T-T0)]

[0196] Where, αT = 1.05 × 10 -5 / ℃ is the comprehensive temperature influence coefficient;

[0197] T0=20℃ is the reference temperature, and the cable tension error after compensation is ≤0.5%.

[0198] 3. Model Construction and Algorithm Design

[0199] 3.1 Transformer Error Prediction Model

[0200] 3.1.1 Network Architecture

[0201] Input layer:

[0202] X = [ΔH(n-5), ΔH(n-4), ..., ΔH(n-1), T, v, SegType] contains the cumulative error of the first five sections, current temperature, wind speed, and section type (first section / standard section);

[0203] Multi-head attention mechanism:

[0204] MultiHead(Q,K,V)=Concat(head1,...,head h )·W O

[0205] Single-head attention:

[0206]

[0207] Scaled Dot Product Attention: dk=64 is the key dimension.

[0208] 3.1.2 Error prediction formula

[0209]

[0210] Feedforward Neural Network:

[0211] FFN(x)=max(0,xW1+b1)W2+b2

[0212] dh=256 is the hidden layer dimension.

[0213] 3.1.3 Parameter Configuration

[0214] parameter definition Value Number of attention heads h＝4 4-8 Positional encoding dimension dpos=256 128-512 Learning rate λ=5e-4 1e-4-1e-3 Dropout rate p＝0.15 0.1-0.2 Prediction step length k＝6 3-10

[0215] 3.2 Neural Network Pose Adjustment Model

[0216] 3.2.1 Convolutional LSTM Network Structure

[0217] Feature extraction layer:

[0218] C t =σ(W c,i ·[h t-1 ,x t ]+W c,f ·C t-1 +b c ) The input gate It, forget gate Ft, and output gate Ot have similar structures; the convolution kernel size is 3×3, which extracts the spatial features of the steel anchor beam posture.

[0219] Status Update:

[0220] ht=Ot⊙tanh(Ct)+(1-Ot)⊙ht-1, where ht is the hidden state at time t, capturing time series dependency.

[0221] 3.2.2 Adjustment calculation model

[0222]

[0223] Input parameters:

[0224] Design deviation [Δxd, Δyd, Δzd], predicted cumulative error

[0225] Historical adjustment record Hist_adj=[Δx t-1 ,Δy t-1 ,Δz t-1 ];

[0226] DNN structure: 3 fully connected layers, number of neurons [128, 64, 3], activation function ReLU.

[0227] 3.3 Cable force optimization algorithm

[0228] 3.3.1 Cable force-displacement mapping model

[0229] Simplified finite element model:

[0230] d=K -1 F+d0

[0231] Where K is the overall stiffness matrix of the bridge (obtained through ANSYS modal analysis);

[0232] d0 is the initial displacement, which is calculated from the bridge state.

[0233] Neural Network Optimization Mapping:

[0234] F=NN(d target ,KPCA )

[0235] K PCA The parameters of the stiffness matrix after dimensionality reduction are compressed to 16 dimensions using principal component analysis (PCA).

[0236] 3.3.2 Multi-objective optimization function

[0237] min J=w1·||FF d || 2 +w2·||dd target || 2 +w3·max(σ beam )

[0238] in,

[0239] Constraints:

[0240]

[0241] Among them, the weight coefficients w1 = 0.5, w2 = 0.3, w3 = 0.2; Fmin / max is ± 20% of the design value of the cable force, ΔFs t e p =5%Fd is the upper limit of single-step adjustment.

[0242] 3.3.3 Cable force optimization solution Fopt = argminJ(F), using the Adam algorithm for iterative solution

[0243] Learning rate λ = 1e-3, number of iterations N = 200, convergence threshold 1e-4.

[0244] 3.4 Gradient Boosting Tree Cable Force Prediction

[0245] 3.4.1 Ensemble Learning Model F^(x)=∑t=1Tαt·ht(x)+β0

[0246] Where ht(x) is a decision tree of depth 3, αt is the tree weight, and β0 is the initial value;

[0247] Loss function: Regularization coefficient λ = 0.01.

[0248] 3.4.2 Cable force and temperature coupling prediction

[0249]

[0250] Considering the three factors of temperature T, load Load and construction stage; prediction accuracy: root mean square error RMSE ≤ 1.5% Fd.

[0251] 4. Cable Force Optimization Execution Process

[0252] 4.1 Optimization of steel anchor beam posture

[0253] 4.1.1 Error Prediction Process

[0254] Input data preprocessing:

[0255] Standardization of historical cumulative error ΔH(n-5:n-1);

[0256] Temperature T and wind speed v are normalized to [0,1].

