A girder erecting elevation adjustment method based on machine learning
By combining the LSTM-Transformer hybrid model and the XG-Boost algorithm, a reinforcement learning agent is constructed, which solves the problem of precise control of the elevation of the formwork and achieves efficient, accurate and stable adjustment of elevation prediction, which is suitable for complex construction environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA HIGHWAY ENG CONSULTING GRP CO LTD
- Filing Date
- 2025-12-22
- Publication Date
- 2026-06-26
AI Technical Summary
Traditional methods for cantilever construction of main beams in cable-stayed/suspension bridges suffer from insufficient data utilization, lack of nonlinear modeling, and poor dynamic adaptability in accurately controlling the formwork elevation. This leads to deviations from actual values, requiring repeated manual adjustments and resulting in low efficiency.
A hybrid LSTM-Transformer model is used to process time series data, and the XG-Boost algorithm is combined to analyze nonlinear variables. A reinforcement learning agent is constructed, and the elevation of the erected mold is adjusted through a hydraulic system. The elevation adjustment amount is optimized by using a spatiotemporal database and the reinforcement learning agent to achieve real-time dynamic adjustment.
It significantly improves the accuracy and efficiency of elevation prediction and adjustment, reduces the need for manual intervention, enhances the adaptability and stability of elevation adjustment, and improves prediction accuracy by 30%.
Smart Images
Figure CN121502229B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of elevation adjustment technology, specifically relating to a machine learning-based method for elevation adjustment of main beam formwork. Background Technology
[0002] In the cantilever construction of the main girder of cable-stayed / suspension bridges, the precise control of the formwork elevation directly affects the final bridge alignment. Traditional methods rely on empirical formulas or linear regression models, which have significant drawbacks: insufficient data utilization: only considering single-point data such as temperature and load, ignoring the temporal correlation of historical construction data; lack of nonlinear modeling: the complex nonlinear relationship between factors such as concrete creep, cable relaxation, and elevation is difficult to express analytically; poor dynamic adaptability: changes in the construction environment (such as sudden winds, fluctuations in material properties) cause predicted values to deviate from reality, requiring repeated manual adjustments, which is inefficient.
[0003] Therefore, a machine learning-based method for adjusting the elevation of the main beam formwork is needed to solve the above problems. Summary of the Invention
[0004] In view of this, the purpose of this invention is to provide a machine learning-based method for adjusting the elevation of the main beam formwork, which aims to improve the prediction accuracy and adjustment efficiency of the elevation of the formwork in complex construction environments.
[0005] To achieve the above objectives, the present invention provides the following technical solution:
[0006] This invention provides a machine learning-based method for adjusting the elevation of main beam formwork, comprising the following steps:
[0007] S1: Collect data during the construction period of the cable-free section, including ambient temperature, concrete elastic modulus, load history, formwork deformation time series data and measured elevation deviation, and construct a spatiotemporal database;
[0008] S2: The LSTM-Transformer hybrid model is used to process time series data, extract long-period features, and output the initial predicted elevation values. The LSTM layer captures local temporal dependencies, while the Transformer self-attention mechanism models global correlations.
[0009] S3: Analyze the nonlinear variable set using the XG-Boost algorithm. Inputs include cable stress change rate, solar radiation gradient, and material age. Output is the nonlinear compensation factor. ;
[0010] S4: Fusion and Generate joint predictions ;
[0011] S5: Construct a reinforcement learning agent, using the construction phase as the state space and the elevation adjustment amount as the action space, with the goal of minimizing... Using construction stability as the reward function, output adjustment instructions. According to the adjustment instructions The hydraulic system is driven to adjust the elevation of the vertical mold and update the database in real time.
[0012] Furthermore, the LSTM-Transformer hybrid model adopts a hierarchical extraction architecture for spatiotemporal features, specifically including:
[0013] The input encoding layer divides the temperature load and deformation time series data into 144 continuous time windows according to construction stages. Each window undergoes Z-score normalization and a sinusoidal position encoding vector is added, generating a dimension of... The input tensor;
[0014] The bidirectional gated recurrent layer adopts a two-layer stacked Bi-LSTM structure with 256 hidden layer neurons. The forward LSTM captures the history dependence of early concrete shrinkage, and the backward LSTM models the prestressing tension hysteresis effect, outputting a 512-dimensional bidirectional feature vector.
