Railway bridge deflection intelligent prediction method based on multi-feature fusion and dynamic feedback

Abstract

The invention discloses a railway bridge deflection intelligent prediction method based on multi-feature fusion and dynamic feedback, and the method comprises the following steps: obtaining multi-dimensional time series data, and carrying out the preprocessing; data is divided into a training set, a verification set and a test set during preprocessing; constructing a deep learning model; training the model by using the training set to obtain a trained model; and performing rolling prediction on the test set by using the trained model. According to the method, unified modeling is carried out on 12 key environment physical field variables influencing the deflection of the bridge and the deflection, a complete environment excitation-structure response mapping relation is constructed, and the limitation that a traditional method is single in consideration factor is overcome. According to the method, the model has a solid physical basis and is excellent in generalization ability. According to the invention, key and timely data support is provided for safe operation management, preventive maintenance and speed limit decision of the high-speed railway bridge under extreme weather or abnormal load.

Description

Technical Field

[0001] This invention belongs to the field of railway bridge monitoring technology, specifically involving an intelligent prediction method for railway bridge deflection based on multi-feature fusion and dynamic feedback. Background Technology

[0002] According to incomplete statistics, my country currently has over 8,000 continuous beam bridges for high-speed railways and over 200 long-span bridges exceeding 200 meters. Due to the extremely complex temperature deformation of civil engineering structures and numerous influencing factors, current predictions for the deflection and temperature deformation of railway bridges only include the railway itself, resulting in low short-term temperature deformation forecast accuracy. Each forecast process is extremely complex and requires comprehensive analysis. Furthermore, the increasing frequency of extreme weather phenomena, previously rare, significantly increases the difficulty of forecasting. Summary of the Invention

[0003] This invention is proposed to address the problems existing in the prior art, and its purpose is to provide an intelligent prediction method for railway bridge deflection based on multi-feature fusion and dynamic feedback.

[0004] The technical solution of this invention is: an intelligent prediction method for railway bridge deflection based on multi-feature fusion and dynamic feedback, comprising the following steps: A. Acquire multidimensional time series data and perform preprocessing; B. Construct a CNN deep learning model; C. Train the model using the training set to obtain a trained model; D. Use the trained model to make rolling predictions on the test set.

[0005] Furthermore, step A involves obtaining multidimensional time series data, the specific process of which is as follows: First, the bridge deflection is measured using a hydrostatic level to obtain the bridge deflection. Then, we obtain environmental data about the area surrounding the bridge.

[0006] Furthermore, the environmental data surrounding the bridge is as follows: First, environmental data around the bridge was obtained using nearby weather stations; Then, the environmental data around the bridge include surface pressure, air temperature, surface temperature, precipitation, relative humidity, evaporation, potential evaporation, north-south wind speed component, east-west wind speed component, surface solar downward radiation, surface net radiation, and direct radiation across the entire sky.

[0007] Furthermore, step A involves preprocessing the data, and the preprocessed data is then packaged into a batch data loader.

[0008] Furthermore, step D uses the trained model to perform rolling predictions on the test set, as detailed below: First, load the pre-trained prediction model; Then, the test dataset is divided into multiple consecutive subsequences; Finally, the subsequences are predicted one by one to complete the rolling prediction.

[0009] Furthermore, the subsequence contains consecutive samples of a preset length.

[0010] Furthermore, in the rolling prediction process, the first sample in the subsequence is first input into the prediction model to obtain the first predicted value.

[0011] Furthermore, during the rolling prediction process, the first predicted value is used as part of the input features for the next sample to construct a new sample.

[0012] Furthermore, new samples are constructed as input and fed into the prediction model to obtain a second predicted value.

[0013] Furthermore, the second predicted value is also used as part of the input features for the next sample, thus forming a rolling prediction.

[0014] The beneficial effects of this invention are as follows: This invention integrates the 12 key environmental physical field variables that affect bridge deflection with the deflection itself into a unified model, constructing a complete environmental excitation-structural response mapping relationship, overcoming the limitation of traditional methods that consider only one factor.

[0015] The features selected in this invention are directly targeted at the actual situation of bridge engineering, covering key environmental effects such as temperature gradient, solar radiation thermal effect, and wind and rain coupling, so that the model has a solid physical foundation and excellent generalization ability.

