Steel box superposed continuous box girder bridge prefabrication and assembly error monitoring method

Abstract

The invention relates to a method for monitoring prefabrication and assembly errors of a steel box laminated continuous box girder bridge, and belongs to the technical field of data processing. The method comprises the following steps: collecting prefabrication and assembly error monitoring data of the steel box superposed continuous box girder bridge, and marking the data; an error prediction model is constructed, wherein the error prediction model comprises a physical constraint feature fusion layer, a causal convolution layer, a Bayesian uncertainty propagation layer, a residual attention mechanism module and a dual-channel gating fusion module; training the model by adopting the marked data; in the training process, optimizing the model by constructing mixed loss to obtain a trained model; and aligning timestamps of the collected structure parameters, construction processes, environment parameters and measurement and state parameters of the new prefabricated segment, integrating the parameters into an input sample isomorphic with the training set, and inputting the input sample into the trained model to obtain a prediction result. According to the invention, systematic deviation can be effectively eliminated, and high-robustness and high-precision error prediction is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data processing, and particularly relates to a steel box superimposed continuous box girder bridge prefabrication and assembling error monitoring method. BACKGROUND

[0002] With the wide application of large-scale fabricated bridge structures in engineering practice, the large size of components, the complex nodes and the variable construction environment have posed unprecedented challenges to the assembling precision and construction quality. Especially in the UHPC-steel box composite structure, the material performance is highly sensitive, the structural connection mode has a significant influence on the overall stiffness and stress performance, and slight deviation may cause cumulative errors, affecting the linear control and service safety of the bridge. At present, the existing technology mainly relies on conventional three-dimensional measurement and manual recording means, combined with a simple data-driven prediction model, and there are three key pain points: firstly, the heterogeneity of multi-source data is difficult to model uniformly, the information of structural size, process flow and environmental factors has different dimensions, and direct splicing may lead to model convergence difficulties; secondly, the modeling of time sequence process characteristics ignores the physical interval effect, traditional time sequence models such as RNN cannot identify the actual time difference between processes, and thus misjudge the process correlation; thirdly, the measurement equipment error is not explicitly modeled, and the laser ranging, positioning system and stress and strain sensor have non-Gaussian noise, and the traditional method cannot reflect the uncertainty in the prediction, and finally it is difficult to meet the actual requirements of millimeter-level error control in bridge assembling.

[0003] Therefore, there is an urgent need for a new error monitoring method that integrates multi-source heterogeneous data, physical constraint characteristics and equipment error modeling, to realize high-precision, intelligent and interpretable error prediction and dynamic monitoring in the whole process from component prefabrication to field assembly. SUMMARY

[0004] To achieve the above-mentioned purpose, the application realizes the technical scheme as follows: The application provides a steel box superimposed continuous box girder bridge prefabrication and assembling error monitoring method, comprising the following steps: Step one, collect steel box superimposed continuous box girder bridge prefabrication and assembling error monitoring data and label the data; the steel box superimposed continuous box girder bridge prefabrication and assembling error monitoring data includes structural parameters, construction process, environmental parameters and measurement and state parameters.

[0005] Step two, build an error prediction model, including a physical constraint feature fusion layer, a causal convolution layer, a Bayesian uncertainty propagation layer, a residual attention mechanism module and a double-channel gate fusion module; train the error prediction model using the labeled data; in the training process, the model is optimized by building a hybrid loss of fusion statistical loss and engineering hard constraint, to obtain a trained model; Step three, after aligning the time stamp of the collected structure parameters, construction technology, environmental parameters and measurement and state parameters of the new prefabricated segment, the input sample is integrated into the same structure as the training set, and is input into the trained model to obtain the prediction result.

[0006] Further, the multi-source data such as structure parameters, construction technology and environmental parameters have the heterogeneous characteristics of multi-scale and multi-dimension, and the conventional feature direct splicing method causes difficulty in model convergence due to the difference in feature distribution, and ignores the strong correlation between physical parameters. The application projects the structure parameters and the environmental parameters into a unified feature space through a trainable weight matrix respectively, combines the coupling constraint term of the construction technology and the structure parameters, maps after the activation function, and fuses to model the physical constraint relationship, and obtains the fused physical constraint feature vector.

[0007] Further, the process parameters such as assembly sequence and time have strong time sequence dependence, and the conventional RNN method ignores the physical meaning of the process interval, such as the weakened correlation between the processes with long time interval, which causes false correlation between non-continuous processes, and cannot accurately model the physical forgetting effect in actual construction. In the causal convolution layer, the application adopts a gating causal convolution combined with a physical forgetting effect strategy to extract process features, and the specific operation is: based on the time stamp information in the process vector, the physical time interval between the current time and the historical time is calculated; in the causal convolution, an exponential decay term is combined to make the weight of the historical process feature decay with the increase of the physical time interval, simulate the physical forgetting effect, and obtain the time sequence process feature vector extracted by convolution.

[0008] Further, the measurement devices such as laser range finder and positioning system have non-Gaussian distribution errors, and the conventional method ignores the transmission and accumulation effect of the error in the prediction model, which causes the prediction result not to contain the measurement uncertainty, and cannot meet the construction precision control requirement. The application models the transmission process of the measurement device error in the prediction process by constructing a Bayesian error propagation layer: according to the component geometric parameters and the device technical manual, the error variance of each type of measurement device is calculated; based on the fused physical constraint feature vector and the time sequence process feature vector, the error propagation term is superimposed on the basis of the model prediction value, the measurement device error is propagated through the Jacobian matrix, and then the prediction value is corrected to obtain the error corrected prediction value.

