CN-122020265-A - Bridge damage identification and life prediction method integrating deep learning

CN122020265ACN 122020265 ACN122020265 ACN 122020265ACN-122020265-A

Abstract

The application relates to the technical field of bridge engineering and structural health monitoring, and discloses a bridge damage identification and life prediction method integrating deep learning, which comprises the steps of firstly calculating effective temperature of a structure and comprehensive risk potential energy indexes of nodes based on a heat conduction hysteresis rule and accumulated fatigue damage degree; the method comprises the steps of verifying actual measurement of the non-closure degree of a stagnation ring by using a theoretical non-closure degree prediction model, effectively eliminating sensor drift noise, constructing a dynamic weighting adjacent matrix according to the numerical value difference of node comprehensive risk potential energy among all nodes, carrying out feature aggregation on an effective variable data set, finally, dividing hidden layer feature vectors and structural effective temperature input condition variations serving as condition constraints from an encoder, generating a reconstruction residual error, and comparing the reconstruction residual error with a dynamic alarm threshold. According to the application, through data cleaning and risk homogeneity topology construction driven by a physical rule, environmental thermal effect interference is effectively decoupled, and accurate state evaluation and life prediction of the bridge life cycle are realized.

Inventors

Tong Hanyuan
WEN ZHIPENG
HU SHENGLIANG
LIU CHENGHAI
Xiang Yinsong
Xiao gui
XIE BIN
ZOU JIAYI
ZHU KAI
XU YONGQING
HE FENG
LU JIANQING
XIAO TAO
CAO FANGLIANG

Assignees

南昌市城市规划设计研究总院集团有限公司

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (10)