[0257] Transformer predictions:

[0258]

[0259] Output the cumulative error for the next 6 sections with a confidence level ≥ 90%.

[0260] Adjustment calculation:

[0261]

[0262] That is, the transverse direction / elevation adjustment amount of the bridge, with an accuracy of ±0.5mm.

[0263] 4.2 Cable-force collaborative optimization

[0264] 4.2.1 Determination of target cable force

[0265] (1) Design displacement mapping:

[0266] F target =NN(d design ,K PCA )

[0267] d design The design displacement for the completed bridge is provided by the design institute.

[0268] (2) Multi-objective optimization:

[0269] F opt =min J(F)st After the constraint conditions are optimized, the cable force deviation is ≤3% and the displacement deviation is ≤2mm.

[0270] 4.2.2 Execution of cable force adjustment

[0271] Adjustment calculation:

[0272]

[0273] Jacobian matrix Obtained through finite element sensitivity analysis.

[0274] Real-time feedback correction:

[0275] Ffinal =F opt +GBT(Δd measured )

[0276] The robustness is improved by correcting the measured displacement deviation Δdmeasured.

[0277] 4.3 Closed-loop control indicators

[0278]

[0279] Experimental verification

[0280] Verification Target

[0281] Steel anchor beam installation accuracy: transverse / along the bridge deviation ≤3.5mm, elevation deviation ≤1.8mm;

[0282] Cable force optimization accuracy: cable force deviation ≤ 2.5%, displacement control ≤ 2mm;

[0283] System response time: key decisions ≤ 30s, overall process efficiency improved ≥ 40%.

[0284] Verification Object

[0285] A 500m-long cable-stayed bridge (main span 260m) has 28 sections of steel anchor beams, each weighing 8.9t, and 48 stay cables (maximum cable force 5800kN).

[0286] Comparison plan:

[0287] Traditional solution: manual measurement + experience adjustment + static cable force optimization;

[0288] This solution: intelligent measurement + Transformer prediction + multi-objective cable tension collaborative optimization.

[0289] 1.3 Verification indicator system, as shown in Table 1

[0290] Table 1

[0291]

[0292]

[0293] Transformer error prediction experiment

[0294] Input data: cumulative error, temperature, and wind speed data of the first 20 sections of steel anchor beams (200 sets of samples);

[0295] Forecasting task: Compare the cumulative error of 6-hour forecasts in advance with the traditional ARIMA model and the Transformer model.

[0296] 2.1.2 Experimental results, as shown in Table 2

[0297] Table 2

[0298]

[0299] When the 15th section was installed, the traditional model predicted a cumulative error of +3.2mm after 6 sections, but it actually reached +5.8mm. In this solution, the Transformer predicted +5.5mm, compensated 0.9mm / section in advance, and the final cumulative error was +1.2mm.

[0300] Cable-force collaborative optimization experiment

[0301] Experimental design

[0302] Working conditions: Cable tension adjustment during the bridge completion stage, target displacement deviation ≤ 2mm, cable tension deviation ≤ 3%;

[0303] Comparison method: Traditional trial algorithm: manual adjustment + finite element verification; This solution: multi-objective optimization + gradient boosting tree correction.

[0304] 2.2.2 Experimental results are shown in Table 3

[0305] Table 3

[0306]

[0307]

[0308] Stress cloud map comparison

[0309] Traditional solution: The local stress concentration area of ​​the steel anchor beam reaches 210MPa (allowable stress 235MPa);

[0310] This solution: The stress distribution is more uniform, the maximum stress is 195MPa, which is reduced by 7.0%.

[0311] Closed-loop control real-time experiment

[0312] Experimental design

[0313] Disturbance input: simulate sudden strong winds (wind speed increases from 8m / s to 12m / s) and monitor system response;

[0314] Indicators: the time from the occurrence of disturbance to the issuance of adjustment instructions, and the deviation convergence speed within 10 minutes after adjustment.

[0315] 2.3.2 Experimental Results

[0316]

[0317] 3. Engineering Application Verification

[0318] Installation accuracy verification

[0319] Statistics of deviation of 28 steel anchor beams in the entire bridge

[0320] Cross bridge direction:

[0321] The maximum deviation of the traditional solution is 6.5mm, and the average is 4.8mm;

[0322] The maximum deviation of this solution is 3.8mm, the mean is 3.2mm, and the compliance rate (≤3.5mm) is 78.6%;

[0323] Elevation: The maximum deviation of the traditional solution is 3.1mm, and the average is 2.3mm; the maximum deviation of this solution is 1.8mm, and the average is 1.5mm, with a compliance rate of 100%.