[0015] The attention enhancement module adds a learnable position embedding matrix in front of the Transformer encoder, enabling the self-attention mechanism to distinguish the deformation mode differences between the vibration stage and the static curing stage of the hydraulic pump.
[0016] The multi-head spatiotemporal perceptron is equipped with 8 independent attention heads, of which 4 attention heads are used for temperature-strain cross-correlation analysis, and the other 4 attention heads focus on cable force-displacement transmission relationship. The output layers of the attention heads are normalized and then stitched together to form a global feature map.
[0017] The anti-interference output layer is an embedded Dropout layer with a random inactivation rate of 0.3 in the fully connected network of the prediction head, and an adaptive gradient pruning strategy is adopted during backpropagation.
[0018] Furthermore, in step S3, the nonlinear compensation factor The generation includes the following steps:
[0019] A1: Calculate the concrete creep coefficient based on material age, generate the east-west template deformation differential characteristics by combining the solar radiation gradient, and construct the energy accumulation characteristics using the second derivative of the cable force change rate.
[0020] A2: In XG-Boost, a splitting criterion based on construction phase awareness is introduced, using entropy-weighted gain for samples with an age of less than 3 days and mean square error gain for samples with an age of more than 7 days.
[0021] A3: Mechanical Constraint Compensation: Apply material mechanical constraints to the output values of the decision tree leaf nodes to ensure... The absolute value does not exceed the threshold determined by the elastic modulus and the coefficient of thermal expansion;
[0022] A4: When the measured elevation deviation exceeds 5mm twice consecutively, the feature importance reassessment mechanism is triggered, and the feature weight ratio of cable force change rate and solar gradient is dynamically adjusted.
[0023] Furthermore, the reinforcement learning agent in step S5 adopts a hierarchical decision-making architecture, including:
[0024] The upper-level state sensor discretizes the construction stage into three state subspaces: foundation pouring, prestressing tensioning, and system transformation. Each subspace defines an independent set of state variables.
[0025] The motion space constraint module dynamically limits the range of single adjustment based on the current concrete slump. When the slump is greater than 180mm, the single adjustment is limited to ≤3mm. When the curing time is less than 12 hours, cross-stage formwork adjustment is prohibited.
[0026] The multi-objective reward function has a basic reward term of 1 / (1+|Pjoint-measured value|), and the stability reward term introduces the variance penalty coefficient of the mold adjustment action. An additional cable force uniformity reward is added during the system conversion stage.
[0027] The adversarial training mechanism injects historical maximum temperature difference ±15℃ disturbance data into the simulation environment and uses a near-end strategy optimization algorithm to train the agent to resist the impact of sudden weather.
[0028] Furthermore, the hydraulic drive system in step S6 includes the following cooperative control mechanism:
[0029] exist Before sending the data to the hydraulic actuator, a transient mechanical simulation of the template system to be adjusted is performed using a finite element model. If the maximum stress is greater than 30% of the concrete compressive strength, an alarm is triggered.
[0030] The displacement difference of the four corners is monitored in real time by a high-precision laser rangefinder. When the lag at a single point is greater than 0.5mm, the flow distribution of the hydraulic valve is adjusted to achieve dynamic correction.
[0031] After the elevation adjustment is completed, the actual displacement and... The error data is used to train an error compensation neural network and pre-compensate the system's mechanical hysteresis during the next action;
[0032] When the ambient wind speed is greater than level 8 or the deformation difference between adjacent measuring points is greater than 2mm, the system automatically switches to pressure holding mode and locks the hydraulic valve.
[0033] Furthermore, the construction of the spatiotemporal database in step S1 includes:
[0034] Vibrating wire strain sensors, fiber optic temperature sensors, and hydraulic force gauges are deployed at key sections of the cable-free template to collect data synchronously at a sampling rate of 10Hz.
[0035] An improved dynamic time warping algorithm was used to align temperature and humidity data with deformation data at different sampling rates, and the spatial coordinates of the measurement points were calibrated using the BeiDou positioning system.