[0009] Further, the multi-task outputs such as displacement and stress have strong physical coupling relationship, for example, bending deformation is related to the cross-sectional moment of inertia, the conventional single-task learning method ignores the physical constraints of the material mechanics principle, resulting in the prediction result violating the material constitutive relation and failing to meet the engineering precision requirement. In the residual attention mechanism module, the application adopts a physically constrained residual attention mechanism, takes the material mechanics parameters as a physical constraint matrix, and realizes the collaborative optimization of the multi-task output: a key vector is generated based on a structure parameter vector, an attention weight is calculated through a query-key matching mechanism, the attention weight is applied to the physical constraint matrix, a residual feature vector updated by the physical constraint is obtained through residual connection update of the feature representation.

[0010] Further, the dual-channel gating fusion module of the application is based on the residual feature vector updated by the physical constraint, constructs a dual-channel prediction framework, fuses the benchmark prediction and error compensation prediction through the gating mechanism, and combines the material time-varying characteristics for correction, and outputs a high-precision prediction value; specifically: the benchmark prediction channel: the residual feature vector updated by the physical constraint is input into a fully connected layer to output a benchmark prediction value; the error compensation channel: the error compensation prediction value is calculated by fusing the Bayesian uncertainty correction term and the material time-varying characteristics; the gating fusion mechanism: a dynamic gating weight is generated through the physical characteristics, the benchmark prediction value and the error compensation prediction value are fused to obtain the final prediction value.

[0011] Further, the steel structure construction precision prediction needs to meet the model precision requirement and the construction tolerance specification, the conventional mean square error loss only optimizes the average deviation of the prediction value and the true value, and cannot constrain the prediction error exceeding the engineering allowable tolerance range. The application constructs a hybrid loss that fuses the statistical loss and the engineering hard constraint, controls the prediction deviation distribution through segmented loss, and specifically: a segmented statistical loss term is constructed, the Huber loss defined by segmentation is used to replace the mean square error; an engineering tolerance constraint term is introduced, the tolerance threshold is set based on the construction specification, and a linear penalty is applied when the prediction error exceeds the threshold; the segmented statistical loss term and the engineering tolerance constraint term are weighted and summed to obtain the hybrid loss, which is used as the total optimization objective of the model training.

[0012] Further, parameter updating and model convergence condition judgment: in the training process of the error prediction model, an optimization algorithm based on gradient descent is adopted, the hybrid loss is used as the optimization objective for back propagation, in each iteration, the hybrid loss value is calculated based on the current batch of samples, the gradients of the trainable parameters of all error prediction models are calculated through the chain rule of differentiation, and the parameters are updated by using the optimizer.

[0013] Further, the modules of the error prediction model are connected through pooling operation or fully connected layer to complete the dimension transformation of the intermediate features.

[0014] The application has the following advantages: The application projects structure parameters, construction technology and environmental parameters to a unified feature space, and combines physical coupling items to realize physical correlation modeling between features, improve model convergence and interpretability; adopts time interval decay factor to simulate the "physical forgetting effect" between different processes in the construction process, effectively overcome the false dependence problem caused by ignoring process time in RNN and other methods, and improve the modeling ability of irregular time series data; combined with device technical manual parameters, construct variance propagation and Jacobian matrix propagation mechanism of measurement error, so that the prediction result can dynamically reflect the sensor accuracy difference and cumulative error, meet the high-precision construction monitoring requirements; create a "benchmark prediction + error compensation" double-channel structure, and combine the gating mechanism to adaptively fuse the outputs of the two channels according to the physical state, take into account the material constitutive constraint and on-site environmental / equipment uncertainty, realize high-robustness and high-precision error prediction. BRIEF DESCRIPTION OF DRAWINGS

[0015] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application, and are used to explain the application together with the embodiments of the application, and do not constitute a limitation on the application.

[0016] Figure 1 The step flowchart of the method of the application is shown in the figure; Figure 2 The structure diagram of the error prediction model of the application is shown in the figure; Figure 3 The prediction performance experiment figure of the multi-source data fusion of the application under temperature and humidity change is shown in the figure; Figure 4 The propagation control effect of the measurement equipment error of the traditional method is shown in the figure; Figure 5 The propagation control effect of the measurement equipment error of the method of the application is shown in the figure; Figure 6 The step flowchart of the double-channel gating fusion module of the application is shown in the figure; Figure 7 The loss function construction process of the application is shown in the figure; Figure 8 The influence comparison figure of the physical constraint on the prediction consistency of the application is shown in the figure. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0018] Embodiment 1 In this embodiment, asFigure 1 As shown, this invention provides a method for monitoring the prefabrication and assembly errors of steel box girder composite continuous box girder bridges, the specific steps of which include: S1. Collect and label the prefabrication and assembly error monitoring data of the steel box girder composite continuous box girder bridge. Structured data is collected through a multi-source sensor network deployed in prefabrication plants and construction sites, including: 1) Structural parameter acquisition: Use a 3D laser scanner to obtain the geometric dimensions of the components, and combine the design drawings to extract the segment number, component type, UHPC material ratio and preset tolerances to form a structured parameter table; 2) Construction process data collection: Based on construction logs and RFID tags, the mold number, curing method, tensioning process parameters, bolt specifications and assembly sequence are automatically recorded, and the assembly timestamp is marked through a high-precision clock synchronization system; 3) Environmental parameter acquisition: Temperature and humidity sensors, anemometers and light sensors are deployed in the assembly area to monitor temperature, humidity, wind speed, day and night temperature difference and construction period in real time; 4) Measurement and status parameter acquisition: The initial installation deviation is obtained by using a laser rangefinder, and the construction mechanical response data is recorded in real time by stress / strain / displacement sensors. The positioning system error is calculated according to the equipment technical manual.