[0324] Comparison of typical segment deviations

[0325]

[0326] Verification of cable force optimization effect

[0327] Cable force deviation distribution of the entire bridge

[0328] Traditional solution: The number of cables with deviations > 3% accounts for 23%, and the maximum deviation is 5.7%;

[0329] This solution: The number of cables with deviations > 2.5% accounts for 5%, the maximum deviation is 2.1%, and the cable force uniformity is improved by 35%.

[0330] Tower displacement monitoring

[0331] Construction phase: The maximum displacement of the traditional solution was 8.3mm, while that of this solution was 3.7mm, a reduction of 55.4%;

[0332] Bridge construction stage: When the temperature changes by 20°C, the displacement fluctuation of the traditional solution is 4.1mm, while that of this solution is 1.8mm, a reduction of 56.1%.

[0333] Comparative experiments and mechanism analysis

[0334] Error prediction mechanism verification

[0335] Transformer model: The attention weight of the temperature feature reaches 0.28, which is significantly higher than the linear correlation of the ARIMA model; the weight of the recent error (n-1 section) is 0.35, and the weight of the long-term error (n-5 section) is 0.12, which conforms to the error attenuation law.

[0336] Error transmission path

[0337] Traditional solution: processing error → transportation deformation → installation accumulation, no active compensation;

[0338] This solution: After 6 compensation cycles, the error attenuation rate reaches more than 80%.

[0339] Verification of cable force optimization mechanism

[0340] Multi-objective weighted sensitivity analysis

[0341] Weight adjustment experiment: When the stress control increases from 0.2 to 0.4, the maximum stress decreases by 12%, but the cable force deviation rate increases from 1.8% to 2.3%;

[0342] The optimal weight combination (w1=0.5, w2=0.3, w3=0.2) can achieve the best overall performance.

[0343] Cable force-displacement coupling relationship

[0344] Finite element simulation: A 1% adjustment in cable force corresponds to a tower displacement of 0.15-0.3mm. The prediction error of this model is ≤5%. The traditional trial algorithm ignores nonlinear coupling, and the prediction error reaches 15-25%.

[0345] Example 2

[0346] This embodiment provides a cable force optimization system for steel anchor beams of a long-span cable-stayed bridge, comprising:

[0347] The data acquisition module is configured as

[0348] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a method for optimizing the cable force of a steel anchor beam of a long-span cable-stayed bridge.

[0349] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded and executed by the processor, a method for optimizing the cable force of steel anchor beams of a long-span cable-stayed bridge.

[0350] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for optimizing cable force of steel anchor beams of a long-span cable-stayed bridge, characterized in that: include: Obtain original data of steel anchor beams of long-span cable-stayed bridges; Perform data preprocessing on the original data of the steel anchor beam of the long-span cable-stayed bridge; Use Transformer to capture long-range error dependencies and predict the cumulative error of future segments; Use convolutional LSTM networks to generate optimal adjustments based on accumulated errors; The optimal cable tension scheme is solved using the cable tension optimization algorithm.

2. The method for optimizing cable force of steel anchor beams of a long-span cable-stayed bridge according to claim 1, characterized in that: The method of obtaining the original data of the steel anchor beam of a large-span cable-stayed bridge includes obtaining the original accuracy data of the steel anchor beam during the processing stage, judging the processing accuracy by measuring the geometric dimensions of the loose parts, and quantifying the cumulative error effect by cumulatively calculating the assembly height differences of adjacent segments; obtaining pre-assembly data on the construction site, including the linear deviation of the steel corbel embedded plate, the contact surface gap and the height difference of the four corner points; obtaining the real-time posture data of the steel anchor beam during installation to provide real-time feedback for precise positioning and cable force optimization; and also including cable force monitoring data, specifically including basic parameters of the inclined cable and real-time monitoring data.

3. The method for optimizing cable force of steel anchor beams of a long-span cable-stayed bridge according to claim 2, characterized in that: The data preprocessing of the obtained raw data of the steel anchor beam of the long-span cable-stayed bridge includes solving the rotation matrix and translation vector by the least square method, establishing a coordinate transformation relationship, eliminating the difference between the tower column coordinate system and the local coordinate system of the steel anchor beam, and ensuring that the design value and the measured value are compared under the same benchmark; The outliers in the cable force monitoring data were eliminated, and the temperature correction coefficient was used to eliminate the influence of temperature change on the cable force measurement.