[0036] Sensor failure data is detected based on the isolated forest algorithm. When five consecutive samples deviate from the mean by three times the standard deviation, Kalman filtering is activated to repair the failure.
[0037] Automatically identify pouring start and tensioning completion events in the construction log and associate them with the corresponding timestamp of the mechanical state snapshot in the database.
[0038] Furthermore, the fusion process in step S4 employs an adaptive weighting strategy:
[0039] Confidence assessment: calculation In recent K Average absolute error of each construction stage Simultaneously calculate the nonlinear compensation factor. Standard deviation of similar materials at different ages ;in,
[0040]
[0041]
[0042] In the formula, For the number of stages, For the first i Predicted values for the stage, For the first i The actual value of the stage; For the sample size, For the first i one sample value, for N The average of the samples;
[0043] Dynamic weight allocation: joint predicted value ,in The weighting coefficients are generated using the logistic function: In the formula, k The correlation coefficient is the material type.
[0044] Cross-phase transfer: When a system transition event is detected, the data from the end of the previous construction phase will be transferred. The value is used as a priori to initialize the weights for the current stage;
[0045] Residual error feedback: Define the weight adjustment amount In the formula, To adjust the step size, take , For residuals, , indicating the first j Each stage, the fused predicted value Compared with the true value The difference between them; if the residual If the same sign appears consecutively, then follow the rules. Update the weights.
[0046] Furthermore, in step A3, the implementation of mechanical constraint compensation includes:
[0047] B1: Based on laboratory data, fit the temperature-expansion rate curve to generate the maximum allowable expansion amount under different mixing ratios. ;
[0048] B2: Obtain the hydraulic system pressure-stroke conversion matrix through jack calibration tests, and constrain... The component caused by cable force in the middle;
[0049] B3: XG-Boost Raw Output First, input the material property modifier. If... The output is then compressed using the Sigmoid function and then input into the cable tension channel limiter.
[0050] B4: Automatically outputs statistics on the number of constraint triggers, types, and adjustment amounts for construction quality traceability.
[0051] Furthermore, it also includes an optimization mechanism for adjusting instructions verification:
[0052] Digital twin sand table: Construct a parametric template system model in the BIM platform and import actual sensor data to drive the virtual template movement in real time;
[0053] Pre-execution simulation module: Input the sand table to perform multiphysics coupling simulation and predict the creep development path within 6 hours after adjustment;
[0054] Decision optimization engine: If simulation shows that the strain rate of a critical section exceeds a threshold, a progressive leveling sequence is generated. Replace the original single action; satisfy and mm; They are respectively to The height change of the first and second leveling commands after splitting;
[0055] Knowledge accumulation system: Achieving better-than-expected results from actual adjustments. The values and their operating characteristics are stored in the case library to optimize the reinforcement learning strategy network.
[0056] The beneficial effects of this invention are as follows:
[0057] 1. The method proposed in this invention, by combining the LSTM-Transformer hybrid model and the XG-Boost algorithm, can effectively capture the influence of long-period time-series characteristics and nonlinear variables, thereby significantly improving the accuracy of elevation prediction. Especially in complex construction environments, this method can analyze multi-dimensional factors such as ambient temperature and humidity, cable tension change rate, and solar radiation gradient in real time, outputting accurate compensation factors. By optimizing the reinforcement learning agent, the system can dynamically adjust the elevation adjustment amount. This ensures the real-time accuracy and stability of the model elevation. Compared to traditional methods, this method improves prediction accuracy by approximately 30% while significantly reducing the need for manual intervention.
[0058] 2. In the elevation adjustment process, this invention not only considers traditional environmental temperature and humidity and historical load data, but also introduces nonlinear factors such as cable force change rate and solar radiation gradient. By modeling these complex factors using the XG-Boost algorithm, it can effectively capture nonlinear relationships that are difficult to identify using traditional methods. Furthermore, the LSTM-Transformer hybrid model, when processing time-series data, can simultaneously capture local and global temporal dependencies, further enhancing the model's generalization ability. This multi-dimensional comprehensive modeling method significantly improves the comprehensiveness and adaptability of elevation adjustment, making it suitable for various complex construction environments.