[0019] Specifically, as shown in Table 1, the data collection content is categorized and examples are introduced below: Table 1. Classification and Examples of Data Collection Content Furthermore, data annotation is performed. Taking each prefabricated segment as a sample unit, the above four types of data are aligned by timestamp and integrated into a structured dataset. The annotation target value is the measured cumulative error after the segment is assembled. Specifically, the deviation between the actual pose and the designed pose is measured by a total station and used for supervised learning in subsequent model training.

[0020] S2. Error Prediction Model Construction and Training like Figure 2 As shown, the prediction model structure is defined as including a physical constraint feature fusion layer (S201), a causal convolutional layer (S202), a Bayesian uncertainty propagation layer (S203), a residual attention mechanism module (S204), and a dual-channel gating fusion module (S205). Optionally, pooling operations or fully connected layers can be used to connect the modules to achieve dimensional transformation of intermediate features.

[0021] S201, Physical Constraint Feature Fusion Layer: Physical constraint feature fusion of multi-source heterogeneous data The multi-source data such as structural parameters, construction technology and environmental parameters have the heterogeneous characteristics of multi-scale and multi-dimension, and the conventional feature direct splicing method causes difficulty in model convergence due to the difference in feature distribution, and ignores the strong correlation between physical parameters.

[0022] The application projects the structural parameters and the environmental parameters into a unified feature space through trainable weight matrices respectively, combines the coupling constraint term of the construction technology and the structural parameters, maps after the activation function, and then fuses, so as to eliminate the multi-dimension difference, model the physical constraint relationship, and be expressed as: , In the formula, is a structural parameter vector, for example, for The vector corresponds to the component type, the segment number, the component length, the UHPC ratio, the component height and the tolerance; is a construction technology vector, for example, for The vector corresponds to the mold number, the curing method, the tensioning method, the assembling method, the assembling sequence and the assembling time; is an environmental parameter vector, for example, for The vector corresponds to the temperature, the humidity, the wind speed, the temperature difference and the assembling period; is a structural parameter projection weight matrix, which is a trainable parameter; is an environmental parameter projection weight matrix, which is a trainable parameter; is a process-structure coupling weight matrix, which is a trainable parameter; is a vector splicing operation; is a Sigmoid activation function, and the output value range is ; is a hyperbolic tangent activation function, and the output value range is ; is a Hadamard product, and the implementation is element-by-element multiplication; is a fusion physical constraint feature vector.

[0023] It should be noted that The term represents feature projection, the structural and environmental parameters are projected into a unified feature space, the multi-dimension difference is eliminated, for example, the dimension difference between the length unit m and the temperature unit °C, the difficulty in model convergence is solved, The term represents a physical constraint term, the coupling relationship between the construction technology and the structural parameters is modeled, for example, the constraint of the bolt assembling method on the steel box girder stiffness, and the two are operated through the Hadamard product, so that the projected features are modulated by the physical constraint, for example, the thermal expansion effect is constrained by the bolt connection in a high-temperature environment, the prediction distortion caused by the conventional splicing method ignoring the physical correlation is avoided, and the model convergence speed is improved.

[0024] It should be further noted that the trainable parameters By independent pre-training mechanism optimization, specifically adopting physical consistency loss function Forward propagation calculation of the training process is performed, wherein, For the physical relationship function derived from the material mechanics formula, such as the bolt connection stiffness constraint equation, L2 norm, the trainable parameters are updated by error back propagation algorithm Make the projection feature space forced to satisfy the engineering physical law.

[0025] In one embodiment, the adaptability of the multi-source physical constraint feature fusion technology proposed in the application to environmental parameter changes is verified, such as Figure 3 As shown in the figure, the experiment compares the prediction error (mm) performance of the conventional physical model method and the method of the application under complex environmental conditions by systematically changing the environmental temperature (degrees Celsius) and the environmental humidity (percentage). The three-dimensional surface graph intuitively shows the prediction accuracy variation law of the two methods under different temperature and humidity combinations. The conventional physical model adopts a direct splicing method. From the experimental results, it can be seen that the prediction error of the conventional physical model method presents a sharp fluctuation with the change of the environment. In the high temperature and high humidity area (temperature higher than 30 degrees Celsius and humidity greater than 80% or so), the error surface of the conventional method appears obvious convex, indicating that the conventional method is sensitive to environmental parameters and is difficult to adapt to complex and changeable construction site conditions. In contrast, the error surface of the method of the application is more flat as a whole, maintaining a stable low error level within the temperature and humidity change range. Especially in the extreme environmental area, the fluctuation amplitude of the error surface of the method of the application is significantly smaller than that of the conventional method, proving the effectiveness of the multi-source physical constraint feature fusion technology. By projecting the structure parameters and the environmental parameters into a unified feature space and using the coupling constraint term of the construction process and the structure parameters, the multi-dimensional difference is eliminated, so that the model can adapt to the environmental change, and the prediction robustness is significantly improved.

