Abnormal deflection monitoring method of cable-stayed bridge main beam based on deep learning model
A deep learning model with a convolutional and LSTM neural network architecture enhances the extraction of local and temporal information to accurately identify abnormal deflections in cable-stayed bridges, addressing interpretability and parameter debugging issues in traditional methods.
Patent Information
- Application Number
- CN202111185481.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-12
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2041-10-12
AI Technical Summary
Traditional regression analysis and mathematical modeling struggle to accurately express the complex, nonlinear relationship between temperature and temperature-induced deflection in cable-stayed bridges, leading to difficulties in identifying abnormal deflections due to poor model interpretability and parameter debugging in deep learning neural networks.
A deep learning model is developed using a convolutional neural network layer followed by a stacked long short-term memory (LSTM) neural network layer to process temperature data, enhancing local and temporal information extraction, and a linear activation function to obtain a normalized regression value, which is then denormalized for precise deflection identification.
The model achieves highly accurate and stable deflection regression values, enabling direct quantitative identification of abnormal deflections with a residual-based anomaly indicator, ensuring reliable bridge status assessment.
Smart Images

Figure CN114037002B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of bridge structure performance monitoring and is a model for identifying abnormal deflection values of a cable-stayed bridge main beam based on bridge monitoring data. Specifically, it relates to a modeling of a reference value of the deflection of a cable-stayed bridge main beam based on time series modeling and deep learning, and uses the reference value to directly identify abnormal deflection values of the main beam. Background Art
[0002] The deflection of the main beam of a cable-stayed bridge is the most intuitive expression of the stiffness performance of the cable-stayed bridge. Affected by the complex temperature field acting on the cable-stayed bridge, the main beam of the cable-stayed bridge will produce significant deflection changes, that is, the temperature-induced deflection of the cable-stayed bridge. Therefore, a mapping model is established for the input-output relationship between the temperature field of the cable-stayed bridge and the temperature-induced deflection of the main beam. The temperature data can be brought into the model and the regression value of the temperature-induced deflection can be output. This regression value can be used as a reference value to characterize the linear shape of the cable-stayed bridge, and then the abnormal deflection can be identified based on this reference value, so as to judge the state of the bridge, perform bridge maintenance and prevent accidents. However, the temperature field acting on the cable-stayed bridge is extremely complex, and the relationship between temperature and the temperature-induced deflection of the cable-stayed bridge shows a strong nonlinear and fuzzy relationship. Traditional regression analysis and mathematical modeling cannot accurately express the mapping relationship between the two, which makes it very difficult to directly identify the abnormal deflection of the main beam.
[0003] The birth of deep learning technology has provided tools for modeling highly nonlinear problems. Based on the advancement of computer technology, big data on bridge monitoring has been accumulated, providing a data-driven foundation for the mature application of deep neural networks. However, the mapping model built on neural networks is black-boxed and lacks interpretability, so it is difficult to ensure the reliability of the model. At the same time, the simple data-driven model makes it difficult to set model parameters, which further makes it difficult to build a high-precision neural network.
[0004] Relying on bridge monitoring big data, deep learning is undoubtedly the best solution to establish a highly nonlinear relationship between the temperature and temperature-induced deflection of cable-stayed bridges, and has the potential to provide high-precision deflection benchmark values. However, there is no mature precedent for applying deep learning technology to the identification of abnormal deflections of cable-stayed bridges. Although deep neural networks have excellent performance, as a data-driven black box model with complex hyperparameters and numerous neural units, it is difficult to build a deep model of temperature-induced deflection with high precision and high stability. Therefore, in order to achieve high-precision identification of abnormal deflection of the main beam of a cable-stayed bridge, it is necessary to create a temperature→temperature-induced deflection mapping model with high precision and high output stability for the temperature-induced deflection modeling task of cable-stayed bridges, and then appropriately process the output data in order to directly identify abnormal deflections. Therefore, two problems must be solved:
[0005] (1) Deep learning neural networks that rely solely on data-driven learning are typical black box models. Parameter debugging is cumbersome, it is difficult to achieve ideal accuracy, and the model has poor interpretability and reliability.