[0016] This invention enables high-precision deflection prediction up to 24 hours in advance, providing crucial and timely data support for the safe operation management, preventive maintenance, and speed limit decisions of high-speed railway bridges under extreme weather or abnormal loads. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This refers to the hourly data of bridge deflection collected by this invention; Figure 3 This is hourly surface air pressure data from the meteorological station of this invention; Figure 4 The data represents hourly air temperature data from the weather station of this invention. Figure 5 This invention provides hourly precipitation data from the meteorological station. Figure 6The data represents the hourly relative humidity data from the weather station of this invention. Figure 7 The data represents hourly surface temperature data from the meteorological station of this invention. Figure 8 This is hourly evaporation data from the weather station of this invention; Figure 9 This invention provides potential evaporation data from the meteorological station. Figure 10 The north-south wind speed of the meteorological station of this invention; Figure 11 The east-west wind speed of the weather station of this invention; Figure 12 Solar radiation on the surface of the weather station of this invention (downward direction); Figure 13 The surface solar radiation (net) of the weather station of this invention; Figure 14 The meteorological station of this invention provides direct radiation across the entire sky. Figure 15 This invention provides a prediction of bridge deflection for the next hour. Figure 16 This invention provides a prediction of bridge deflection for the next 24 hours. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and embodiments: like Figures 1 to 16 As shown, the intelligent prediction method for railway bridge deflection based on multi-feature fusion and dynamic feedback includes the following steps: A. Acquire multidimensional time series data and perform preprocessing; B. Constructing deep learning models; C. Train the model using the training set to obtain a trained model; D. Use the trained model to make rolling predictions on the test set.

[0019] Step A involves obtaining multidimensional time series data, and the specific process is as follows: First, the bridge deflection is measured using a hydrostatic level to obtain the bridge deflection. Then, we obtain environmental data about the area surrounding the bridge.

[0020] The environmental data surrounding the bridge is as follows: First, environmental data around the bridge was obtained using nearby weather stations; Then, the environmental data around the bridge include surface pressure, air temperature, surface temperature, precipitation, relative humidity, evaporation, potential evaporation, north-south wind speed component, east-west wind speed component, surface solar downward radiation, surface net radiation, and direct radiation across the entire sky.

[0021] Step A involves preprocessing the data, and the preprocessed data is then encapsulated into a batch data loader.

[0022] Specifically, in step B, the deep learning model uses a one-dimensional convolutional neural network (1D-CNN) as the core feature extractor. The deep learning model contains several cascaded one-dimensional convolutional layers (Conv1D) and pooling layers (MaxPooling1D).

[0023] In the aforementioned cascaded one-dimensional convolutional layers, the first convolutional layer uses multiple convolutional kernels (e.g., 64), and subsequent convolutional layers use stacked, deeper convolutional layers (e.g., 32 or 16), sliding along the time dimension to extract local correlation patterns and trends of different features within a short time window.

[0024] The pooling layer is used to gradually reduce the size of the time dimension, thereby reducing computational complexity and enhancing the model's robustness to small changes in time position.

[0025] The deep learning model converts the high-dimensional feature map output by the last convolutional / pooling layer into a one-dimensional feature vector through a flattening layer; then, it connects one or more fully connected layers (dense layers).

[0026] Before training, the deep learning model defines the model's loss function (such as mean squared error MSE, suitable for regression prediction tasks) and optimizer (such as Adam optimizer), and optionally sets the evaluation metric (such as mean absolute error MAE).

[0027] Step C involves training the model using the training set to obtain a trained model. The specific process is as follows: First, the multidimensional time series dataset, which has been preprocessed and packaged in step A, is divided into training set, validation set and test set in chronological order. Next, set the training parameters, including batch size and number of training epochs. Initialize the CNN model built in step B; Then, iterative training and validation are performed; Then, when the validation loss no longer decreases significantly over several consecutive epochs, training is terminated early to prevent overfitting. Throughout the training process, snapshots of the optimal model parameters on the validation set are continuously saved as the final "trained model". Finally, after training, the training and validation loss curves are analyzed to ensure that the model learning process is stable and effective. The final saved model will be used for rolling prediction in step D.

[0028] Specifically, during the iterative training and validation process, a batch of training data is first input into the model, and the predicted deflection value is obtained through network calculation.

[0029] Specifically, during the iterative training and validation process, a preset loss function value is calculated based on the model output and the actual deflection label.

[0030] Specifically, during the iterative training and validation process, the backpropagation algorithm is executed through the optimizer to calculate the gradient of the loss with respect to the parameters of each layer of the model, and the model weights are updated to minimize the loss.

[0031] Specifically, during the iterative training and validation process, after each training epoch, the validation set is used to evaluate the current model performance, calculate the validation loss, and set metrics. This step is used to monitor whether the model is overfitting or underfitting and to determine the convergence status.

[0032] Step D uses the trained model to perform rolling predictions on the test set. The specific process is as follows: First, load the pre-trained prediction model; Then, the test dataset is divided into multiple consecutive subsequences; Finally, the subsequences are predicted one by one to complete the rolling prediction.

[0033] The subsequence contains consecutive samples of a preset length.

[0034] In the rolling prediction process, the first sample in the subsequence is first input into the prediction model to obtain the first predicted value.