[0059] 3. By constructing a reinforcement learning agent, this invention transforms the elevation adjustment process into an optimization problem, aiming to minimize the deviation between predicted and measured values while considering construction stability. This intelligent optimization method can dynamically adjust the elevation adjustment amount. This method ensures that the formwork elevation remains optimal across different construction stages and environments. Compared to traditional rule-based control methods, this method significantly improves the stability and response speed of elevation adjustments, reducing repeated adjustments and resource waste during construction.
[0060] Other advantages, objectives, and features of the invention will be set forth in the following description and will be apparent to those skilled in the art in some respects, or may be learned by practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0061] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration:
[0062] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation
[0063] like Figure 1 As shown, this invention provides a machine learning-based method for adjusting the elevation of a main beam formwork, comprising the following steps:
[0064] S1: Collect data during the construction period of the cable-free section, including ambient temperature, concrete elastic modulus, load history, formwork deformation time series data and measured elevation deviation, and construct a spatiotemporal database;
[0065] S2: The LSTM-Transformer hybrid model is used to process time series data, extract long-period features, and output the initial predicted elevation values. The LSTM layer captures local temporal dependencies, while the Transformer self-attention mechanism models global correlations.
[0066] S3: Analyze the nonlinear variable set using the XG-Boost algorithm. Inputs include cable stress change rate, solar radiation gradient, and material age. Output is the nonlinear compensation factor. ;
[0067] S4: Fusion and Generate joint predictions ;
[0068] S5: Construct a reinforcement learning agent, using the construction phase as the state space and the elevation adjustment amount as the action space, with the goal of minimizing... Using construction stability as the reward function, output adjustment instructions. According to the adjustment instructions The hydraulic system is driven to adjust the elevation of the vertical mold and update the database in real time.
[0069] In one embodiment of the present invention, the LSTM-Transformer hybrid model adopts a spatiotemporal feature hierarchical extraction architecture, specifically including:
[0070] The input encoding layer divides the temperature load and deformation time series data into 144 continuous time windows according to construction stages. Each window undergoes Z-score normalization and a sinusoidal position encoding vector is added, generating a dimension of... The input tensor; where, This refers to the batch size, which is the number of data samples processed at one time. For time step; The number of features is 5 at each time step;
[0071] The bidirectional gated recurrent unit adopts a two-layer stacked Bi-LSTM structure with 256 hidden layer neurons. The forward LSTM learns the historical dependence of early concrete shrinkage, and the backward LSTM models the prestressing tension hysteresis effect, outputting a 512-dimensional bidirectional feature vector.
[0072] The attention enhancement module adds a learnable position embedding matrix in front of the Transformer encoder, enabling the self-attention mechanism to distinguish the deformation mode differences between the vibration stage and the static curing stage of the hydraulic pump.
[0073] The multi-head spatiotemporal sensor is equipped with eight independent attention heads, four of which calculate the temperature-strain cross matrix. The other four attention heads calculate the cable force-displacement transfer matrix. ,Will and The output layers are normalized and concatenated, then compressed into a 256-dimensional global feature vector through a fully connected layer.
[0074]
[0075]
[0076] In the formula, This is a temperature query vector. For strain bond vectors, Let be the dimension of the key vector. For the cable query vector, For the displacement bond vector; This is the normalization function;
[0077] The anti-interference output layer embeds a random deactivation module with a random deactivation rate of 0.3 in the fully connected layer and employs an adaptive gradient pruning threshold during backpropagation. Limit weight oscillations caused by sudden construction noise.
[0078]
[0079] In the formula, Let L2 be the gradient of the loss function, and let represent the magnitude of the gradient. To train step count, It is an exponential function, increasing with the number of training steps. As the threshold increases, the denominator gradually approaches 1, causing the threshold to gradually increase.
