Multi-source data and mechanical mechanism driven arch bridge support state prediction method and system

Abstract

The invention relates to a multi-source data and mechanical mechanism driven arch bridge support state prediction method and system, and the method comprises the steps: 1, analyzing construction conditions through building a fine finite element model, and determining the most unfavorable position of a support; 2, a plurality of sensors are arranged at the most unfavorable positions to acquire strain, temperature and displacement data, and a multi-source monitoring data stream is formed; 3, preprocessing the data, including data cleaning, time alignment and normalization, and constructing a supervised learning data set; 4, generating a sample set by using a sliding window method, and dividing the sample set into a training set and a test set by adopting layered sampling; 5, constructing a multi-output prediction model based on attention-LSTM (Long Short Term Memory), and 6, training by using the training set, setting a weighted mean square error loss function and an Adam optimizer, and applying a learning rate attenuation strategy until the performance of the model is converged. And 7, evaluating the performance of the model by using the test set, and setting an early warning threshold value so as to realize real-time accurate prediction of the state of the arch bridge support and provide intelligent guarantee for construction safety.

Description

Technical Field

[0001] This invention belongs to the field of bridge engineering construction monitoring and safety control technology, and specifically relates to a method and system for predicting the state of arch bridge supports driven by multi-source data and mechanical mechanisms. Background Technology

[0002] In the cast-in-place construction of the main arch ring of a tied arch bridge, the scaffolding, as a temporary support structure, directly affects the success or failure of the entire project. Traditional monitoring methods mainly rely on manual, periodic data collection, which suffers from low data collection frequency and response lag, making it difficult to achieve real-time perception and early warning of the scaffolding's condition. Existing prediction models are mostly based on single-type data (such as displacement or stress alone), failing to fully utilize the inherent correlations between multi-source heterogeneous data (stress, temperature, displacement) generated simultaneously during construction. Furthermore, conventional time series prediction models (such as ARIMA and simple RNNs) struggle to effectively capture the long-term impact of critical time points such as sudden changes in construction loads and temperature on the scaffolding's condition, resulting in insufficient prediction accuracy and an inability to provide forward-looking judgments on the scaffolding's safety status. Therefore, a novel prediction method is urgently needed that can integrate multi-source data, incorporate mechanical mechanisms, and possess adaptive critical event perception capabilities. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides a method and system for predicting the state of arch bridge supports driven by multi-source data and mechanical mechanisms. The aim is to achieve real-time and accurate prediction of the multi-variable states of arch bridge supports, such as strain, displacement, and temperature, during construction by integrating mechanical mechanisms and data-driven technology, thereby providing intelligent protection for construction safety.

[0004] To achieve the above objectives, the specific solution of the present invention is as follows:

[0005] A method for predicting the state of arch bridge supports driven by multi-source data and mechanical mechanisms includes the following steps:

[0006] Step 1, Mechanical Mechanism Analysis and Monitoring Point Location: Based on the design drawings and actual conditions, establish a detailed finite element model of the tie-arch bridge support construction. By simulating the entire process of segmented casting of the main arch ring, analyze the mechanical response of the support under different construction conditions and determine the most unfavorable load-bearing position of the support during the entire construction period.

[0007] Step 2, Multi-source monitoring data acquisition: At the most unfavorable location determined in Step 1, various sensors, including vibrating wire strain gauges and 360° prisms, are deployed to construct a multi-source monitoring system. Strain and temperature data are collected synchronously through the vibrating wire strain gauges, and displacement data is measured using a total station and 360° prisms. The vertical displacement data of the support is obtained by calculating the elevation at each moment. The collected strain, temperature, and displacement data are integrated to form a multi-source monitoring data stream.

[0008] Step 3, data preprocessing: data cleaning and time alignment are performed on the raw monitoring data stream of the multi-source monitoring data in step 2 to form a regular multi-dimensional time series data set, and normalization processing is performed on the regular multi-dimensional time series data set to obtain a normalized multi-source monitoring data set, the multi-source monitoring data set including normalized strain, temperature and displacement data;

[0009] Step 4, supervised learning data set construction: the normalized multi-source monitoring data in step 3 is processed using a sliding window method to generate a sample set for supervised learning. In the sliding window method, the strain, temperature and displacement data of the past N time steps are set as input features, where each time step includes normalized strain, temperature and displacement, and the strain, temperature and displacement variables of the future M time steps are set as prediction targets to constitute prediction target variables. The sliding window is slid over the entire normalized multi-source data set, one time step at a time, to generate a series of sample pairs. All generated sample pairs are combined to form a sample set. The sample set is divided into a training set and a test set in a 70:30 ratio using a stratified sampling method. The training set is used for training the attention-LSTM multi-output prediction model, and the test set is used to evaluate the model performance.

[0010] Step 5, construction of attention-LSTM multi-output prediction model: using the training set divided in step 4, an attention-LSTM multi-output prediction model containing an input layer, two layers of stacked LSTM layers, an attention mechanism layer and a multi-output regression layer is constructed.

[0011] Step 6: model training and optimization: the attention-LSTM multi-output prediction model constructed in step 5 and the training set divided in step 4 are used for training. The loss function is set as weighted mean square error, and the weights are set according to the influence of strain, temperature and displacement on engineering safety. The Adam algorithm is selected as the optimizer, and learning rate decay strategy is applied. Iterative training is performed until the model performance on the validation set converges.

[0012] Step 7: model performance evaluation and early warning: the test set divided in step 4 is used to evaluate the performance of the deep learning model trained in step 6. Root mean square error, mean absolute error and coefficient of determination are used as precision measurement standards for the model. According to the model prediction results and combining with the actual engineering, a warning threshold is set. When the predicted value exceeds the threshold, an early warning is triggered.

[0013] Further, the formula for calculating the elevation change to obtain the vertical displacement data of the support in step 2 is as follows:

[0014] ,

[0015] In the formula, represents the vertical displacement of the support at a certain moment in time; is the initial reference period prism center point elevation, is the t period prism center point elevation.