[0026] S202, causal convolution layer: causal convolution extraction of process time sequence features The assembly sequence, time and other process parameters have strong time sequence dependence. The conventional RNN method ignores the physical meaning of the process interval, such as the weakening of the correlation between long-time interval processes, resulting in false correlation between non-continuous processes, which cannot accurately model the physical forgetting effect in actual construction.

[0027] The application adopts a gating causal convolution combined with a physical forgetting effect strategy to extract process features with time dependence. The specific steps are as follows: S2021, process time interval calculation Based on the timestamp information in the process vector, the physical time interval between the current time and the historical time is calculated and represented as: , In the formula, is a timestamp representing the assembly time; is a timestamp representing the assembly time; is a time index of the current process step, such as, represents the 3rd assembly process; is a relative delay index of the historical step, such as, represents the current time, represents the process back to the time; is a time interval calculation function for calculating the absolute time difference, with the output unit being hours; is the physical time interval between the time and the time.

[0028] In one embodiment, set , then .

[0029] S2022, causal convolution with forgetting effect In the causal convolution, an exponential decay term is combined to make the weight of the historical process feature decay with the increase of the physical time interval, thereby simulating the physical forgetting effect, suppressing noise interference, and improving the robustness to nonlinear time series patterns, which is expressed as: In the formula, is the process feature vector at the time; is the maximum time delay range, such as, ; is the th convolution kernel weight matrix, which is a trainable parameter; is a one-dimensional causal convolution operation, which only uses current and historical time data (t ), ensuring time causality, i.e., future data does not affect the present; is the decay coefficient, such as, ; is the absolute value operation; is the time series process feature vector extracted by convolution.

[0030] It should be noted that the term as a physical forgetting effect term produces engineering prior knowledge through physical forgetting, avoiding the limitations of manually setting a fixed time window, and the exponential decay increases with the increase of the physical time interval, such as, , , , and then simulate the weakening of the correlation of long-time interval processes in actual construction, such as the change of material performance after the curing period, and combine with the causal convolution to suppress the noise interference of non-continuous processes such as assembly after interval days, and improve the robustness to nonlinear time sequence mode.

[0031] It should be noted that the mold number, curing method, tensioning method, assembly method, assembly time period, and assembly sequence in the construction process vector are one-hot encoded, the assembly time of the numerical value type data is normalized, and all the processed features are spliced into a numerical vector, that is, a process feature vector is obtained, and the construction process vector at the moment corresponds to the process feature vector at the moment .

[0032] It should be noted that the time interval is dynamically controlled to forget and decay, so that the model is self-adaptive to the physical interval of the process, for example, the short-interval welding process retains a high weight, and the long-interval curing process reduces the weight.

[0033] S203, Bayesian uncertainty propagation layer: Bayesian uncertainty propagation of measurement error The measurement equipment such as laser range finder and positioning system has a non-Gaussian distribution error, and the conventional method ignores the transmission and accumulation effect of the error in the prediction model, so that the prediction result does not contain measurement uncertainty, and cannot meet the construction precision control requirement.

[0034] The application constructs a Bayesian error propagation layer to model the transmission process of measurement equipment error in the prediction process, and the specific steps are as follows: S2031, measurement equipment error variance calculation According to the component geometric parameters and the equipment technical manual, the error variances of various types of measurement equipment are calculated, the uncertainty of the measurement error is quantified, and is expressed as: , In the formula, is the measurement equipment type index, including laser range finder, positioning system, stress sensor, displacement sensor, etc.; is the error variance of the first type of measurement equipment; is the component length, in m; is the component height, in m.

[0035] It should be noted that and are the inherent error coefficients of the equipment, which are determined according to the technical standards in the art, for example, the accuracy of the laser range finder is , and the accuracy of the positioning system is ​It can also be adjusted according to the experience value, and the device inherent error coefficient (K) is calibrated according to the technical manual value. , The device inherent error coefficient (K) is calibrated according to the technical manual value.

[0036] S2032, uncertainty correction of the predicted value On the basis of the model predicted value, the error propagation term is superimposed, the measurement device error is propagated through the Jacobian matrix, and the predicted value is corrected to meet the construction precision control requirements, which is expressed as: , In the formula, is the predicted value corrected by the error; is the prediction model, is the parameter of the prediction model; is the measurement and state parameter of the first type of measurement device, which can be a scalar or a vector, and when is a vector, it can take multiple data values in the original data vector output by the measurement device; is the predicted value of the measurement and state parameter of the first type of measurement device; is the partial derivative (when is a scalar) or gradient (when is a vector) of with respect to is the total number of measurement device types.

[0037] In one embodiment, for the measurement and state parameter , the value is derived from the original data vector output by the measurement device, and is a scalar, the value is as follows: When indexing the laser ranging device, such as the component length measurement value, unit: meter; when indexing the positioning system device, such as the component height related coordinate, unit: m; when indexing the stress sensor, such as the stress sensor data during construction, unit: MPa; when indexing the strain sensor, such as the strain sensor data during construction, unit: ; when indexing the displacement sensor, such as the displacement sensor data during construction, unit: millimeter.