[0080] In one embodiment of the present invention, the nonlinear compensation factor in step S3 The generation includes the following steps:
[0081] A1: Physical Characteristics and Structure: Calculation of Concrete Creep Coefficient Based on Material Age ,in, The material age is expressed in days; the template deformation differential characteristics are generated by combining the east-west solar radiation gradient. ,in, In response, Using the position coordinates, the second derivative of the rate of change of cable force is used. Constructing energy accumulation characteristics: ,in, For time;
[0082] A2: Dynamically Weighted Tree Sets: Defining Construction Phase-Aware Splitting Criteria in XGBoost
[0083] If age For 3 days, entropy-weighted gain is used: ,in, This is the temperature correlation coefficient. and These are the entropies of the left and right subtrees, respectively.
[0084] If age For 7 days, the mean square error gain is used: ,in, For the sample size, For the true value, This is a predicted value;
[0085] A3: Mechanical Constraint Compensation: Setting Compensation Factors Boundary conditions, ,in, As the maximum compensation factor, In the formula, The coefficient of thermal expansion is For temperature difference, The length of the beam segment. This is the proportionality coefficient. For cable force, The elastic modulus; when the original output of XG-Boost When the limit is exceeded, according to Compress;
[0086] A4: Online incremental learning: When two consecutive measured elevation deviations... At 5mm, feature weight update is triggered: ,in, For the first Each feature weight, For learning rate, For loss function Weights The partial derivatives; For the first Each feature weight in The value at the next iteration For the first Each feature weight in The value at the next iteration.
[0087] In one embodiment of the present invention, the reinforcement learning agent in step S5 adopts a hierarchical decision-making architecture, including:
[0088] The upper-level state perceptron is used for state space partitioning, discretizing the construction phase into three subspaces: ,in, These are respectively foundation pouring, prestressing tensioning, and system transformation; each subspace defines an independent set of state variables. ,in, For temperature, For the rate of change of cable force, In response, The material age is expressed in days.
[0089] The motion space constraint module dynamically limits the range of single adjustment based on the current concrete slump. When the slump is greater than 180mm, the single adjustment is limited to ≤3mm. When the curing time is less than 12 hours, cross-stage formwork adjustment is prohibited.
[0090] Multi-objective reward function, total reward Defined as: ,in, , During the system transition phase, a uniform reward for cable force is added. ;
[0091] Adversarial training mechanism, injecting temperature perturbations into the simulation environment. The Proximal Policy Optimization (PPO) algorithm is used to update the policy network parameters. The objective function is:
[0092]
[0093] In the formula, The reward is for the accuracy of the agent's predictions compared to the actual values. The predicted value of the agent. For the agent's true value, For balance coefficient, To measure the stability of the agent's modulus changes over the past four steps. To start from time step arrive The modulus change sequence, Let the standard deviation of the cable force be , Let PPO be the pruning loss function. For strategy ratio, For the sake of advantage estimation, The cropping area; Indicates the strategy ratio Limited to the range Inside;
[0094] In one embodiment of the present invention, the hydraulic drive system in step S6 includes the following cooperative control mechanism:
[0095] Mechanical safety pre-inspection will Input a parametric finite element model for transient analysis, if the template stress This will trigger an alarm, among which, It refers to the compressive strength of concrete;
[0096] Multi-cylinder synchronous control, with real-time monitoring of four-corner displacement via laser rangefinder. If the displacement difference between any two cylinders exceeds 0.5mm, then it shall be handled according to... Adjust the hydraulic flow rate, among which, This is the hydraulic flow adjustment amount for cylinder k. This is the proportionality coefficient. This represents the average value of the four corner displacements. For the first k The displacement of the cylinder;
[0097] Hysteresis compensation, establishing an error-compensation mapping network By inputting historical error sequences Output the compensation amount for the next instruction. ;
[0098] The environmental interlock strategy automatically switches to pressure-holding mode and locks the hydraulic valve when the ambient wind speed is greater than level 8 or the deformation difference between adjacent measuring points is greater than 2mm.
[0099] In one embodiment of the present invention, the construction of the spatiotemporal database in step S1 includes:
[0100] Vibrating wire strain sensors, fiber optic temperature sensors, and hydraulic force gauges are deployed at key sections of the cable-free template to collect data synchronously at a sampling rate of 10Hz.
[0101] An improved dynamic time warping algorithm was used to align temperature and humidity data with deformation data at different sampling rates, and the spatial coordinates of the measurement points were calibrated using the BeiDou positioning system.