[0006] (3) Complex abnormal deflection indicators make it impossible to directly identify abnormal values. Summary of the invention
[0007] The purpose of the present invention is to use bridge monitoring data to identify the service status of the main beam of a cable-stayed bridge, and to provide a deep learning model for monitoring abnormal deflection of the main beam of a cable-stayed bridge. Specifically, it relates to a temperature-induced deflection modeling scheme for cable-stayed bridges based on a deep learning neural network and quantitative identification of deflection anomalies.
[0008] To achieve the above object, the technical solution adopted by the present invention is: a deep learning model for monitoring abnormal deflection of the main beam of a cable-stayed bridge is proposed, and the deflection reference value is obtained by using the model to complete abnormal deflection identification, including the following steps:
[0009] (1) The temperature field data collected by the cable-stayed bridge monitoring system is converted into time series data of n temperature variables, such as the average temperature of the main beam, the vertical temperature difference of the main beam, and the temperature of the cable tower. Based on the time-frequency characteristics of temperature, the time series data of the temperature-induced deflection in the change of the main beam deflection is extracted;
[0010] (2) The input-output mapping mode of temperature → temperature-induced deflection is set to input temperature time history data with i data points and output a single temperature-induced deflection data point at the current moment, thereby solidifying the input data specification;
[0011] (3) Normalize the input and output data in (2);
[0012] (4) Build a black-box deep neural network mapping model, add a convolutional neural network layer with n convolution kernels after the input layer, strengthen the local information of the temperature data, and pass the output matrix of the convolutional neural network layer to the next layer of stacked long short-term memory (LSTM) neural network layer. The output of the stacked long short-term memory neural network layer is still a matrix with two dimensions of time and feature. Then set up a convolutional neural network layer with the same size as the stacked long short-term memory neural network layer, integrate the two-dimensional output results of the stacked long short-term memory neural network layer into a single value, and then pass it through a linear activation function to obtain a normalized regression value;
[0013] (5) Denormalize the normalized regression value to obtain the temperature-induced deflection regression value;
[0014] (6) Prepare sufficient data sets to train deep neural networks to achieve excellent practical performance;
[0015] (7) The residual between the temperature-induced deflection regression value output by the mapping model and the actual temperature-induced deflection measured by the bridge monitoring system is used as the value of the abnormal deflection to complete the abnormal deflection monitoring of the main beam of the cable-stayed bridge.
[0016] Preferably, the step (2) comprises:
[0017] The input data of the model is an n×i data matrix composed of the time series data of all temperature variables, and the model output is the single-point temperature-induced deflection at the current moment.
[0018] Preferably, the step (4) comprises:
[0019] The first core layer is a convolutional neural network layer with n convolution kernels, each of which is n×m in size. Each convolution kernel traverses the input data matrix once along the time dimension, transforming the n×i temperature data matrix into an n×(i-m+1) matrix to enhance local information. The second core layer is a stacked long short-term memory neural network layer with two layers of LSTM hidden units, which is used to learn the temporal correlation characteristics of temperature data and inherit the matrix of size n×(i-m+1) output from the upper layer. Each layer contains p LSTM units, so the information output by this core layer is a p×(i-m+1) matrix. The third core layer is a p×(i-m+1) convolution kernel. After one convolution operation, the matrix of size p×(i-m+1) output from the upper layer is integrated into a single value through convolution operation, that is, the regression value of temperature-induced deflection.
[0020] Beneficial effects:
[0021] Compared with the prior art, the present invention has the following beneficial effects:
[0022] (1) De-blackboxing: The deep learning neural network model is a data-driven black box model with difficult parameter adjustment. At the same time, due to the poor interpretability of the principle, it is difficult to ensure the reliability of the model. The present invention determines the construction paradigm of the input data, which is interpretable. The clear data planning optimizes the construction target of the neural network, improves the model fitting performance, and realizes the de-blackboxing of deep neural network modeling.