[0016] Further, the data cleaning in step 3 is for missing values caused by transmission and storage of raw data in multi-source monitoring data stream, and linear interpolation or time series interpolation method is used for filling; for abnormal values obviously deviating from the normal range, the quartile range method is combined for identification and elimination; the time alignment is to resample and align all sensor multi-source monitoring data streams with a unified time reference to form a regular multi-dimensional time series data set; the data normalization processing adopts Min-Max normalization method, and the value of each feature variable is mapped to the interval [0, 1], and the expression of data normalization processing is as follows:

[0017] ,

[0018] In the formula: and ; represents the feature value in the original sample data, which belongs to the vector in the real number field ; represents the minimum value of the feature in all samples, that is ; represents the maximum value of the feature in all samples, that is ; represents the normalized feature value, which belongs to the vector in the real number field , and its value range is [0, 1] or [-1, 1].

[0019] Further, the input layer in step 5 is used to receive a tensor with a shape of (batch size , N, 3), where N is the historical time step, and 3 represents three features of strain, temperature and displacement;

[0020] The LSTM layer adopts a two-layer stacked LSTM network to learn the complex time dependence in the input sequence deeply and efficiently;

[0021] The attention mechanism layer is used to automatically learn and assign the importance weight of different historical moments to the current prediction, and the calculation steps of the attention mechanism layer are as follows:

[0022] Step 51, calculate the attention score: take the last hidden state h N as the query, and calculate the hidden state h N and each historical hidden state h tThe formula for the relevance score of the key is as follows:

[0023] ,

[0024] In the formula, e t W represents the attention score at time step t; α U α These are the learnable weight matrices; v α It is a learnable weight vector; T represents the transpose of a vector or matrix; That is to say, for the weight vector v α Perform transpose; tanh(·) is the hyperbolic tangent activation function; h t h represents the historical hidden state at time step t; N This represents the hidden state at the last time step;

[0025] Step 52, Normalization to Weights: Use the Softmax function to normalize the scores of all time steps into a probability distribution to obtain the attention weights α. t ;

[0026]

[0027] In the formula, This represents the attention weight corresponding to the t-th time step; N represents the total length of the input sequence, i.e., the total number of time steps. This represents the attention score at time step t. The result of the exponentiation operation; exp(e j ) represents the attention score at time step j. The result of the exponential operation; This represents the summation of attention scores over all time steps (from step 1 to step N) after an exponential operation.

[0028] Step 53, Generate context vector: All weights α t With the corresponding hidden state h t The values ​​are weighted and summed to obtain a context vector C, and the formula for the context vector C is as follows:

[0029] ,

[0030] In the formula, N represents the total length of the input sequence, i.e., the total number of time steps; α t h represents the attention weight at the i-th time step; t This represents the hidden historical state at time step t.

[0031] The multi-output regression layer is used to transform and integrate the context vector through a shared fully connected layer, and the integrated features are sent to three independent output layers to generate the predictions of stress, temperature and displacement at M future time steps;

[0032] The formula of the feature transformation and integration is as follows:

[0033] ,

[0034] In the formula, C' represents the new features output after the feature transformation and integration; W S represents the learnable weight matrix corresponding to the shared fully connected layer; C represents the context vector; b s represents the learnable bias vector corresponding to the shared fully connected layer; ReLU(⋅) represents the linear rectified activation function, which is defined as ReLU(x)=max(0,x);

[0035] The prediction of the stress: ,

[0036] The prediction of the temperature: ,

[0037] The prediction of the displacement: ,

[0038] In the formula, W , respectively represent the learnable weight matrix corresponding to the stress, temperature and displacement prediction branch; respectively represent the learnable bias vector corresponding to the stress, temperature and displacement prediction branch;

[0039] Further, the loss function in step 6 is defined as:

[0040] ,

[0041] In the formula, Total Loss represents the loss function; w Ɛ , w t , w D are the weights respectively given to the strain, temperature and displacement variables according to the importance of engineering safety; MSE( Ɛture , Ɛpred ), MSE(t true , t pred ), MSE(D true , D pred ) respectively represent the mean square error loss between the predicted value and the true value of the strain, temperature and displacement;

[0042] The formula of the mean square error is:

[0043] ,

[0044] The formula of the average absolute error is:

[0045] ,

[0046] The formula of the determination coefficient is:

[0047] ,

[0048] In the formula, MSE is the mean square error; MAE is the mean absolute error; R 2 represents the determination coefficient; n is the total number of samples; y i is the measured value of the ith sample; is the measured mean value; p i is the predicted value of the ith sample by the model.

[0049] Further, the early warning mechanism in step 7 includes a first yellow warning, a second orange warning, and a third red warning;

[0050] First yellow warning: when the predicted displacement value exceeds 70% of the design allowable value or the predicted stress value exceeds 70% of the material allowable strain, a general warning is given;

[0051] Second orange warning: when the predicted displacement value exceeds 80% of the design allowable value or the predicted stress value exceeds 80% of the material allowable strain, a high-risk warning is given, prompting attention to a specific area;

[0052] Third red warning: when the predicted displacement value exceeds 90% of the design allowable value or the predicted stress value exceeds 90% of the material allowable strain, a high-risk warning is given, and the dangerous rod number is accurately located, suggesting immediate intervention measures.

[0053] An arch bridge support state prediction system implementing the method comprises:

[0054] A mechanism analysis module for establishing a refined finite element model, simulating the segmented pouring construction process, and identifying the most unfavorable position of the support;

[0055] A data acquisition and processing module for laying multiple source sensors, collecting strain, temperature, and displacement data, performing data cleaning, alignment, and normalization processing, and constructing a sliding window sample;

[0056] The attention-LSTM prediction model module comprises an input layer, an LSTM layer, an attention mechanism layer and a multi-output regression layer, the input layer is used for receiving normalized multi-source monitoring data, the LSTM layer is used for learning time dependence in the input sequence, the attention mechanism layer is used for automatically learning and assigning importance weights of different historical moments to the current prediction, and the multi-output regression layer is used for performing feature transformation and integration of the context vector through a shared fully connected layer, and sending the integrated features to three independent output layers of strain, temperature and displacement respectively to generate stress, temperature and displacement predictions of future M time steps for generating strain, temperature and displacement predictions of future time steps.

[0057] The safety warning module is used for inputting real-time data, outputting prediction results and triggering graded warnings according to threshold values.