[0102] Sensor failure data is detected based on the isolated forest algorithm. When five consecutive samples deviate from the mean by three times the standard deviation, Kalman filtering is activated to repair the failure.
[0103] Automatically identify pouring start and tensioning completion events in the construction log and associate them with the corresponding timestamp of the mechanical state snapshot in the database.
[0104] In one embodiment of the present invention, the fusion process in step S4 employs an adaptive weighting strategy:
[0105] Confidence assessment: calculation In recent K Average absolute error of each construction stage Simultaneously calculate the nonlinear compensation factor. Standard deviation of similar materials at different ages ;in,
[0106]
[0107]
[0108] In the formula, For the number of stages, For the first i Predicted values for the stage, For the first i The actual value of the stage; For the sample size, For the first i one sample value, for N The average of the samples;
[0109] Dynamic weight allocation: joint predicted value ,in The weighting coefficients are generated using the logistic function: In the formula, k The correlation coefficient is the material type.
[0110] Cross-phase transfer: When a system transition event is detected, the data from the end of the previous construction phase will be transferred. The value is used as a priori to initialize the weights for the current stage;
[0111] Residual error feedback: Define the weight adjustment amount In the formula, To adjust the step size, take , For residuals, , indicating the first j Each stage, the fused predicted value Compared with the true value The difference between them; if the residual If the same sign appears consecutively, then follow the rules. Update the weights; where, It is a symbolic function.
[0112] This adaptive weighting strategy evaluates... and By dynamically adjusting their weights in the fusion process based on their confidence level, and combining cross-stage propagation and residual feedback mechanisms, the accuracy of the predicted values and the stability of the system are ensured.
[0113] In one embodiment of the present invention, the implementation of mechanical constraint compensation in step A3 includes:
[0114] B1: Fitting temperature-expansion curves based on laboratory data In the formula, The coefficient of thermal expansion is The amount of temperature increase The length of the beam end. These are nonlinear correction coefficients;
[0115] B2: Establishing a pressure-stroke conversion matrix based on jack calibration data Constraint cable force components satisfy:
[0116] mm;
[0117] B3: Dual-channel correction:
[0118] B3.1: Material Corrector: If Then compress according to the following formula: ;
[0119] B3.2: Cable Limiter: If If the value is in mm, then scale according to the following formula: ,in, This is the final corrected compensation factor;
[0120] B4: Interpretable output, generating a compensation factor decomposition report: The contribution rate of each component is labeled, where, This is the compensation amount due to temperature. The compensation amount caused by cable tension. The amount of compensation caused by other factors.
[0121] In one embodiment of the present invention, a verification and optimization mechanism for adjustment instructions is included:
[0122] Digital twin sand table: Construct a parametric template system in the BIM platform to map sensor data to the virtual model in real time;
[0123] Pre-execution simulation: Input a sand table for creep prediction, and the strain rate threshold determination condition is as follows:
[0124]
[0125] In the formula, , For creep parameters, For age period, For stress;
[0126] If creep strain rate When this happens, an alert is triggered;
[0127] Gradual leveling: When a warning is triggered, the original command will be adjusted. Split into ,satisfy and mm; They are respectively to The height change of the first and second leveling commands after splitting;
[0128] Knowledge Accumulation: Building a Case Library and Optimizing Instructions and eigenvectors A reinforcement learning sample augmentation dataset is generated through cluster analysis; where, For temperature, For traffic, For age period, This is the target value.
[0129] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.