[0023] (2) Excellent and stable output accuracy: The present invention adds a convolutional neural network layer with the same number of convolution kernels as the number of variables after the output layer of the neural network, which strengthens the extraction of local temperature information. The stacked long short-term memory unit layer has a strong temporal feature expression. A convolution kernel is set before the output layer, which can integrate dual-dimensional information. Therefore, the deflection regression value output by the mapping model constructed by the present invention is highly accurate and stable;
[0024] (3) Direct identification of outliers: Thanks to the high-precision and high-stability mapping model, the residual between the deflection regression value and the measured value can be directly used as the deflection anomaly indicator, which can achieve direct quantitative identification of abnormal deflection. Therefore, the present invention has the potential to clearly judge the status of the bridge. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1It is an algorithm flow chart of the method of the present invention;
[0026] Figure 2 A neural network structure for modeling temperature-induced deflection mapping of cable-stayed bridges proposed by the method of the present invention;
[0027] Figure 3 The data flow of data in the deep neural network in the method of the present invention;
[0028] Figure 4 Schematic diagram of the deflection anomaly value and allowable error obtained according to the method of the present invention. DETAILED DESCRIPTION
[0029] The following will refer to Figure 1 to Figure 4 , the technical solution of the present invention is described in detail.
[0030] The present invention is a method for monitoring abnormal deflection of main beam of cable-stayed bridge based on deep learning model. Figure 1 As shown, the method specifically comprises the following steps:
[0031] (1) The temperature field data collected by the cable-stayed bridge monitoring system is converted into time series data of n temperature variables, such as the average temperature of the main beam, the vertical temperature difference of the main beam, and the temperature of the cable tower. Based on the time-frequency characteristics of temperature, the time series data of the temperature-induced deflection in the change of the main beam deflection is extracted;
[0032] (2) The input-output mapping mode of temperature → temperature-induced deflection is set to input temperature time history data with i data points and output the temperature-induced deflection data point at the current moment, thereby solidifying the input data specification;
[0033] (3) Normalize the data in (2);
[0034] (4) Build a mechanism-data dual-driven black-box deep neural network mapping model. Add a convolutional neural network layer with n convolution kernels after the input layer, substitute the output matrix of this layer into the next layer of stacked long short-term memory (LSTM) neural network, and then bring the output matrix of the stacked long short-term memory network into a convolutional neural network layer to complete the integration of the matrix into a single value, and then pass it through a linear activation function to obtain a normalized regression value;
[0035] (5) Denormalize the data to obtain the regression value of temperature-induced deflection;
[0036] (6) Prepare sufficient data sets to train the black-box deep neural network mapping model so that it has excellent practical performance;
[0037] (7) The value of abnormal deflection is directly identified by using the residual between the temperature-induced deflection regression value output by the mapping model and the temperature-induced deflection measured by the bridge monitoring system;
[0038] Preferably, the step (2) comprises:
[0039] The input data of the model is an n×i data matrix composed of the time series data of all temperature variables, and the model output is the single-point temperature-induced deflection at the current moment.
[0040] Preferably, the step (4) comprises:
[0041] The first core layer is a convolutional neural network layer with three convolution kernels. The size of each convolution kernel is n×m. Each convolution kernel traverses the input data matrix once along the time dimension, transforms the n×i temperature data matrix into an n×(i-m+1) matrix, and completes local information enhancement. The second core layer is a stacked long short-term memory neural network layer with two layers of LSTM hidden units, which is used to learn the temporal correlation characteristics of temperature data and inherit the matrix of size n×(i-m+1) output from the upper layer. Each core layer contains p LSTM units, so the information output by this layer is a p×(i-m+1) matrix. The third core layer is a p×(i-m+1) convolution kernel. After a convolution operation, the matrix of size p×(i-m+1) output from the upper layer is integrated into a single value through convolution operation, which is the regression value of temperature-induced deflection.
[0042] Example:
[0043] Based on the monitoring data of the temperature field and the mid-span deflection of the main span of the Tongling highway-railway cable-stayed bridge in Anhui Province over nine months, the specific implementation process of the present invention is described.
[0044] (1) The temperature field data of the Tongling Yangtze River Bridge were extracted into three temperature characteristics that affect the deflection of the cable-stayed bridge, namely the average temperature of the main beam, the vertical temperature difference of the main beam and the tower temperature, thereby obtaining the time series data of the three temperature variables; using the wavelet multi-scale decomposition method, the temperature-induced deflection part of the main beam deflection of the Tongling Bridge caused by temperature was extracted, thereby obtaining the time series data of the temperature-induced deflection.