[0035] During the rolling prediction process, the first predicted value is used as part of the input features for the next sample to construct a new sample.

[0036] The new sample is constructed as input and fed into the prediction model to obtain a second prediction value.

[0037] The second predicted value is also used as part of the input features for the next sample, thus forming a rolling prediction.

[0038] Specifically, after completing the rolling prediction, the prediction results of all subsequences are collected to obtain the final multi-step prediction sequence, such as... Figure 15 and Figure 16 As shown.

[0039] Specifically, the predicted sequence is compared with the actual value sequence, and evaluation indicators are calculated and visualized.

[0040] Specifically, the deep learning model mentioned in step B is CNN.

[0041] Specifically, preprocessing is performed in step A, and the specific process is as follows: First, load the time series dataset; More specifically, the dataset contains a target variable and multiple related feature variables; Then, the dataset is divided into training set, validation set and test set according to the preset ratio; More specifically, the preset ratio is 0.7:0.2:0.1; Next, normalized parameters are calculated based on the training set; Then, the sliding window method is used to combine the historical values ​​of the target variable and related feature values ​​within each time window into an input sequence; Next, the input sequence and the corresponding target variable labels for the next time step are converted into tensor formats that can be processed by the deep learning model; Finally, the processed data is packaged into a batch data loader.

[0042] This invention integrates the 12 key environmental physical field variables that affect bridge deflection with the deflection itself into a unified model, constructing a complete environmental excitation-structural response mapping relationship, overcoming the limitation of traditional methods that consider only one factor.

[0043] The features selected in this invention are directly targeted at the actual situation of bridge engineering, covering key environmental effects such as temperature gradient, solar radiation thermal effect, and wind and rain coupling, so that the model has a solid physical foundation and excellent generalization ability.

[0044] This invention enables high-precision deflection prediction up to 24 hours in advance, providing crucial and timely data support for the safe operation management, preventive maintenance, and speed limit decisions of high-speed railway bridges under extreme weather or abnormal loads.

Claims

1. A method for intelligent prediction of railway bridge deflection based on multi-feature fusion and dynamic feedback, characterized in that: Includes the following steps: A. Acquire multidimensional time series data and perform preprocessing; B. Constructing deep learning models; C. Train the model using the training set to obtain a trained model; D. Use the trained model to make rolling predictions on the test set.

2. The intelligent prediction method for railway bridge deflection based on multi-feature fusion and dynamic feedback as described in claim 1, characterized in that: Step A involves obtaining multidimensional time series data, and the specific process is as follows: First, the bridge deflection is measured using a hydrostatic level to obtain the bridge deflection. Then, we obtain environmental data about the area surrounding the bridge.

3. The intelligent prediction method for railway bridge deflection based on multi-feature fusion and dynamic feedback according to claim 2, characterized in that: The environmental data surrounding the bridge is as follows: First, environmental data around the bridge was obtained using nearby weather stations; Then, the environmental data around the bridge include surface pressure, air temperature, surface temperature, precipitation, relative humidity, evaporation, potential evaporation, north-south wind speed component, east-west wind speed component, surface solar downward radiation, surface net radiation, and direct radiation from the entire sky.

4. The intelligent prediction method for railway bridge deflection based on multi-feature fusion and dynamic feedback according to claim 1, characterized in that: Step A involves preprocessing the data, and the preprocessed data is then encapsulated into a batch data loader.

5. The intelligent prediction method for railway bridge deflection based on multi-feature fusion and dynamic feedback according to claim 1, characterized in that: Step D uses the trained model to perform rolling predictions on the test set. The specific process is as follows: First, load the pre-trained prediction model; Then, the test dataset is divided into multiple consecutive subsequences; Finally, the subsequences are predicted one by one to complete the rolling prediction.

6. The intelligent prediction method for railway bridge deflection based on multi-feature fusion and dynamic feedback according to claim 5, characterized in that: The subsequence contains consecutive samples of a preset length.

7. The intelligent prediction method for railway bridge deflection based on multi-feature fusion and dynamic feedback according to claim 6, characterized in that: In the rolling prediction process, the first sample in the subsequence is first input into the prediction model to obtain the first predicted value.

8. The intelligent prediction method for railway bridge deflection based on multi-feature fusion and dynamic feedback according to claim 7, characterized in that: During the rolling prediction process, the first predicted value is used as part of the input features for the next sample to construct a new sample.

9. The intelligent prediction method for railway bridge deflection based on multi-feature fusion and dynamic feedback as described in claim 8, characterized in that: The new sample is constructed as input and fed into the prediction model to obtain a second prediction value.

10. The intelligent prediction method for railway bridge deflection based on multi-feature fusion and dynamic feedback according to claim 9, characterized in that: The second predicted value is also used as part of the input features for the next sample, thus forming a rolling prediction.

Images