Claims
1. A method for adjusting the elevation of a main beam formwork based on machine learning, characterized in that: Includes the following steps: S1: Collect data during the construction period of the cable-free section, including ambient temperature, concrete elastic modulus, load history, formwork deformation time series data and measured elevation deviation, and construct a spatiotemporal database; S2: The LSTM-Transformer hybrid model is used to process time series data, extract long-period features, and output the initial predicted elevation values. The LSTM layer captures local temporal dependencies, while the Transformer self-attention mechanism models global correlations. S3: Analyze the nonlinear variable set using the XG-Boost algorithm. Inputs include cable stress change rate, solar radiation gradient, and material age. Output is the nonlinear compensation factor. ; S4: Fusion and Generate joint predictions ; S5: Construct a reinforcement learning agent, using the construction phase as the state space and the elevation adjustment amount as the action space, with the goal of minimizing... Using construction stability as the reward function, output adjustment instructions. According to the adjustment instructions The hydraulic system is driven to adjust the elevation of the vertical mold and update the database in real time; Among them, the reward function Defined as: ,in, , During the system transition phase, a uniform reward for cable force is added. ; In the formula, The reward is for the accuracy between the agent's predicted value and the actual value. The predicted value of the agent. For the agent's true value, For balance coefficient, To measure the stability of the agent's modulus changes over the past four steps. To start from time step arrive The modulus change sequence, Let be the standard deviation of the cable force.
2. The machine learning-based method for adjusting the elevation of the main beam formwork as described in claim 1, characterized in that: The LSTM-Transformer hybrid model adopts a hierarchical extraction architecture for spatiotemporal features, specifically including: The input encoding layer divides the temperature load and deformation time series data into 144 continuous time windows according to construction stages. Each window undergoes Z-score normalization and a sinusoidal position encoding vector is added, generating a dimension of... The input tensor; The bidirectional gated recurrent layer adopts a two-layer stacked Bi-LSTM structure with 256 hidden layer neurons. The forward LSTM captures the history dependence of early concrete shrinkage, and the backward LSTM models the prestressing tension hysteresis effect, outputting a 512-dimensional bidirectional feature vector. The attention enhancement module adds a learnable position embedding matrix in front of the Transformer encoder, enabling the self-attention mechanism to distinguish the deformation mode differences between the vibration stage and the static curing stage of the hydraulic pump. The multi-head spatiotemporal perceptron is equipped with 8 independent attention heads, of which 4 attention heads are used for temperature-strain cross-correlation analysis, and the other 4 attention heads focus on cable force-displacement transmission relationship. The output layers of the attention heads are normalized and then stitched together to form a global feature map. The anti-interference output layer is an embedded Dropout layer with a random inactivation rate of 0.3 in the fully connected network of the prediction head, and an adaptive gradient pruning strategy is adopted during backpropagation.
3. The machine learning-based method for adjusting the elevation of the main beam formwork as described in claim 1, characterized in that: In step S3, the nonlinear compensation factor The generation includes the following steps: A1: Calculate the concrete creep coefficient based on material age, generate the east-west template deformation differential characteristics by combining the solar radiation gradient, and construct the energy accumulation characteristics using the second derivative of the cable force change rate. A2: In XG-Boost, a splitting criterion based on construction phase awareness is introduced, using entropy-weighted gain for samples with an age of less than 3 days and mean square error gain for samples with an age of more than 7 days. A3: Mechanical Constraint Compensation: Apply material mechanical constraints to the output values of the decision tree leaf nodes to ensure... The absolute value does not exceed the threshold determined by the elastic modulus and the coefficient of thermal expansion; A4: When the measured elevation deviation exceeds 5mm twice consecutively, the feature importance reassessment mechanism is triggered, and the feature weight ratio of cable force change rate and solar gradient is dynamically adjusted.
4. The machine learning-based method for adjusting the elevation of the main beam formwork as described in claim 1, characterized in that: The reinforcement learning agent in step S5 adopts a hierarchical decision-making architecture, including: The upper-level state sensor discretizes the construction stage into three state subspaces: foundation pouring, prestressing tensioning, and system transformation. Each subspace defines an independent set of state variables. The motion space constraint module dynamically limits the amount of adjustment per cycle based on the current concrete slump. The range is limited when the slump is >180mm. ≤3mm, cross-stage mold adjustment is prohibited when the age is <12 hours; The multi-objective reward function has a basic reward term of 1 / (1+|Pjoint-measured value|), and the stability reward term introduces the variance penalty coefficient of the mold adjustment action. An additional cable force uniformity reward is added during the system conversion stage. The adversarial training mechanism injects historical maximum temperature difference ±15℃ disturbance data into the simulation environment and uses a near-end strategy optimization algorithm to train the agent to resist the impact of sudden weather.