4. The method for optimizing cable force of steel anchor beams of a long-span cable-stayed bridge according to claim 3, characterized in that: The method uses Transformer to capture long-distance error dependencies and predict the cumulative errors of future segments, including using the Transformer's self-attention mechanism to capture long-distance error dependencies based on preprocessed data, thereby improving the prediction accuracy of multi-segment cumulative errors. The input layer contains the cumulative errors, temperature, and wind speed features of the first five segments. The multi-head attention mechanism is used to calculate the correlation weights between the features to indicate the degree of influence of the processing error of a certain segment on the subsequent segments. The position encoding gives the sequence data time series information to prevent the model from confusing the error features of different segments.

5. The method for optimizing cable force of steel anchor beams of a long-span cable-stayed bridge according to claim 4, characterized in that: The method uses Transformer to capture long-distance error dependencies and predict the cumulative errors of future segments, and also includes predicting the cumulative errors of future segments based on historical errors and environmental parameters. The encoder-decoder architecture is used to first map the input features to a high-dimensional latent space, and then the predicted values ​​are output through a feedforward network. Residual connections are used to avoid the vanishing gradient of deep networks, and layer normalization is used to stabilize the training process.

6. The method for optimizing cable force of steel anchor beams of a long-span cable-stayed bridge according to claim 5, characterized in that: The method uses a convolutional LSTM network to generate an optimal adjustment amount based on the accumulated error, including using the convolutional LSTM network to fuse spatial feature extraction and time series analysis to achieve accurate calculation of the steel anchor beam posture adjustment amount, wherein the spatial features of the steel anchor beam posture are extracted by the convolution layer, the time dependency of historical adjustments is captured by the LSTM layer, and key information is selectively retained by the gating mechanism. By learning the adjustment priority experience of the transverse bridge direction and the longitudinal bridge direction elevation, the adjustment amount in this direction is automatically calculated first when the transverse bridge direction deviation is large.

7. The method for optimizing cable force of steel anchor beams of a long-span cable-stayed bridge according to claim 6, characterized in that: The method uses a convolutional LSTM network to generate an optimal adjustment amount based on the accumulated error, and also includes generating an optimal adjustment amount based on the design deviation, prediction error and historical record, wherein a deep neural network (DNN) is used to map the multi-dimensional input design deviation, prediction error and historical adjustment to a three-dimensional adjustment amount, and the weight is optimized through back propagation to minimize the final deviation of the output adjustment amount, and the nonlinear activation function of the hidden layer is used to learn complex adjustment rules, including nonlinear compensation for the accumulated error.

8. The method for optimizing cable force of steel anchor beams of a long-span cable-stayed bridge according to claim 7, characterized in that: The method of solving the optimal cable tension scheme by using a cable tension optimization algorithm includes establishing a quantitative relationship between cable tension and structural displacement to provide physical constraints for cable tension optimization. Specifically, a simplified finite element model based on the principles of structural mechanics is constructed, and the mapping of cable tension to displacement is described by a stiffness matrix. A neural network is used to approximate the nonlinear part of the finite element model, and a feasible optimization scheme is generated by seeking a balance between cable tension accuracy, displacement control and structural safety. Finally, an optimal cable tension scheme that meets multiple objective constraints is found through iterative calculation.

9. The method for optimizing cable force of steel anchor beams of a long-span cable-stayed bridge according to claim 8, characterized in that: The method seeks a balance between cable tension accuracy, displacement control and structural safety to generate a feasible optimization scheme, including constructing an objective function including cable tension deviation, displacement deviation and maximum stress, and reflecting the engineering priority through weight coefficients; finding the optimal cable tension scheme that meets multiple objective constraints through iterative calculation, wherein the Adam algorithm is combined with momentum and adaptive learning rate to efficiently search the optimal solution space, continuously adjusts the cable tension vector during the iteration process to minimize the objective function, and sets a convergence threshold to ensure the accuracy of the solution and avoid computational redundancy.

10. A cable force optimization system for steel anchor beams of a long-span cable-stayed bridge, characterized in that: include: The data acquisition module is configured to acquire original data of the steel anchor beam of the long-span cable-stayed bridge; The preprocessing module is configured to perform data preprocessing on the acquired raw data of the steel anchor beam of the long-span cable-stayed bridge; The error module is configured to use Transformer to capture long-range error dependencies and predict the cumulative errors of future segments; An adjustment module is configured to generate an optimal adjustment amount based on the accumulated error using a convolutional LSTM network; The optimization module is configured to solve the optimal cable tension scheme using a cable tension optimization algorithm.

Images