[0058] Advantages of the present application

[0059] 1. Significant advantages of multi-source data fusion: The present application fully utilizes multi-source heterogeneous data such as stress, temperature, displacement, etc., and deeply mines the internal correlation between variables, thereby comprehensively and multi-angulary perceiving the state of the arch bridge support, greatly improving the comprehensiveness of state perception, avoiding the one-sidedness possibly caused by a single data source, and providing a richer information basis for accurate prediction.

[0060] 2. Deep fusion of mechanism and data driving: The present application accurately determines key monitoring points through finite element analysis, makes data acquisition more targeted, and ensures that the collected data has more value. At the same time, the model prediction has better interpretability due to the incorporation of physical mechanism, solving the deficiency of pure data-driven methods in physical interpretability, and making the prediction results not only accurate but also easy to understand and apply.

[0061] 3. Strong self-adaptive key event perception capability: The present application introduces an LSTM model with attention mechanism, which can adaptively focus on historical key periods such as concrete pouring period and temperature sudden change moment. This feature significantly improves the prediction accuracy of load mutation and other special working conditions, enhances the robustness of the model in complex construction environment, and makes it better cope with various unexpected situations in actual construction.

[0062] 4. Strong practicality of multi-variable synchronous prediction: The present application adopts a multi-output regression structure to realize synchronous prediction of stress, temperature, displacement and other multi-variables, which is more in line with the actual monitoring needs of engineering. This synchronous prediction method not only improves the prediction efficiency, but also facilitates construction personnel to comprehensively master the real-time state of the arch bridge support, providing more comprehensive and timely support for construction decision-making.

[0063] 5. Engineering applicability flexibility and high efficiency: the application can flexibly adjust the model optimization direction according to the engineering safety emphasis through the carefully designed weighted loss function. This flexibility makes the application widely applicable to different engineering scenes, highlights the prediction accuracy of key variables according to the specific needs of different projects, and has very high practical value.

[0064] 6. Excellent prediction performance: the application realizes the synchronous and accurate prediction of the quasi-static support system state (stress, displacement) and its multivariate affected by temperature, not only improves the comprehensiveness of prediction, but also enhances the applicability of prediction results in engineering practice, provides a precise, real-time and forward-looking prediction method for arch bridge support construction safety, and effectively promotes the intelligent development of bridge construction monitoring, which has important practical significance and broad application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 The flow chart of the arch bridge support construction state prediction method driven by multi-source data and mechanical mechanism.

[0066] Figure 2 The flow chart of the arch bridge support construction state prediction method driven by multi-source data and mechanical mechanism. Figure 1

[0067] The flow chart of the arch bridge support construction state prediction method driven by multi-source data and mechanical mechanism. Figure 3 Figure 1 The flow chart of the arch bridge support construction state prediction method driven by multi-source data and mechanical mechanism.

[0068] Figure 4 Figure 1 The flow chart of the arch bridge support construction state prediction method driven by multi-source data and mechanical mechanism.

[0069] Figure 5 The flow chart of the arch bridge support construction state prediction method driven by multi-source data and mechanical mechanism. Figure 1

[0070] The flow chart of the arch bridge support construction state prediction method driven by multi-source data and mechanical mechanism. Figure 6

[0071] The flow chart of the arch bridge support construction state prediction method driven by multi-source data and mechanical mechanism. Figure 7 Figure 6 The flow chart of the arch bridge support construction state prediction method driven by multi-source data and mechanical mechanism.

[0072] Figure 8 The flow chart of the arch bridge support construction state prediction method driven by multi-source data and mechanical mechanism.

[0073] The application will be further explained and described below in combination with the drawings and specific embodiments, and it should be noted that the specific embodiments are not used to limit the scope of the application.

[0074] As Figures 1 to 5 ​​​​The multi-source data and mechanical mechanism driven arch bridge support state prediction method provided by the embodiment comprises the following steps:

[0075] Step 1, mechanical mechanism analysis and monitoring point positioning: according to the design drawing and the actual situation, a fine finite element model of the tied-arch bridge support construction is established, the whole process of the segmented pouring of the main arch ring is simulated, the mechanical response of the support under different construction conditions is analyzed, and the most unfavorable position of the support in the whole construction period is determined; the most unfavorable position is the focus of subsequent sensor layout and state prediction. The finite element analysis not only considers the static load, but also takes into account the time-varying effects such as temperature change, concrete shrinkage and creep, so as to ensure the accuracy and representativeness of the identified key position.

[0076] Step 2, multi-source monitoring data acquisition: a plurality of sensors including a vibrating wire strain gauge and a 360 prism are arranged at the most unfavorable position determined in step 1 to build a multi-source monitoring system, strain data and temperature data are synchronously collected by the vibrating wire strain gauge, displacement data are measured by the total station and the 360 prism, the three-dimensional coordinates of the prism are automatically measured at the same time interval as the vibrating wire strain gauge, the vertical displacement data of the support are obtained by calculating the elevation at each time, and the collected strain, temperature and displacement data are integrated to form a multi-source monitoring data stream;

[0077] The calculation formula of the elevation at each time is as follows:

[0078] (1),

[0079] In the formula, represents the prism center point elevation; represents the known elevation of the total station station; represents the instrument height, that is, the height from the station point to the horizontal axis of the total station; represents the slant distance from the total station to the prism; represents the vertical angle, that is, the height angle, represents the prism height, that is, the height from the measuring point to the prism center.

[0080] Based on the elevation at each time obtained by the above calculation, the vertical displacement of the support at a certain time is obtained as follows:

[0081] (2),

[0082] In the formula, is the initial reference period prism center point elevation, is the t period prism center point elevation.

[0083] ​​Step 3, data preprocessing: data cleaning and time alignment are performed on the original monitoring data stream of the multi-source monitoring data described in step 2 to form a regular multi-dimensional time series data set, and normalization processing is performed on the regular multi-dimensional time series data set to obtain a normalized multi-source monitoring data set, the multi-source monitoring data set including normalized strain, temperature and displacement data;

[0084] The data cleaning is for the original data abnormality or missing problem caused by uncontrollable factors such as network signal loss, sensor failure, etc. The missing values caused by transmission and storage of the original data in the multi-source monitoring data stream are filled by linear interpolation or time series interpolation method. For abnormal values obviously deviating from the normal range, such as spikes caused by transient interference of equipment, the four quantile distance method is used for identification and elimination. The interpolation or marking processing is combined with the construction log to ensure that the data quality meets the model training requirements.