1. The bridge damage identification and life prediction method integrating deep learning is characterized by comprising the following steps of: Acquiring original stress time-course data and surface temperature data acquired by a bridge embedded sensor network, converting the surface temperature data into structural effective temperature based on a heat conduction hysteresis rule, and calculating a node comprehensive risk potential energy index evolving along with service time by combining structural design parameters and accumulated fatigue damage degree; Establishing a theoretical hysteresis loop non-closure degree prediction model by utilizing the node comprehensive risk potential energy index, comparing the actually measured hysteresis loop non-closure degree of the original stress time-course data with the output value of the theoretical hysteresis loop non-closure degree prediction model, removing sensor drift noise and outputting a cleaned effective strain data set; calculating the numerical value difference of node comprehensive risk potential energy indexes among all nodes in the sensor network, constructing a dynamic weighted adjacent matrix based on risk homogeneity according to the numerical value difference, carrying out feature aggregation on an effective strain data set based on the dynamic weighted adjacent matrix by using a graph convolutional neural network, and outputting a hidden layer feature vector; Inputting the hidden layer feature vector and the structural effective temperature serving as a condition constraint into a condition variation self-encoder to generate a reconstruction residual error, adjusting a reference alarm threshold according to the average level of the node comprehensive risk potential energy indexes of all the nodes of the bridge to generate a dynamic alarm threshold, and comparing the reconstruction residual error with the dynamic alarm threshold to output damage identification and service life prediction results.
2. The method for identifying bridge damage and predicting service life based on fusion deep learning of claim 1, wherein the step of converting the surface temperature data into the structural effective temperature based on the thermal conduction hysteresis law comprises the following steps: calculating the structure effective temperature at the current sampling time by using the structure effective temperature at the last sampling time and the surface temperature data at the current sampling time by adopting a discrete time recursion correction model; The discrete time recursive correction model defines a dimensionless thermal hysteresis coefficient that is used to low pass filter and phase delay compensate for high frequency fluctuations of the surface temperature data.
3. The method for identifying and predicting the service life of the bridge damage by fusion deep learning according to claim 1, wherein the step of calculating the node comprehensive risk potential energy index evolving with the service time by combining the structural design parameter and the accumulated fatigue damage degree comprises the following steps: Performing rain flow counting on the original stress time interval data, extracting stress circulation, and calculating the accumulated fatigue damage degree of each node by combining a material fatigue life curve; Obtaining the structural design parameters corresponding to each node, wherein the structural design parameters comprise structural design stress concentration coefficients; and carrying out weighted summation on the structural design stress concentration coefficient and the accumulated fatigue damage degree by using a time weight function to obtain the node comprehensive risk potential energy index, wherein the time weight function is configured to be monotonically decreased along with the service time.
4. The method for identifying bridge damage and predicting service life by fusion deep learning according to claim 1, wherein the step of establishing a theoretical hysteresis loop non-closure prediction model by using the node comprehensive risk potential energy index specifically comprises the following steps: Extracting the difference value between the peak value and the valley value of the effective temperature of the structure as effective temperature amplitude variation by taking a single day as a period; Acquiring a nonlinear relation among a reference deformation coefficient, a risk sensitivity index and a temperature activating factor; Taking the node comprehensive risk potential energy index, the effective temperature amplitude and the nonlinear relation at the current sampling moment as inputs, taking the theoretical hysteresis loop non-closure degree under single temperature cycle as output, and constructing a theoretical hysteresis loop non-closure degree prediction model; The theoretical hysteresis loop non-closure degree is the theoretical residual deformation of the structure due to the viscoplastic damping or microcrack effect under the current risk potential energy level and temperature driving.
5. The method for identifying bridge damage and predicting service life by fusion deep learning according to claim 4, wherein the step of comparing the actually measured hysteresis loop non-closure degree of the original stress time course data with the output value of the theoretical hysteresis loop non-closure degree prediction model specifically comprises the following steps: calculating the difference value of the head and tail end data of the original stress time interval data in the same temperature cycle period as the actually measured hysteresis loop non-closure degree; calculating an absolute deviation between the measured hysteresis loop non-closure degree and the theoretical hysteresis loop non-closure degree; When the absolute deviation is smaller than a preset allowable error threshold, judging that a current data segment corresponding to the original stress time-course data is a structure real response, and reserving the current data segment and marking the current data segment as effective strain data; And when the absolute deviation is greater than or equal to the allowable error threshold, judging that the current data segment contains sensor drift noise, and executing data rejection or baseline reset operation.
6. The method for identifying bridge damage and predicting service life by fusion deep learning according to claim 1, wherein the step of constructing a dynamic weighted adjacency matrix based on risk homogeneity according to the numerical difference comprises the following steps: for any two nodes in the sensor network, calculating the absolute value of the numerical difference of the node comprehensive risk potential energy index of the two nodes at the current sampling moment; mapping the absolute value of the numerical value difference into a connection weight by using an exponential decay function, wherein the smaller the absolute value of the numerical value difference is, the larger the connection weight is; Setting a sparsification cutoff threshold, resetting the corresponding connection weight to zero when the absolute value of the numerical difference exceeds the sparsification cutoff threshold, only preserving node connection with similar risk characteristics, and generating a dynamic topological structure evolving along with the node comprehensive risk potential energy index, wherein the dynamic topological structure is the dynamic weighting adjacent matrix.
7. The method for identifying bridge damage and predicting service life by fusion deep learning according to claim 1, wherein in the step of feature-aggregating the effective strain data set based on the dynamic weighted adjacency matrix by using a graph convolution neural network, the specific process of constructing the input feature matrix of the graph convolution neural network comprises the following steps: setting a time window with a preset length; extracting time sequence data of each node in the time window from the effective data set to be used as an effective time sequence vector; Acquiring time sequence data of the effective temperature of the structure corresponding to each node in the time window, and taking the time sequence data as a time sequence vector of the effective temperature of the structure; Splicing the effective strain time sequence vector of each node and the structural effective temperature time sequence vector in a channel dimension to obtain node feature vectors of each node, wherein the input feature matrix is formed by the node feature vectors of all nodes; And the graph convolution neural network takes the input feature matrix as input, utilizes the dynamic weighting adjacent matrix to aggregate the features of neighbor nodes with similar risk potential energy through multi-layer graph convolution operation, and outputs the hidden layer feature vector containing structural space-time coupling information.
8. The method for identifying bridge damage and predicting life by fusion depth learning according to claim 1, wherein the step of dividing the hidden layer feature vector and the structural effective temperature input condition variation as a condition constraint from an encoder to generate a reconstructed residual, specifically comprises: constructing the conditional variation self-encoder comprising an encoder and a decoder, taking the hidden layer feature vector as input data and taking the effective temperature of the structure as a conditional constraint; training the conditional variation self-encoder by minimizing reconstruction errors and potential distribution divergences; and reconstructing the hidden layer feature vector input in real time by using the trained conditional variation self-encoder, calculating the difference metric between the reconstructed feature and the original input feature, and taking the difference metric as a reconstructed residual.
9. The method for identifying bridge damage and predicting service life by fusion deep learning according to claim 1, wherein the step of adjusting the reference alarm threshold according to the average level of the node comprehensive risk potential energy index of all the nodes of the bridge to generate the dynamic alarm threshold specifically comprises the following steps: Calculating the weighted average value of the node comprehensive risk potential energy indexes of all the nodes of the bridge at the current sampling moment, and taking the weighted average value as the full-bridge average risk level; And adjusting a reference alarm threshold according to a nonlinear attenuation model and a full-bridge average risk level based on a structural damage limiting potential energy reference value to obtain the dynamic alarm threshold, wherein the dynamic alarm threshold is reduced along with the increase of the full-bridge average risk level, and the nonlinear attenuation model controls the speed and the amplitude of the dynamic alarm threshold along with the decrease of the full-bridge average risk level through an adjusting coefficient and a sensitivity shape factor.
10. The method for identifying and predicting the life of the bridge damage by fusion deep learning according to claim 1, wherein the step of outputting the result of identifying and predicting the life of the bridge damage specifically comprises the following steps: when the reconstructed residual error of any node exceeds a dynamic alarm threshold value of the current sampling moment, judging that the node is abnormal, and executing the following processing: mapping according to the identification of the node to obtain an abnormal physical coordinate; Determining a dynamic risk level according to a numerical interval of the node comprehensive risk potential energy index of the node; Based on the historical growth rate of the node comprehensive risk potential energy index of the node, predicting the time span of the potential energy of the node reaching the structural damage limit potential energy reference value by utilizing a time sequence extrapolation algorithm, and outputting the residual life trend.