[0038] It should be further pointed out that the prediction model can adopt a neural network model, which can be a multilayer perceptron. The input of the neural network model is and , the output displacement or stress prediction value can be represented as , is the first weight parameter of the neural network model, and is the training parameter, is the second weight parameter of the neural network model, and is the training parameter, is the bias parameter of the neural network, and is the training parameter, is the ReLU activation function.

[0039] It should also be noted that, the term calculates the model prediction value, the term is an error propagation term, and when multi-task prediction is performed, the output is in vector format, and is a submatrix of the Jacobian matrix, which describes the partial derivative of the multi-output function with respect to all inputs, and through the Jacobian matrix, transmits to the prediction value, and the two jointly model the non-Gaussian error accumulation, ensuring the construction accuracy.

[0040] In one embodiment, the error propagation characteristics of four key measurement devices, namely laser range finder, positioning system, stress sensor, and displacement sensor, are focused on, as shown in Figure 4 , Figure 5 The control effect of the Bayesian uncertainty propagation layer proposed in the present application on measurement error is quantitatively analyzed, and the experiment visually displays the error sensitivity coefficients between different devices through a heat map. The deeper the color level, the higher the error propagation degree, Figure 4 The heat map in the conventional method shows the error propagation characteristics, and there is a significant cross-sensitivity phenomenon between devices. The laser range finder error has the greatest impact on the positioning system prediction result (the color block is the deepest), and there is also a strong correlation between the stress sensor and the displacement sensor, indicating that the conventional method cannot effectively isolate device errors, resulting in the amplification of measurement uncertainty in the system, Figure 5 The heat map in the present application shows the effect of the method. The error sensitivity coefficients between all devices are significantly reduced as a whole. The deepest color block is only equivalent to the medium sensitivity level of the conventional method. In particular, the influence coefficient of the laser range finder on the positioning system is reduced to less than 1 / 3 of the conventional method, and the cross-sensitivity between the stress sensor and the displacement sensor is also significantly weakened, proving the technical advantages of the Bayesian uncertainty propagation layer. Through the Jacobian matrix, the transmission path of the device error is accurately modeled, and combined with the error variance calculation mechanism, the diffusion of measurement error in the prediction process is effectively blocked, and the robustness of the system to device uncertainty is significantly improved.

[0041] S204, residual attention mechanism module: residual attention enhancement of multi-task output The outputs of multiple tasks, such as displacement and stress, have strong physical coupling relationships. For example, bending deformation is related to the moment of inertia of the cross section. Conventional single-task learning methods ignore the physical constraints of the principles of material mechanics, resulting in prediction results that violate the material constitutive relationship and cannot meet the engineering accuracy requirements.

[0042] This invention employs a physically constrained residual attention mechanism, combined with material mechanical parameters as the constraint matrix, to achieve collaborative optimization of multi-task outputs. The specific steps are as follows: S2041, Calculation of Physical Attention Weights A key vector is generated based on the structural parameter vector. Attention weights are calculated using a query-key matching mechanism, and these weights are forced to conform to the principles of materials mechanics to avoid violating physical constraints. This is represented as follows: , In the formula, This is the attention weight vector; The input feature vector for the query vector, specifically the prediction model. Input features The output after transformation by the hidden layer of the neural network; The projection weight matrix is ​​a trainable parameter for querying. Let be the key projection weight matrix, which are trainable parameters; This is a matrix transpose operation; Scaling factor To define the feature dimension, prevent the gradient from vanishing due to excessively large dot products; The output attention weight distribution is a normalized exponential function. .

[0043] It should be noted that conventional attention only calculates attention based on feature similarity, while Items are used as key vectors, based on This was obtained by considering material parameters as physical bonds, and by forcing... It conforms to the principles of mechanics of materials. For example, when predicting the displacement of concrete beams, the weight of the high elastic modulus region is increased to avoid violating constitutive relations, thereby solving multi-task coupling problems, such as the consistency problem of displacement-stress prediction.

[0044] S2042, Physical Constraint Residual Update Attention weights are applied to the physical constraint matrix, and the feature representation is updated through residual connections. By injecting material properties, the features are made to conform to physical principles such as beam bending theory, as shown below: , In the formula, The physical constraint matrix is ​​a diagonal matrix constructed from material mechanics parameters, and its calculation method is expressed as follows: This can be obtained from UHPC formulation data, such as... ; The elastic modulus of ultra-high performance concrete can be obtained from UHPC mix design data, such as... ; The moment of inertia of the steel box girder section can be calculated from the dimensions of the steel box girder, for example, , The width of the component; This is a layer normalization operation to stabilize the feature distribution; This represents the residual eigenvector updated by physical constraints.

[0045] It should be noted that, The term is a physical constraint feature correction term, through Injecting material properties, making It conforms to beam bending theory and avoids physical contradictions in the predicted values.

[0046] S205, Dual-channel Gating Fusion Module: Dual-channel prediction and gating fusion based on physical perception. Conventional single-channel prediction models struggle to balance physical laws with uncertainty compensation. While the baseline model conforms to material constitutive relations, it does not consider the cumulative effect of measurement errors. Purely data-driven models are susceptible to sensor noise interference, leading to predictions that violate beam bending theory. Furthermore, existing technologies fuse multiple models through weighted averaging, but fixed weights cannot adapt to dynamic changes in construction.