[0045] (2) The time interval between the temperature and temperature-induced deflection data collected in this embodiment is 10 minutes. If the time series data of 5 hours is selected as the input, the data volume of the input time series data of the temperature variable i=30. Assume that the average temperature of the main beam X G , the vertical temperature difference of the main beam is X GD , the tower temperature is X T , the temperature-induced deflection is Y, and the time-shift mapping mode is explained by taking time t as an example. The same can be applied to other time periods. The mapping mode of the 3×30 temperature matrix input at time t and the output temperature-induced deflection is shown in formula (1):
[0046]
[0047] (3) Normalize the data and set the average temperature of the main beam after normalization as x G , the vertical temperature difference of the main beam is x GD , the tower temperature is x T , the temperature-induced deflection is y.
[0048] (4) Build a neural network, such as Figure 2 For the neural network architecture of this embodiment, a convolution layer 1 with three convolution kernels is set after the input layer. The three convolution kernels A / B / C of this layer are all 3×6 in size. After convolution layer 1, there is a stacked long short-term memory layer (LSTM layer). This layer is composed of two stacked LSTM hidden layers. Each layer has 64 LSTM units. After the stacked LSTM layer, a convolution layer 2 is set. This layer has only one convolution kernel D of size 25×64. The size of the convolution kernel is the same as the output matrix of the stacked LSTM layer. Taking the regression value of the temperature-induced deflection at the 30th moment as an example, the data flow of the black-box neural network constructed in this embodiment is explained, as shown in FIG. Figure 3 is the data flow of temperature data in the neural network, and the input layer is a normalized 3×30 temperature matrix, as shown in formula (2):
[0049]
[0050] After convolution kernels A / B / C, local information enhanced data is formed after convolution processing. The convolution kernel moves 25 times along the time dimension to form 25 new data. The convolution kernel can be represented by a matrix. Taking convolution kernel A as an example, it can be represented by matrix M KA Indicates that M KA As shown in formula (3):
[0051]
[0052] In the formula, a 11 A ,a 12 A ,a 13 A ,...,a 34 A , a 35 A , a 36 A is the matrix M KA The elements in .
[0053] Suppose the new data formed by the three convolution kernels are x' A 、x' B and x' C , the data x' obtained by convolution kernel A in the first step of convolution operation A(1) Take an example to illustrate, x' A(1)The calculation formula is as follows:
[0054] x' A(1) =x G(1) a 11 A +x G(2) a 12 A +…+x G(6) a 16 A +x GD(1) a 21 A +x GD(2) a 22 A +
[0055] …+x GD(6) a 26 A +x T(1) a 31 A +x T(2) a 32 A +…+x T(6) a 36 A (4)
[0056] The other operation steps are the same, and a total of 25 new data x' are obtained A(1) , x' A(2) ,...,x' A(25) , using the matrix M of convolution kernel B and convolution kernel C KB With M KC The same operation is performed, and finally a new data matrix of 3×25 formed by new data is obtained as shown in formula (5):
[0057]
[0058] Substitute the matrix shown in formula (5) into the stacked LSTM layer and output a 25×64 matrix as shown in formula (6):
[0059]
[0060] Convolution kernel D(M KD ) is a 25×64 matrix as shown in formula (7):
[0061]
[0062] In the formula, a 1(1) D ,a 1(2) D ,a 1(3) D ,...,a64(23) D ,a 64(24) D ,a 64(25) D is the matrix M KD The elements in .
[0063] The matrix shown in formula (6) and the matrix M shown in formula (7) KD Perform convolution operation to obtain a single value O KD As shown in formula (8):
[0064]
[0065] O KD Substitute the linear activation function to obtain the normalized regression value y' at the 30th moment 30 , as shown in formula (9):
[0066] y′ 30 =W KD O KD +b KD (9)
[0067] Where W KD is the weight, b KD For bias.
[0068] (5) The output normalized regression value y' is denormalized to obtain the temperature-induced deflection regression value Y'.