[0085] The time alignment is to resample and align all sensor multi-source monitoring data streams with a unified time reference (such as UTC time) to form a regular multi-dimensional time series data set {(Ɛ i ,T i ,D i )}, wherein i is the time point index.

[0086] Since the dimensions and orders of magnitude of strain, temperature and displacement data are different, standardization processing is required to eliminate the dimension effect. The Min-Max data normalization method is used to map the value of each feature variable to the interval [0, 1]. This process unifies the data scale, standardizes the variable dimension, ensures that each feature can make balanced and effective contribution in the subsequent prediction model, and accelerates the convergence process of model training.

[0087] The expression of data normalization processing is as follows:

[0088] (3),

[0089] In the formula: and ; denotes the feature value in the original sample data, which belongs to the vector in the real number field ;

[0090] denotes the minimum value of the feature in all samples, that is, ;

[0091] denotes the maximum value of the feature in all samples, that is, ;

[0092] The normalized eigenvalues ​​belong to the real number field. The vector in the vector has a value range of [0,1] or [-1,1].

[0093] Step 4, Supervised Learning Dataset Construction: The normalized multi-source monitoring data from Step 3 is processed using the sliding window method to generate a sample set suitable for supervised learning. In the sliding window method, strain, temperature, and displacement data from the past N time steps are set as input features, where each time step includes normalized strain, temperature, and displacement. Strain, temperature, and displacement variables from the next M time steps are set as prediction targets, constituting the prediction target variables. For example, if N = 48 hours and M = 12 hours, then each sample uses data from the past 48 hours to predict the multivariate state for the next 12 hours. The sliding window slides across the entire normalized multi-source dataset, moving one time step at a time, generating a series of sample pairs. All generated samples are combined into a sample set. A stratified sampling method is used on the sample set, dividing it into a training set and a test set at a ratio of 70% and 30%. The training set is used to train the attention-LSTM-based multi-output prediction model, and the test set is used to evaluate the model performance.

[0094] Step 5, Construct a multi-output prediction model based on attention-LSTM: Using the training set divided in Step 4, construct a multi-output prediction model based on attention-LSTM that includes an input layer, two stacked LSTM layers, an attention mechanism layer, and a multi-output regression layer.

[0095] The input layer is used to receive batches of shape (batch) size ,N,3) is a tensor X, where N is the history time step and 3 represents the three features of strain, temperature and displacement;

[0096] The LSTM layer employs a two-layer stacked LSTM network for deep and efficient learning of complex temporal dependencies in the input sequence. The LSTM's gating mechanism effectively captures long-term dependencies, avoiding the vanishing or exploding gradient problem.

[0097] The attention mechanism layer is used to automatically learn and assign importance weights for different historical moments to the current prediction. Introducing the attention mechanism after the LSTM layer enables the attention-LSTM multi-output prediction model to automatically learn and assign importance weights for different historical moments to the current prediction. For example, the attention-LSTM multi-output prediction model focuses on the time period when concrete pouring causes a sudden increase in load, thereby improving the prediction accuracy for critical conditions. Visualizing the attention weights can also provide decision-making references for engineers.

[0098] The computational steps of the attention mechanism layer are as follows:

[0099] Step 51, Calculate the attention score: for the last hidden state h N As a query, calculate the hidden state h. N With each historical hidden state h in the sequence t The formula for the relevance score of the key is as follows:

[0100] (4),

[0101] In the formula, e t W represents the attention score at time step t; α U α These are the learnable weight matrices; v α It is a learnable weight vector; T represents the transpose of a vector or matrix; That is to say, for the weight vector v α Perform transpose; tanh(·) is the hyperbolic tangent activation function; h t h represents the historical hidden state at time step t; N This represents the hidden state at the last time step;

[0102] Step 52, Normalization to Weights: Use the Softmax function to normalize the scores of all time steps into a probability distribution to obtain the attention weights α. t :

[0103]

[0104] In the formula, This represents the attention weight corresponding to the t-th time step; N represents the total length of the input sequence, i.e., the total number of time steps. This represents the attention score at time step t. The result of the exponentiation operation; exp(e j ) represents the attention score at time step j. The result of the exponential operation; This represents the summation of attention scores over all time steps (from step 1 to step N) after an exponential operation.

[0105] Step 53, Generate context vector: All weights α t With the corresponding hidden state h t The values ​​are weighted and summed to obtain a context vector C, and the formula for the context vector C is as follows:

[0106] (6),

[0107] In the formula, N represents the total length of the input sequence, that is, the total number of time steps; the attention weight of the t-th time step; the history hidden state of the t-th time step.

[0108] The multi-output regression layer is used to transform and integrate the context vector through a shared fully connected layer, and the integrated features are sent to three independent output layers to generate the prediction of stress, temperature and displacement in the next M time steps;

[0109] This multi-output structure realizes multi-variable synchronous prediction, ensuring the synchronization and coordination of variable prediction.

[0110] The formula of the feature transformation and integration is as follows:

[0111] (7),

[0112] In the formula, C' represents the new feature representation output after feature transformation and integration; W S represents the learnable weight matrix corresponding to the shared fully connected layer; C represents the context vector; b s represents the learnable bias vector corresponding to the shared fully connected layer; ReLU(·) represents the linear rectification activation function, which is defined as ReLU(x) = max(0, x);

[0113] The prediction of stress: (8),

[0114] The prediction of temperature: (9),

[0115] The prediction of displacement: (10),

[0116] In the formula, , W Ɛ , W t , W D respectively represent the learnable weight matrix corresponding to the stress, temperature and displacement prediction branch; b Ɛ , b t , b D respectively represent the learnable bias vector corresponding to the stress, temperature and displacement prediction branch.

[0117] Step 6: Model training and optimization: use the deep learning model constructed in step 5 and the training set divided in step 4 for training; set the loss function as weighted mean square error, and the weight is set according to the influence degree of strain, temperature and displacement on engineering safety; select Adam algorithm as the optimizer, and apply learning rate decay strategy; perform iterative training until the performance of the model on the validation set converges;

[0118] The loss function is defined as:

[0119] (11),

[0120] Total Loss is expressed as a loss function; w Ɛ , w t , w D are weights given to strain, temperature, displacement variables according to engineering safety importance; MSE(s Ɛture , Ɛpred ), MSE(t true , t pred ), MSE(D true , D pred ) respectively represent the mean square error loss between the predicted value and the true value of strain, temperature and displacement; for example, if strain safety is the most critical, w Ɛ may be set slightly larger than other weights to guide the model to prioritize the accuracy of stress prediction in training. The optimizer uses Adam and cooperates with the learning rate decay strategy to stabilize the training process and improve the model generalization ability.