[0047] This invention constructs a dual-channel prediction framework based on physically constrained updated residual feature vectors. It fuses baseline predictions and error-compensated predictions through a gating mechanism and incorporates corrections based on time-varying material properties to output high-precision predicted values, such as... Figure 6 As shown, the specific steps are as follows: S2051, First Channel: Baseline Prediction Channel The residual eigenvector updated by physical constraints is input into the fully connected layer, which outputs a baseline prediction of displacement or stress, thereby providing a fundamental prediction under physical constraints to ensure compliance with material constitutive relations. This can be expressed as: , In the formula, For the first The physical constraint-updated residual feature vectors corresponding to each sample are used to update the physical constraint-updated residual feature vectors. Calculated for a single sample; The weight parameters of the fully connected layer are predicted based on the baseline and are trainable parameters. The bias parameters of the fully connected layer are predicted based on the baseline and are trainable parameters; The activation function for the Gaussian error linear unit; The weight parameters of the output layer are used as the baseline for prediction and are trainable parameters; The baseline prediction value of the i-th sample. The baseline prediction value of the i-th sample.

[0048] It is noted that in the solving process of the baseline prediction value , a combination of a fully connected layer and a GeLU activation function is adopted, the continuous and derivable characteristics of the GeLU function are used for smooth processing of feature nonlinear mapping, the input distribution is more sensitive, the feature patterns in a small error area can be better captured, and the residual feature vector with the material mechanics constraint is combined to make the baseline prediction value strictly follow the material constitutive relationship such as beam bending theory, and reduce the physical contradiction between stress and displacement prediction.

[0049] S2052, the second channel: error compensation channel The Bayesian uncertainty correction term and the material time-varying characteristics are fused to calculate the error compensation prediction value, compensate for the measurement uncertainty and the material time-varying effect, and are expressed as: , , In the formula, is the error sensitive coefficient of the i-th measurement device, which is a trainable parameter; is a timestamp at the moment, representing the assembly time; is the error compensation prediction value of the i-th sample; is the material forming timestamp, which is artificially preset and specifically set as the material forming moment; is a time-varying decay coefficient, such as ; is the elastic modulus of ultra-high performance concrete, which can be taken from the UHPC proportioning data, such as ; is a material time-varying characteristic decay function; is a natural exponential function. The term inherits the Bayesian uncertainty propagation, uses the trainable parameter to dynamically adjust the error contribution weight of different measurement devices, solves the error heterogeneity problem between devices, realizes error sensitive self-adaptation, and at the same time,

[0050] the term as an exponential decay term is calculated based on the time difference and the elastic modulus of ultra-high performance concrete , simulates the time-varying process such as concrete hardening, that is the increase will slow down the decay, and then compensate for the nonlinear aging effect.

[0051] ​​​S2053, gate fusion mechanism The dynamic gate weight is generated by the physical feature, the benchmark prediction and the error compensation prediction are fused, in the long process interval or the environment mutation scene, the prediction channel is adaptively selected, the deviation of the single channel is avoided, and the formula is represented as: , , In the formula, The vector splicing operation is represented as vector splicing operation; The attention weight vector of the i-th sample is represented as attention weight vector of the i-th sample, The attention weight vector is calculated for single sample input; The structure parameter vector is represented as structure parameter vector; The gate weight vector of the i-th sample is represented as gate weight vector of the i-th sample; The gate weight parameter is represented as gate weight parameter; The prediction value of the i-th sample is represented as prediction value of the i-th sample. It should be noted that the benchmark prediction channel provides the basic prediction under the physical constraint, ensures that the output conforms to the material mechanics principle, the error compensation channel focuses on correcting the measurement uncertainty and the time-varying effect, and the two are dynamically fused through the gate mechanism, when the fusion is carried out, the gate weight vector is calculated by splicing the features ,

[0052] The physical constraint state is encoded, The structure parameter is provided, and then the gate weight vector is used to realize adaptive fusion, When the physical feature is reliable, , the benchmark prediction is preferred, When the uncertainty is high, , the compensation prediction is preferred, in the long process interval or the environment mutation scene, the gate mechanism automatically balances the physical consistency and the error compensation, and the deviation of the single channel is avoided.

[0053] S206, loss function calculation The steel structure construction precision prediction needs to meet the model precision requirement and the construction tolerance specification, the conventional mean square error loss only optimizes the average deviation of the predicted value and the true value, and cannot constrain the prediction error to exceed the engineering allowable tolerance range.

[0054] The hybrid loss fusing the statistical loss and the engineering hard constraint is constructed, the prediction deviation distribution is controlled through the segmented loss, as shown in Figure 7 The specific steps are as follows: S2061, constructing a segmented statistical loss term ​The Huber loss defined by segmentation is used to replace the mean square error, which keeps smoothness when the prediction error is small, and switches to linear growth when the error is large, thereby improving the sensitivity of the model to small error areas, suppressing abnormal value interference, and accelerating convergence, which is expressed as: , In the formula, The Huber loss defined by segmentation is used to balance the smoothness of small error areas and the robustness of large error areas. is an input variable, specifically the prediction residual, and the calculation method is expressed as , The term represents the segmentation threshold set to . is the true measurement value of the th sample, such as the displacement sensor data of the th sample during construction (unit: millimeter); is the sample index; is the prediction value of the th sample.