[0069] (6) The first 75% of the collected data is used as a training set to train the neural network. The relevant hyperparameters are set as follows: Batch size is 10, learning rate (Ir) is 0.0001, number of iterative training (Epoch) is 100, and Adam optimization algorithm is used as the back propagation optimization algorithm.
[0070] (7) Using the regression value of temperature-induced deflection D' and the measured value of temperature-induced deflection D m The residual R AD As an expression index of abnormal deflection, the residual R AD The calculation formula is as follows:
[0071] R AD =D m -D′ (10)
[0072] like Figure 4 is the abnormal deflection value (R) obtained based on the deflection regression value of the high-performance deep neural network proposed in the present invention. AD), at this time we already know in advance that there is no abnormal deflection (that is, the theoretical abnormal deflection is 0), and the average error of the abnormal value obtained by the present invention is only 2.20mm, which is much more accurate than the traditional regression model. With a 95% guarantee rate, the upper and lower limits of the error are set, and the allowable error is only ±1.5mm, and the recognition sensitivity is as high as 3mm, which is only one thirty-thousandth of the span of the main span of the bridge, highlighting the excellent performance of the present invention.
[0073] The above embodiments are only further specific descriptions and examples of the scheme of the present invention. After reading the embodiments of the present invention, ordinary technicians in this field may make various equivalent modifications and substitutions to the present invention, which all fall within the scope of protection defined by the claims of the present invention.
Claims
1. A method for monitoring abnormal deflection of main beam of cable-stayed bridge based on deep learning model, characterized in that: The specific steps include: The temperature field data of the cable-stayed bridge is converted into time series data including n temperature variables, and the time series data of temperature-induced deflection in the change of main beam deflection is extracted based on the time-frequency characteristics of temperature. According to the time series correlation characteristics of temperature, the input and output mapping mode of temperature → temperature-induced deflection is set to input temperature time history data with i data points and output single-point temperature-induced deflection data at the current moment, thereby solidifying the input data specifications; Construct a black-box deep neural network mapping model. A first convolutional neural network layer with n convolutional kernels is added after the input layer of the model. The output of the convolutional neural network layer is passed to the stacked long short-term memory neural network layer. The output of the stacked long short-term memory neural network passes through the second convolutional neural network layer. The output of the second convolutional neural network layer passes through a linear activation function to obtain a normalized regression value. The black-box deep neural network mapping model is trained to obtain the temperature→temperature-induced deflection mapping model; The normalized regression value output by the temperature→temperature-induced deflection mapping model is denormalized to obtain a temperature-induced deflection regression value; The residual between the regression value of temperature-induced deflection and the measured value of temperature-induced deflection is used as the value of abnormal deflection to complete the abnormal deflection monitoring of the main beam of the cable-stayed bridge. Among them, the input data is an n×i temperature data matrix, and the output data is the temperature-induced deflection of a single point at the current moment. At the same time, a time-shifted data set is constructed over time for training and testing the black-box deep neural network mapping model; The core layer of the black-box deep neural network mapping model is set according to the data specifications and improvement requirements. In order to enhance the performance of extracting local temperature information, the first core layer is the first convolutional neural network layer with n convolution kernels, each of which is n×m in size. Each convolution kernel traverses the temperature data matrix once along the time dimension, and convolves the n×i temperature data matrix into an n×(i-m+1) matrix; the second core layer is a stacked long short-term memory neural network layer with two layers of hidden units, which is used to learn the temporal correlation characteristics of temperature data and inherit the n×(i-m+1) data matrix output by the first convolutional neural network layer. Each layer has p LSTM units, and the output of the stacked long short-term memory neural network layer is a p×(i-m+1) matrix; the third core layer is the second convolutional neural network layer with a p×(i-m+1) convolution kernel. After one convolution operation, the p×(i-m+1) matrix output by the stacked long short-term memory neural network layer is integrated into a single value through convolution operation.
2. The method according to claim 1, characterized in that The temperature variables include the average temperature of the main beam, the vertical temperature difference of the main beam and the tower temperature.
3. The method according to claim 1, characterized in that The stacked LSTM neural network layer has the same size as the second CNN layer.