Description

Bridge damage identification and life prediction method integrating deep learning Technical Field The invention relates to the technical field of bridge engineering and structural health monitoring, in particular to a bridge damage identification and life prediction method integrating deep learning. Background In the long-term service process of the bridge structure, the structural performance of the bridge structure is gradually degraded under the coupling effect of traffic load and environmental factors. The building of the structural health monitoring system based on mass monitoring data is a key means for guaranteeing the operation safety of bridges. However, existing data analysis methods still face many challenges in long-period practical monitoring applications. First, the quality of the monitored data is significantly affected by ambient noise and the performance of the sensor itself. The existing data cleaning method mainly relies on the mathematical statistics characteristic to carry out filtering, and a verification mechanism combining physical and mechanical behaviors of a structure is lacked, so that aging drift of a sensor and nonlinear hysteresis response of the structure caused by material fatigue are difficult to distinguish effectively, and effective data containing damage information are easy to delete by mistake or false features are reserved. Second, existing patterning approaches are often limited to the geometric proximity of the physical locations of the sensors when processing the monitored data using deep learning, and in particular, graph neural networks. The static topological structure ignores the internal association of each component of the bridge in the stress mode and risk evolution, so that the nodes which are far away in physical position and in similar high-stress states or damage stages are difficult to perform effective characteristic interaction, and the capturing capability of the model on complex structural anomalies is limited. Furthermore, the effect of ambient temperature on structural strain has significant nonlinearity and hysteresis. The traditional linear regression or simple temperature compensation method is difficult to thoroughly strip the temperature effect, and the existing evaluation system mostly adopts a fixed and invariable alarm threshold value, and does not consider the objective rule of natural decay of the bridge performance in the whole life cycle. The threshold setting mode is extremely easy to generate false alarm in the later period of bridge service, and meanwhile, the residual service life of the structure cannot be accurately quantized. Therefore, the invention provides a bridge damage identification and life prediction method integrating deep learning, which solves the defects in the prior art. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a bridge damage identification and life prediction method integrating deep learning, which solves the problems of low damage identification accuracy and unreliable residual life prediction caused by the fact that the existing bridge health monitoring technology cannot effectively strip environmental thermal effect and sensor drift noise, lacks a feature association mechanism based on risk evolution and the alarm threshold cannot be adaptively adjusted along with structure aging. In order to achieve the purpose, the invention is realized by the following technical scheme that the bridge damage identification and service life prediction method integrating deep learning comprises the following steps: Acquiring original stress time-course data and surface temperature data acquired by a bridge embedded sensor network, converting the surface temperature data into structural effective temperature based on a heat conduction hysteresis rule, and calculating a node comprehensive risk potential energy index evolving along with service time by combining structural design parameters and accumulated fatigue damage degree; Establishing a theoretical hysteresis loop non-closure degree prediction model by utilizing the node comprehensive risk potential energy index, comparing the actually measured hysteresis loop non-closure degree of the original stress time-course data with the output value of the theoretical hysteresis loop non-closure degree prediction model, removing sensor drift noise and outputting a cleaned effective strain data set; calculating the numerical value difference of node comprehensive risk potential energy indexes among all nodes in the sensor network, constructing a dynamic weighted adjacent matrix based on risk homogeneity according to the numerical value difference, carrying out feature aggregation on an effective strain data set based on the dynamic weighted adjacent matrix by using a graph convolutional neural network, and outputting a hidden layer feature vector; Inputting the hidden layer feature vector and the structural effective temperature serving