[0121] The formula of the mean square error is:

[0122] (12),

[0123] The formula of the mean absolute error is

[0124] (13),

[0125] The formula of the determination coefficient is

[0126] (14),

[0127] In the formula, n is the total number of samples, y i is the measured value of the i-th sample, is the mean value of the measured value, p i is the predicted value of the i-th sample by the model.

[0128] Step 7: Model performance evaluation and early warning: use the test set divided in step 4 to evaluate the performance of the deep learning model trained in step 6, use the root mean square error, the mean absolute error and the determination coefficient as the precision measurement standard of the model, the smaller the MSE and MAE value, the lower the prediction error; and R 2 tends to 1, indicating that the model fitting effect is good. On the contrary, if MSE and MAE are large, and R 2 deviates from 1, it indicates that the model precision is insufficient and the generalization ability is poor. According to the model prediction results and combining with the actual engineering, set the early warning threshold, when the predicted value exceeds the threshold, trigger the early warning, realize the active safety control of the support state.

[0129] The early warning mechanism includes a first yellow warning, a second orange warning and a third red warning;

[0130] First yellow warning: when the predicted displacement value exceeds 70% of the design allowable value or the predicted stress value exceeds 70% of the material allowable strain, a general warning is issued;

[0131] Second orange warning: when the predicted displacement value exceeds 80% of the design allowable value or the predicted stress value exceeds 80% of the material allowable strain, a high-risk warning is issued, prompting attention to a specific area;

[0132] Third red warning: when the predicted displacement value exceeds 90% of the design allowable value or the predicted stress value exceeds 90% of the material allowable strain, a high-risk warning is issued, and the dangerous rod number is accurately located, suggesting immediate intervention measures.

[0133] Step 7: Model performance evaluation and early warning: using the test set divided in step 4 to evaluate the performance of the deep learning model trained in step 6, using root mean square error, mean absolute error and determination coefficient as the precision measurement standard of the model, the smaller the MSE and MAE values, the lower the prediction error; and R 2 tends to 1, indicating that the model fitting effect is good. On the contrary, if MSE and MAE are large, and R 2 deviates from 1, indicating that the model precision is insufficient and the generalization ability is poor. According to the model prediction results and combined with the engineering practice, set the early warning threshold, when the predicted value exceeds the threshold, trigger the early warning.

[0134] An arch bridge support state prediction system implementing the above method, comprising:

[0135] A mechanism analysis module for establishing a fine finite element model, simulating the segmented pouring construction process, and identifying the most unfavorable position of the support;

[0136] A data acquisition and processing module for laying multiple source sensors, collecting strain, temperature and displacement data, performing data cleaning, alignment and normalization processing, and constructing a sliding window sample;

[0137] The attention-LSTM prediction model module includes an input layer, an LSTM layer, an attention mechanism layer, and a multi-output regression layer. The input layer is used to receive normalized multi-source monitoring data. The LSTM layer is used to learn the time dependence in the input sequence. The attention mechanism layer is used to automatically learn and assign importance weights of different historical time steps for the current prediction. The multi-output regression layer is used to transform and integrate the context vectors through a shared fully connected layer, and then send the integrated features to three independent output layers for stress, temperature, and displacement, respectively, to generate predictions of stress, temperature, and displacement at future M time steps.

[0138] The safety warning module is used to input real-time data, output prediction results, and trigger graded warnings according to threshold values.

[0139] The above-mentioned multi-source data and mechanical mechanism driven arch bridge support state prediction method and system are used to predict the arch bridge support state of the special animal channel bridge of the QL1 section of the West Land-Sea New Passage (Rail-Land) Canal Bridge Project. The main bridge of the special animal channel bridge is a reinforced concrete box-type arch bridge with a net span L0 = 125 m, a net rise f0 = 34 m, and a rise-span ratio f0 / L0 = 1 / 3.64. The main arch ring line type adopts a catenary, the arch axis coefficient m = 1.55, the arch ring section height is 230 cm, and the section is a single box three-chamber. The top and bottom plate thickness of the arch crown segment is 30 cm, the top and bottom plate thickness of the arch foot segment is 35 cm, and the web thickness is 30 cm. The arch column is a four-column column, with a transverse bridge width of 1.2 m and a longitudinal bridge width of 1.2 m. The arch deck plate is a 7x18.6m fabricated prestressed small box girder. The concrete main arch ring is cast in place by the support method, and the arch ring is poured in rings and segments. The support is composed of steel pipe columns, flat links, I-beams, and steel arches, etc. The single-span steel arch is divided into 11 segments and is symmetrically assembled from the arch foot to the arch crown.

[0140] 1. According to the design drawings and actual conditions, the main structure of the arch bridge support is spatially modeled. The arch bridge support calculation adopts the finite element method, and a large finite element software Midas is used to establish the finite element model of the main arch ring support system of the animal channel bridge. Figure 6 As shown in the figure. The modeling has a total of 20967 units, 11853 nodes, 3100 plate units, and 17867 beam units.

[0141] The whole process of segmental pouring of the main arch ring is simulated. The stress distribution and displacement of the support system under different construction conditions are analyzed. In the segmental pouring construction process of the arch ring, the stress and displacement response of the stand pole position of the numbered group A, B, C and D in the support model is the most significant, and the axial pressure far exceeds that of other rods, so the most unfavorable position of the arch bridge support during the whole construction period is shown in Figure 7 .

[0142] 2. In the arch bridge support stand pole group at the determined most unfavorable position, symmetrical vibration string strain gauges are respectively installed for synchronously collecting strain data and environmental temperature data at the position. The sampling frequency is set to 1 time / minute. Meanwhile, 360 prisms are installed at these positions, and the displacement data thereof are measured by a total station instrument to automatically measure the three-dimensional coordinates of the prisms at the same time interval as the strain gauges, so as to obtain the vertical displacement data of the support by calculating the height change, thereby forming a multi-source monitoring data stream composed of strain, temperature and displacement.