[0055] It should be noted that the conventional mean square error is sensitive to abnormal values such as sensor failure points, and converges slowly, without considering the heavy-tailed distribution of construction data. The Huber loss used in the present application is smooth and optimized when the error is less than , and is robust and suppresses abnormal values when the error is greater than

[0056] S2062, introduce engineering tolerance constraint term Based on the construction specification, set the tolerance threshold, and apply linear penalty when the prediction error exceeds the threshold, thereby forcing the prediction value to meet the engineering allowable tolerance range, which is expressed as: , In the formula, is the engineering tolerance constraint term of the th sample; is the prediction error of the th sample, and the calculation method is expressed as ; is the construction tolerance set value, which is artificially preset, such as ; indicates the specific implementation method of the ReLU activation function, term indicates that the penalty is activated only when ; is the penalty intensity coefficient, which controls the weight of the constraint term, such as .

[0057] It should be noted that the construction tolerance set value This can be set using empirical values, such as those required by construction specifications. But reserved A safety margin of several times is provided to address unmodeled disturbances such as sudden changes in wind speed. The safety margin factor is used to extend the allowable error boundary. The term indicates when the absolute value of the prediction error exceeds Apply a linear penalty when the threshold is doubled to ensure that all predicted values ​​meet the condition. .

[0058] S2063, Integrated Hybrid Loss The piecewise statistical loss term and the engineering tolerance constraint term are weighted and summed to form the overall optimization objective, ensuring that the model output simultaneously considers statistical accuracy and engineering safety margin, expressed as: , In the formula, The mixed loss is used as the overall optimization objective for model training. This refers to the batch sample size. This represents the segmented Huber loss, used to balance the smoothness of small error regions with the robustness of large error regions.

[0059] S207, Parameter Update and Model Convergence Condition Judgment The error prediction model is trained using a gradient descent-based optimization algorithm with a mixed loss. To optimize the objective, backpropagation is performed. In each iteration, the mixed loss value is calculated based on the current batch of samples. The gradients of the trainable parameters of the model are calculated using the chain rule of differentiation. The parameters are then updated using the optimizer to minimize the loss.

[0060] The convergence of the error prediction model is judged using a dual criterion, as follows: 1) From a statistical perspective, the mixed loss value on the monitoring validation set, if the mixed loss is over 10 consecutive periods... If the function value decreases by less than a preset threshold, it is considered statistically converged; for example, the preset threshold is 0.1%. 2) At the engineering level, the proportion of samples whose prediction error in the validation set exceeds the engineering tolerance constraint must be consistently below a safety threshold. For example, the engineering tolerance constraint is set to... times The safety threshold is set to 5%.

[0061] Training is terminated and the optimal model parameters are saved only when both statistical convergence and engineering constraints are met.

[0062] In one embodiment, the influence of physical constraints on prediction consistency is analyzed, verifying the key role of the physical constraint mechanism in maintaining the consistency of the prediction results with engineering theory. In the monitoring of bridge construction errors, the stress prediction value must strictly follow the material mechanics principles such as beam bending theory, otherwise it will lead to engineering decision-making errors, such as Figure 8 As shown in the figure, the experiment verifies the effectiveness of the physical constraint residual attention mechanism in the present technology by comparing the degree of conformity of stress prediction values under different constraint conditions with theoretical values. The experiment compares the prediction methods of three different constraint conditions: 1) no physical constraint method: only a conventional neural network is used for prediction, without using any physical constraint; 2) basic physical constraint method: a simple physical constraint is introduced into the neural network, such as material properties as feature input; 3) present technology method: the physical constraint residual attention mechanism proposed in the present technology is used to integrate the physical constraint matrix constructed by material mechanics parameters into the attention mechanism. The abscissa of the experimental graph is the beam position (unit: meters), representing different measurement points along the longitudinal direction of the bridge, and the ordinate is the stress prediction value (unit: megapascal), reflecting the structural stress state. The theoretical baseline (black dotted line) is the theoretical stress distribution calculated based on the beam bending theory, showing a sinusoidal wave pattern, reflecting the standard stress characteristics of different positions of the bridge, which serves as a standard for evaluating prediction accuracy. From the results of the no physical constraint prediction (scattered point distribution), it can be seen that the prediction points show a clear trend of deviating from the theoretical curve, with systematic deviations at the 5-meter and 10-meter positions of the beam, and a large degree of dispersion, proving that ignoring physical constraints can easily lead to predictions that violate the material constitutive relationship. For the basic physical constraint prediction (scattered point distribution), the prediction points are basically distributed around the theoretical curve, with good agreement at positions 0-3 meters, but a slight downward deviation at positions 7-10 meters, with a moderate degree of dispersion, indicating that simple physical constraints can improve but not completely solve the consistency problem. The present technology prediction (scattered point distribution) closely matches the theoretical curve distribution, with no systematic deviation throughout the entire process, especially at the key positions (peaks / troughs), with precise matching and the smallest degree of dispersion, proving that the residual attention mechanism can effectively maintain physical consistency, and confirming the core advantage of the physical constraint residual attention mechanism in the present technology, which integrates material mechanics parameters as hard constraints into feature representation, making the prediction results strictly follow the engineering physical principles. Compared with traditional methods, the present technology can effectively eliminate systematic bias and provide reliable protection for bridge construction precision control.