5. The machine learning-based method for adjusting the elevation of the main beam formwork as described in claim 1, characterized in that: The hydraulic drive system in step S6 includes the following cooperative control mechanism: exist Before sending the data to the hydraulic actuator, a transient mechanical simulation of the template system to be adjusted is performed using a finite element model. If the maximum stress is greater than 30% of the concrete compressive strength, an alarm is triggered. The displacement difference of the four corners is monitored in real time by a high-precision laser rangefinder. When the lag at a single point is greater than 0.5mm, the flow distribution of the hydraulic valve is adjusted to achieve dynamic correction. After the elevation adjustment is completed, the actual displacement and... The error data is used to train an error compensation neural network and pre-compensate the system's mechanical hysteresis during the next action; When the ambient wind speed is greater than level 8 or the deformation difference between adjacent measuring points is greater than 2mm, the system automatically switches to pressure holding mode and locks the hydraulic valve.
6. The machine learning-based method for adjusting the elevation of the main beam formwork as described in claim 1, characterized in that: The construction of the spatiotemporal database in step S1 includes: Vibrating wire strain sensors, fiber optic temperature sensors, and hydraulic force gauges are deployed at key sections of the cable-free template to collect data synchronously at a sampling rate of 10Hz. An improved dynamic time warping algorithm was used to align temperature and humidity data with deformation data at different sampling rates, and the spatial coordinates of the measurement points were calibrated using the BeiDou positioning system. Sensor failure data is detected based on the isolated forest algorithm. When five consecutive samples deviate from the mean by three times the standard deviation, Kalman filtering is activated to repair the failure. Automatically identify pouring start and tension completion events in the construction log and associate them with the corresponding timestamp of the mechanical state snapshot in the database.
7. The machine learning-based method for adjusting the elevation of the main beam formwork as described in claim 1, characterized in that: The fusion process in step S4 employs an adaptive weighting strategy: Confidence assessment: calculation In recent K Average absolute error of each construction stage Simultaneously calculate the nonlinear compensation factor. Standard deviation of similar materials at different ages ;in, In the formula, For the number of stages, For the first i Predicted values for the stage, For the first i The actual value of the stage; For the sample size, For the first i one sample value, for N The average of the samples; Dynamic weight allocation: joint predicted value ,in The weighting coefficients are generated using the logistic function: In the formula, k The correlation coefficient is the material type. Cross-phase transfer: When a system transition event is detected, the data from the end of the previous construction phase will be transferred. The value is used as a priori to initialize the weights for the current stage; Residual error feedback: Define the weight adjustment amount In the formula, To adjust the step size, take , For residuals, , indicating the first j Each stage, the fused predicted value Compared with the true value The difference between them; if the residual If the same sign appears consecutively, then follow the rules. Update the weights.
8. The machine learning-based method for adjusting the elevation of the main beam formwork as described in claim 3, characterized in that: In step A3, the implementation of mechanical constraint compensation includes: B1: Based on laboratory data, fit the temperature-expansion rate curve to generate the maximum allowable expansion amount under different mixing ratios. ; B2: Obtain the hydraulic system pressure-stroke conversion matrix through jack calibration tests, and constrain... The component caused by cable force in the middle; B3: XG-Boost Raw Output First, input the material property modifier. If... The output is then compressed using the Sigmoid function and then input into the cable tension channel limiter. B4: Automatically outputs statistics on the number of constraint triggers, types, and adjustment amounts for construction quality traceability.
9. The machine learning-based method for adjusting the elevation of the main beam formwork as described in claim 1, characterized in that: It also includes an adjustment instruction verification optimization mechanism: Digital twin sand table: Construct a parametric template system model in the BIM platform and import actual sensor data to drive the virtual template movement in real time; Pre-execution simulation module: Input the sand table to perform multiphysics coupling simulation and predict the creep development path within 6 hours after adjustment; Decision optimization engine: If simulation shows that the strain rate of a critical section exceeds a threshold, a progressive leveling sequence is generated. Replaces the original single action; Knowledge accumulation system: Achieving better-than-expected results from actual adjustments. The values and their operating characteristics are stored in the case library to optimize the reinforcement learning strategy network.