[0143] 3. Data preprocessing:

[0144] (1) Data cleaning: for the missing values possibly caused by data transmission and storage, linear interpolation or time series interpolation method is used for filling; for abnormal values obviously deviating from the normal range (such as spikes caused by instantaneous interference of equipment), four-quartile range method is used for identification and elimination. The steps of the four-quartile range method are as follows:

[0145] arrange the sequence X in ascending order, and find the first quartile Q1 (25% quartile) and the third quartile Q3 (75% quartile), and calculate the interquartile range IQR = Q3 - Q1;

[0146] determine the upper and lower limits of the normal value:

[0147] lower limit = Q1 - k * IQR,

[0148] upper limit = Q3 + k * IQR,

[0149] wherein k is a constant, and k = 1.5 (for identifying moderate abnormality) or k = 3.0 (for identifying extreme abnormality) is usually taken. Any data point lower than the lower limit or higher than the upper limit is determined as an abnormal value. The data points falling outside the range are determined as abnormal values and are eliminated, and then interpolation method is used for filling.

[0150] (2) Time alignment: the data streams of all sensors are resampled and aligned with a unified time reference (such as UTC time) to form a regular multi-dimensional time series data set {(Ɛ i ,T i ,D i )} corresponding to time stamp, strain, temperature and displacement, wherein i is the time point index.

[0151] (3) Data normalization: In order to eliminate the dimensional differences between each feature dimension, avoid the excessive influence of a certain feature on model training due to the large value range, and achieve effective comparison of each variable under the same standard, it is necessary to normalize the original sample data. Therefore, data normalization maps feature data of different dimensions to the same scale, i.e. [0, 1] or [-1, 1] interval, and the mathematical expression is:

[0152] (15),

[0153] In the formula: and .

[0154] The strain data of each vibrating wire strain gauge is recorded as: ,

[0155] The temperature data is recorded as: ,

[0156] The height difference data is recorded as: ,

[0157] The fused data is .

[0158] The vibrating wire strain gauge is from Hunan Sanzhiying Sensor Technology Co., Ltd., model SZZX-B150, gage length 150 mm, sensitivity 1 με, standard range ± 1500 με.

[0159] The total station observation prism is Leica Mini 360 prism. The total station is Leica TS60 automatic measurement robot

[0160] 4. The long time series data is converted into a sample set for supervised learning using the sliding window method. The specific parameter settings are: set the history window length N = 48 (time steps), the prediction window length M = 12 (time steps), and the sliding step length is 1 time step.

[0161] Input features (X): Each sample is composed of 48 consecutive hours of monitoring data, including stress, temperature, and displacement, forming a matrix with a dimension of (48, 3).

[0162] The sample set generated in time sequence is divided into 70% and 30% according to the proportion. Specifically, on the timeline, the first 70% of the samples are used for model training, and the last 30% of the samples are used for model testing.

[0163] A total of 1952 valid samples were generated by the sliding window method, of which 1366 samples (about 70% of the total samples) taken from the front of the timeline were used as the training set for internal training and verification of the model; 586 samples (about 30% of the total samples) taken from the back of the timeline were used as the verification set, which represents the "future" working conditions and is completely invisible during training, and is used exclusively to objectively and fairly evaluate the final prediction performance of the model.

[0164] 5. Multi-output prediction model construction based on attention-LSTM:

[0165] Input tensor: The model receives a tensor X with a shape of (batch size , N, 3). Where N is the historical time step (such as 48 hours), and 3 represents the features: stress, temperature, and displacement. The input sequence can be represented as: X = [x1, x2, …, x n ], where x t ∈ R 3 .

[0166] LSTM layer: for time series feature extraction. LSTM (Long Short-Term Memory Network) is a gating mechanism that introduces forget gate, input gate, and output gate, which can better extract time series dependencies and deep features in data, and is suitable for large-scale monitoring data analysis. It mainly includes forget gate, input gate, output gate, and memory cell.

[0167] Forget gate controls which information in the cell state is retained or discarded; memory cell regulates the forgetting and retention of information through the forget gate; input gate determines which new important information in the current input needs to be written into the cell state; output gate decides which information to output and the hidden state to pass on.

[0168] The forget gate determines the information that needs to be discarded by the previous memory unit:

[0169] (16),

[0170] The input gate controls the information that needs to be updated and generates candidate memory content:

[0171] (17),

[0172] (18),

[0173] The current memory cell state is updated to form a new memory:

[0174] (19),

[0175] The output gate determines the information that needs to be output:

[0176] (20),

[0177] (21),

[0178] In the formula: the subscripts i, f, and o are the input gate, the forget gate, and the output gate, respectively; is a nonlinear activation function; W is a weight matrix; and b is a bias matrix; is a hyperbolic tangent activation function; C t represents the updated memory cell state at the current time step; C t-1 represents the memory cell state at the previous time step, represents the candidate memory content at the current time step, which is new information generated from the current input; h t-1 and h t are the outputs of the LSTM unit at the previous time and the current time; x t is the network input at the current time.

[0179] The output of the second layer LSTM outputs the hidden states of all time steps, forming a sequence H = [h1, h2, …, h N ], where h t ∈ R d , and d is the dimension of the hidden state.

[0180] The attention mechanism layer is used to automatically learn and assign the importance weights of different historical moments to the current prediction, and the calculation steps are as follows:

[0181] (1) Calculate the attention score: for the last hidden state h N as the query), calculate the relevance score of each historical hidden state h t (as the key) in the sequence , where W α , U α are weight matrices, and v α is a weight vector.

[0182] (2) Normalize the weights: use the Softmax function to normalize the scores of all time steps into a probability distribution to obtain the attention weights a t :

[0183] (22),

[0184] (3) Generate the context vector: multiply all weights a t with the corresponding hidden states h t(As values) are weighted summed to obtain a context vector C, which is a weighted summary of the entire history sequence

[0185] (23),

[0186] (4) Multi-output regression layer: Multivariate synchronous prediction

[0187] Feature integration: The context vector C is transformed by a shared fully connected layer (with ReLU activation) to obtain the integrated feature C': (24),

[0188] Multi-task prediction: The integrated feature C' is fed into three independent output layers (with linear activation) to generate predictions for the next M time steps:

[0189] Stress prediction: (25),

[0190] Temperature prediction: (26),

[0191] Displacement prediction: (27),

[0192] where, , denote the learnable weight matrices for the stress, temperature, and displacement prediction branches, respectively; denote the learnable bias vectors for the stress, temperature, and displacement prediction branches, respectively;

[0193] The training set and validation set are input into the multi-output prediction model based on attention-LSTM.