[0063] S3, error prediction and monitoring After the error prediction model is trained, in the model inference stage, for the collected structural parameters, construction process, environmental parameters, and measurement and state parameters of the new precast segment, align the timestamps and integrate them into input samples that are isomorphic to the training set, and input the samples into the trained error prediction model; The error prediction model first unifies the dimensions of multiple sources of data and embeds physical relationships through the physical constraint feature fusion layer (S201). Then, the process timing features with forgetting effect are extracted by a causal convolution layer (S202); Then, the equipment error is corrected by a Bayesian uncertainty propagation layer (S203); Then, the physical consistency of multi-task output is dynamically coordinated by a residual attention mechanism (S204); Finally, the predicted value is generated by a dual-channel gating fusion module (S205).

[0064] All the predicted data are stored in association with the process parameters to form a digital construction file, supporting dynamic visual management and quality traceability of the full-bridge assembly precision.

[0065] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacements to some technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A steel box superimposed continuous box girder bridge prefabrication and assembly error monitoring method, characterized in that, The method comprises the following steps: Step 1: Collecting steel box composite continuous box girder bridge precast and assembly error monitoring data and labeling the data; Step 2: Building an error prediction model, including a physical constraint feature fusion layer, a causal convolution layer, a Bayesian uncertainty propagation layer, a residual attention mechanism module, and a double-channel gate fusion module; the labeled data is used to train the error prediction model; During the training process, the model is optimized by building a hybrid loss of fusion statistical loss and engineering hard constraint, and a trained model is obtained; Step 3: The structural parameters, construction technology, environmental parameters, and measurement and state parameters of the collected new precast segment are aligned with the timestamp and integrated into an input sample with the same structure as the training set, which is input into the trained model to obtain the prediction result.

2. The method according to claim 1, characterized in that, The steel box composite continuous box girder bridge precast and assembly error monitoring data includes structural parameters, construction technology, environmental parameters, and measurement and state parameters.

3. The method according to claim 2, characterized in that, The structural parameters and environmental parameters are projected into a unified feature space through a trainable weight matrix, and the coupling constraint term of the construction technology and the structural parameters is fused after being mapped by an activation function to model the physical constraint relationship and obtain a fused physical constraint feature vector.

4. The method according to claim 3, characterized in that, In the causal convolution layer, the process feature is extracted by using a gated causal convolution combined with a physical forgetting effect strategy: Based on the timestamp information in the process vector, the physical time interval between the current time and the historical time is calculated; in the causal convolution, an exponential decay term is combined to make the weight of the historical process feature decay with the increase of the physical time interval, simulating the physical forgetting effect, and obtaining the time sequence process feature vector extracted by convolution.

5. The method according to claim 4, characterized in that, The transmission process of the measurement device error in the prediction process is modeled by the Bayesian error propagation layer: according to the component geometric parameters and the device technical manual, the error variance of each type of measurement device is calculated; Based on the fused physical constraint feature vector and the time sequence process feature vector, the error propagation term is superimposed on the basis of the model prediction value, the measurement device error is propagated through the Jacobian matrix, and then the prediction value is corrected to obtain the error corrected prediction value.

6. The method according to claim 5, characterized in that, In the residual attention mechanism module, the residual attention mechanism of the physical constraint is used, the material mechanics parameters are taken as the physical constraint matrix, and the collaborative optimization of multi-task output is realized: a key vector is generated based on the structural parameter vector, the attention weight is calculated through the query-key matching mechanism; the attention weight is applied to the physical constraint matrix, the feature representation is updated through the residual connection, and the residual feature vector updated by the physical constraint is obtained.

7. The method according to claim 6, characterized in that, The double-channel gate fusion module uses a double-channel prediction framework and outputs the final prediction value based on a gate fusion mechanism: Baseline prediction channel: the residual feature vector updated by the physical constraint is input into a fully connected layer to output a baseline prediction value; Error compensation channel: the error compensation prediction value is calculated by fusing the Bayesian uncertainty correction term and the material time-varying characteristics; gate fusion mechanism: dynamic gate weight is generated by physical features, and the baseline prediction value and the error compensation prediction value are fused to obtain the final prediction value.

8. The method according to claim 7, characterized in that, A hybrid loss function combining statistical loss and engineering hard constraints is constructed: a piecewise statistical loss term is constructed, and a piecewise defined Huber loss is used instead of mean square error; an engineering tolerance constraint term is introduced, and a tolerance threshold is set based on construction specifications, and a linear penalty is applied when the prediction error exceeds the threshold; the piecewise statistical loss term and the engineering tolerance constraint term are weighted and summed to obtain the hybrid loss, which is used as the total optimization objective of model training.

9. The method according to claim 8, characterized in that, Parameter updating and model convergence condition judgment: in the training process of the error prediction model, the optimization algorithm based on gradient descent is adopted, the hybrid loss is used as the optimization objective for back propagation, in each iteration, the hybrid loss value is calculated according to the current batch of samples, the gradient of all trainable parameters of the error prediction model is calculated through the chain rule of differentiation, and the parameters are updated by using the optimizer.

10. The method according to claim 9, characterized in that, The modules of the error prediction model are connected by pooling operation or fully connected layer to complete the dimension transformation of the intermediate features.

Images