[0194] The specific embodiment is implemented in Google Colab, GPU card: Telsa P100, video memory 16GB, operating system Ubuntu 18.04. The main hyperparameters include LSTM network and multi-attention mechanism. After testing, the hyperparameters are set as batch size 36, hidden layer size 64, dropout rate 0.2, learning rate 0.001, activation function relu, and optimizer Adam.

[0195] 6. Model training and optimization:

[0196] The model training uses weighted mean square error as the loss function, and its calculation formula is:

[0197] (28),

[0198] where, Total Loss is the loss function; w Ɛ , w t , w Dis the weight given to strain, temperature, displacement variable according to the importance of engineering safety; for example, if the strain safety is the most critical, w Ɛ is slightly larger than other weights, so as to guide the model to prioritize the accuracy of stress prediction in training. The optimizer uses Adam and cooperates with the learning rate decay strategy to stabilize the training process and improve the generalization ability of the model. MSE( Ɛture , Ɛpred ), MSE(t true , t pred ), MSE(D true , D pred ) respectively represent the mean square error loss between the predicted values of strain, temperature and displacement and the true values; the smaller the mean square error MSE value, the lower the prediction error. The determination coefficient (R 2 ) tends to 1, indicating that the model fitting effect is good.

[0199] The formula of the mean square error is:

[0200] (29),

[0201] The formula of the mean absolute error is:

[0202] (30),

[0203] The formula of the determination coefficient is:

[0204] (31),

[0205] In the formula, MSE is the mean square error; MAE is the mean absolute error; R 2 is the determination coefficient; n is the total number of samples; y i is the measured value of the ith sample; is the measured mean value; p i is the predicted value of the model for the ith sample.

[0206] In this embodiment, the training of the attention-LSTM multi-output prediction model converges after the 70th iteration, and the training, validation and test loss curves of the attention-LSTM multi-output prediction model are as shown in Figure 8 Figure 8 The mean square error of the model prediction is 0.356, and the R 2 ​0.984, it can be seen that the method of the embodiment has high accuracy in predicting the support construction state. In addition, there is no serious overfitting phenomenon in the training process of the model as a whole. The support construction state intelligent monitoring method based on the attention-LSTM multi-output prediction model provided by the embodiment uses attention-LSTM to train the input multi-source fusion data, achieves prediction of the support construction state, and can also set an early warning threshold, and triggers an early warning when the predicted value exceeds the threshold.

[0207] Primary warning (yellow): when the predicted displacement value exceeds 70% of the design allowable value or the predicted stress value exceeds 70% of the material allowable strain, a general warning is issued;

[0208] Secondary warning (orange): when the predicted displacement value exceeds 80% of the design allowable value or the predicted stress value exceeds 80% of the material allowable strain, a high-risk warning is issued, and attention is drawn to a specific area;

[0209] Tertiary warning (red): when the predicted displacement value exceeds 90% of the design allowable value or the predicted stress value exceeds 90% of the material allowable strain, a high-risk warning is issued, and the dangerous rod number is accurately located, and immediate intervention measures are recommended. The active safety control of the support state is realized. It provides intelligent protection for construction safety and has certain engineering significance.

Claims

1. A multi-source data and mechanics mechanism driven arch bridge support state prediction method, characterized in that, Comprising the following steps: Step 1, mechanical mechanism analysis and monitoring point positioning: according to the design drawing and the actual situation, a fine finite element model of the tied arch bridge support construction is established, the whole process of the main arch ring segmental pouring is simulated, the mechanical response of the support under different construction conditions is analyzed, and the most unfavorable position of the support in the whole construction period is determined; Step 2, multi-source monitoring data acquisition: a plurality of sensors including a vibrating wire strain gauge and a 360 prism are arranged on the most unfavorable position determined in step 1 to build a multi-source monitoring system, strain data and temperature data are synchronously collected by the vibrating wire strain gauge, displacement data are measured by the total station and the 360 prism, vertical displacement data of the support are obtained by calculating the elevation at each time, and the collected strain, temperature and displacement data are integrated to form a multi-source monitoring data stream; Step 3, data preprocessing: the original monitoring data stream of the multi-source monitoring data in step 2 is subjected to data cleaning and time alignment to form a regular multi-dimensional time series data set, and the regular multi-dimensional time series data set is subjected to normalization processing to obtain a normalized multi-source monitoring data set, the multi-source monitoring data set including the normalized strain, temperature and displacement data; Step 4, supervised learning data set construction: the normalized multi-source monitoring data in step 3 is processed by a sliding window method to generate a sample set for supervised learning, in the sliding window method, the strain, temperature and displacement data of the past N time steps are set as input features, wherein each time step includes the normalized strain, temperature and displacement, the strain, temperature and displacement variables of the future M time steps are set as prediction targets to constitute prediction target variables, the sliding window is slid on the entire normalized multi-source data set, one time step is slid each time to generate a series of sample pairs, all generated sample pairs are combined to form a sample set, the sample set is subjected to stratified sampling to divide the sample set into a training set and a test set at a ratio of 70% and 30%, the training set is used for training of the attention-LSTM multi-output prediction model, and the test set is used for evaluating the model performance; Step 5, constructing an attention-LSTM multi-output prediction model: using the training set divided in step 4, an attention-LSTM multi-output prediction model including an input layer, two layers of stacked LSTM layers, an attention mechanism layer and a multi-output regression layer is constructed; Step 6: model training and optimization: the attention-LSTM multi-output prediction model constructed in step 5 and the training set divided in step 4 are used for training; the loss function is set as weighted mean square error, the weights are set according to the influence degree of strain, temperature and displacement on engineering safety; the Adam algorithm is selected as the optimizer, and the learning rate decay strategy is applied; iterative training is performed until the performance of the model on the validation set converges; Step 7: Model performance evaluation and early warning: using the test set divided in step 4 to evaluate the performance of the deep learning model trained in step 6, using root mean square error, mean absolute error and determination coefficient as the precision measurement standard of the model, setting early warning threshold according to the model prediction result and combining engineering practice, triggering early warning when the prediction value exceeds the threshold.

2. The method of claim 1, wherein, The formula for calculating the elevation change in step 2 to obtain the vertical displacement data of the support is as follows: ， In the formula, represents the vertical displacement of the support at a certain time; is the initial reference period prism center point elevation, is the t period prism center point elevation.

3. The method of claim 1, wherein, The data cleaning in step 3 is for missing values caused by transmission and storage of original data in multi-source monitoring data stream, which is filled by linear interpolation or time series interpolation method; For abnormal values obviously deviating from the normal range, identify and eliminate them by quartile range method; The time alignment is to resample and align all sensor multi-source monitoring data streams with a unified time reference to form a regular multi-dimensional time series data set; The data normalization processing adopts Min-Max normalization method to map the value of each feature variable to the interval [0, 1], and the expression of data normalization processing is as follows: ， In the formula: and ; represents the feature value in the original sample data, which belongs to the vector in the real number field ; represents the minimum value of the feature in all samples, that is ; represents the maximum value of the feature in all samples, that is ; represents the normalized feature value, which belongs to the vector in the real number field , and its value range is [0, 1] or [-1, 1].

4. The method of claim 1, wherein, The input layer in step 5 is used to receive a tensor of shape (batch size , N, 3), where N is the history time steps and 3 represents three features of strain, temperature and displacement. The LSTM layer adopts two layers of stacked LSTM network to learn the complex time dependence in the input sequence deeply and efficiently; The attention mechanism layer is used to automatically learn and assign the importance weight of different historical time to the current prediction, and the calculation steps of the attention mechanism layer are as follows: Step 51, compute attention score: for the last hidden state h N As a query, compute hidden state h N With each history hidden state h t As a key, compute relevance score, formula as follows: ， where e t denotes the attention score at the t-th time step; W α , U α are learnable weight matrices; v α is a learnable weight vector; T denotes the transpose operation of a vector or a matrix; i.e., denotes the transpose of the weight vector v α ; tanh(·) is the hyperbolic tangent activation function; h t denotes the history hidden state at the t-th time step; h N denotes the hidden state at the last time step; Step 52, normalize to weights: normalize the scores of all time steps to a probability distribution using the Softmax function to get attention weights a t ; ， In the formula, This represents the attention weight corresponding to the t-th time step; N represents the total length of the input sequence, i.e., the total number of time steps. This represents the attention score at time step t. The result of the exponentiation operation; exp(e j ) represents the attention score at time step j. The result of the exponential operation; This represents the summation of attention scores over all time steps after performing an exponential operation. Step 53, generating context vector: weight all weights a t with the corresponding hidden state h t Weighted sum as values, get a context vector C, context vector C formula as follows: ， where N denotes the total length of the input sequence, i.e., the total number of time steps; a t denotes the attention weight of the t-th time step; h t denotes the history hidden state of the t-th time step; The multi-output regression layer is used to transform and integrate the context vector through a shared fully connected layer, and the integrated features are sent to three independent output layers to generate predictions of stress, temperature and displacement in the next M time steps; The formula of feature transformation and integration is as follows: ， In the formula, C' represents the new features output after feature transformation and integration; W S denotes the learnable weight matrix corresponding to the shared fully connected layer; C denotes the context vector; b s denotes the learnable bias vector corresponding to the shared fully connected layer; ReLU(·) denotes a linear rectifier activation function defined as ReLU(x) = max(0, x); the prediction of the stress: , Prediction of the temperature: , The prediction of the displacement: , In the formula, , respectively represent the learnable weight matrix corresponding to the stress, temperature and displacement prediction branch; respectively represent the learnable bias vector corresponding to the stress, temperature and displacement prediction branch.

5. The method of claim 1, wherein, The loss function in step 6 is defined as: ， where Total Loss is expressed as a loss function; w Ɛ , t , D are weights assigned to strain, temperature, displacement variables according to engineering safety importance; MSE(s Ɛture , Ɛpred ), MSE(t true , t pred ), MSE(D true , D pred ) represent mean square error loss between predicted and true values for strain, temperature, and displacement, respectively. The formula of mean square error is: ， The formula of mean absolute error is: ， The formula of determination coefficient is: ， where MSE is the mean square error; MAE is the mean absolute error; R 2 represents the determination coefficient; n is the total number of samples; y i is the measured value of the ith sample; is the measured mean value; p i is the predicted value of the ith sample by the model.

6. The method of claim 1, wherein, The early warning mechanism in step 7 includes first-level yellow warning, second-level orange warning and third-level red warning; First-level yellow warning: when the predicted displacement value exceeds 70% of the design allowable value or the predicted stress value exceeds 70% of the material allowable strain, a general warning is issued; Second-level orange warning: when the predicted displacement value exceeds 80% of the design allowable value or the predicted stress value exceeds 80% of the material allowable strain, a high-risk warning is issued, prompting attention to a specific area; Third-level red warning: when the predicted displacement value exceeds 90% of the design allowable value or the predicted stress value exceeds 90% of the material allowable strain, a high-risk warning is issued, and the dangerous rod number is accurately located, suggesting immediate intervention measures.

7. An arch bridge support state prediction system that implements the method according to any one of claims 1 to 6, characterized by, It includes: Mechanism analysis module for establishing a fine finite element model to simulate the segmented pouring construction process and identify the most unfavorable position of the support; Data acquisition and processing module for laying multi-source sensors, collecting strain, temperature and displacement data, performing data cleaning, alignment and normalization processing, and constructing a sliding window sample; The attention-LSTM prediction model module comprises an input layer, an LSTM layer, an attention mechanism layer and a multi-output regression layer, the input layer is used for receiving normalized multi-source monitoring data, the LSTM layer is used for learning time dependence in an input sequence; The attention mechanism layer is used for automatically learning and assigning importance weights of different historical moments to current prediction, and the multi-output regression layer is used for transforming and integrating features of a context vector through a shared fully connected layer, and sending the integrated features to three independent output layers of strain, temperature and displacement respectively to generate stress, temperature and displacement predictions of M future time steps for generating strain, temperature and displacement predictions of future time steps; The safety warning module is used for inputting real-time data, outputting prediction results and triggering graded warnings according to thresholds.

Images