Bridge damage identification and life prediction method fusing deep learning
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANCHANG URBAN PLANNING & DESIGN RES INST GRP CO LTD
- Filing Date
- 2026-04-14
- Publication Date
- 2026-06-23
Smart Images
Figure CN122020265B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bridge engineering and structural health monitoring technology, specifically to a bridge damage identification and life prediction method that integrates deep learning. Background Technology
[0002] During long-term service, bridge structures gradually degrade in performance due to the combined effects of traffic loads and environmental factors. Establishing a structural health monitoring system based on massive amounts of monitoring data is a crucial means to ensure the safe operation of bridges. However, existing data analysis methods still face many challenges in long-term practical monitoring applications.
[0003] First, the quality of monitoring data is significantly affected by environmental noise and the performance of the sensors themselves. Existing data cleaning methods mainly rely on mathematical and statistical properties for filtering, lacking a verification mechanism that incorporates the physical and mechanical behavior of the structure. This makes it difficult to effectively distinguish between the aging drift of the sensor itself and the nonlinear hysteresis response of the structure due to material fatigue, easily leading to the accidental deletion of valid data containing damage information or the retention of false features.
[0004] Secondly, when using deep learning, especially graph neural networks, to process monitoring data, existing mapping methods are usually limited to the geometric proximity of the physical locations of sensors. This static topology ignores the inherent correlation between the stress patterns and risk evolution of the various components of the bridge, making it difficult for nodes that are physically far apart but in similar high-stress states or damage stages to interact effectively, thus limiting the model's ability to capture anomalies in complex structures.
[0005] Furthermore, the effect of ambient temperature on structural strain exhibits significant nonlinearity and hysteresis. Traditional linear regression or simple temperature compensation methods struggle to completely eliminate the temperature effect, and existing assessment systems often employ fixed alarm thresholds, failing to consider the objective law of natural performance degradation of bridges throughout their entire lifespan. This threshold setting method is highly prone to false alarms in the later stages of bridge service and also cannot accurately quantify the remaining service life of the structure.
[0006] Therefore, this invention proposes a bridge damage identification and life prediction method that integrates deep learning to address the shortcomings of existing technologies. Summary of the Invention
[0007] To address the shortcomings of existing technologies, this invention provides a bridge damage identification and life prediction method that integrates deep learning. This method solves the problems of low damage identification accuracy and unreliable remaining life prediction in existing bridge health monitoring technologies, which are caused by the inability to effectively separate environmental thermal effects and sensor drift noise, the lack of a feature association mechanism based on risk evolution, and the inability of alarm thresholds to adaptively adjust with structural aging.
[0008] To achieve the above objectives, the present invention provides the following technical solution: a bridge damage identification and life prediction method integrating deep learning, comprising the following steps:
[0009] The original stress time history data and surface temperature data collected by the bridge pre-embedded sensor network are obtained. Based on the heat conduction hysteresis law, the surface temperature data is converted into the effective temperature of the structure. The node comprehensive risk potential energy index that evolves with service time is calculated by combining the structural design parameters and the cumulative fatigue damage degree.
[0010] A theoretical hysteresis loop non-closure prediction model is established using the node comprehensive risk potential energy index. The measured hysteresis loop non-closure of the original stress time history data is compared with the output value of the theoretical hysteresis loop non-closure prediction model. Sensor drift noise is removed and the cleaned effective strain dataset is output.
[0011] The numerical differences of the node comprehensive risk potential index among the nodes in the sensor network are calculated. Based on the numerical differences, a dynamic weighted adjacency matrix based on risk homogeneity is constructed. A graph convolutional neural network is used to aggregate features of the effective strain dataset based on the dynamic weighted adjacency matrix, and the hidden layer feature vector is output.
[0012] The hidden layer feature vector and the effective temperature of the structure as a conditional constraint are input into a conditional variational autoencoder to generate a reconstruction residual. At the same time, the baseline alarm threshold is adjusted according to the average level of the node comprehensive risk potential energy index of all nodes of the bridge to generate a dynamic alarm threshold. By comparing the reconstruction residual with the dynamic alarm threshold, the damage identification and life prediction results are output.
[0013] Preferably, the step of converting the surface temperature data into the effective temperature of the structure based on the heat conduction hysteresis law specifically includes:
[0014] A discrete-time recursive correction model is adopted to calculate the effective temperature of the structure at the current sampling time using the effective temperature of the structure at the previous sampling time and the surface temperature data at the current sampling time.
[0015] The discrete-time recursive correction model defines a dimensionless thermal hysteresis coefficient, which is used to perform low-pass filtering and phase delay compensation on the high-frequency fluctuations of the surface temperature data.
[0016] Preferably, the step of calculating the nodal comprehensive risk potential energy index evolving over service time by combining structural design parameters and cumulative fatigue damage specifically includes:
[0017] Rainflow counting is performed on the original stress time history data to extract stress cycles, and the cumulative fatigue damage of each node is calculated by combining the material fatigue life curve.
[0018] Obtain the structural design parameters corresponding to each node, including the structural design stress concentration factor;
[0019] The stress concentration factor and cumulative fatigue damage of the structure are weighted and summed using a time-weighted function to obtain the comprehensive risk potential energy index of the node; the time-weighted function is configured to decrease monotonically with service time.
[0020] Preferably, the step of establishing a theoretical hysteresis loop non-closure prediction model using the node comprehensive risk potential energy index specifically includes:
[0021] The difference between the peak and valley values of the effective temperature of the structure is extracted as the effective temperature amplitude on a daily basis.
[0022] Obtain the nonlinear relationship between the baseline deformation coefficient, the risk sensitivity index, and the temperature activation factor;
[0023] Using the node comprehensive risk potential energy index at the current sampling time, the effective temperature amplitude, and the nonlinear relationship as inputs, and the theoretical hysteresis loop non-closure degree under a single temperature cycle as output, a theoretical hysteresis loop non-closure prediction model is constructed.
[0024] The theoretical hysteresis loop non-closure degree is the theoretical residual deformation of the structure due to viscoplastic damping or microcrack effects under the current risk potential energy level and temperature drive.
[0025] Preferably, the step of comparing the measured hysteresis loop non-closure degree of the original stress time history data with the output value of the theoretical hysteresis loop non-closure degree prediction model specifically includes:
[0026] The difference between the first and last data of the original stress time history data within the same temperature cycle period is calculated as the measured hysteresis loop non-closure degree.
[0027] Calculate the absolute deviation between the measured hysteresis loop non-closure degree and the theoretical hysteresis loop non-closure degree;
[0028] When the absolute deviation is less than the preset allowable error threshold, the current data segment corresponding to the original stress time history data is determined to be the true structural response, the current data segment is retained and marked as valid strain data;
[0029] When the absolute deviation is greater than or equal to the allowable error threshold, it is determined that the current data segment contains sensor drift noise, and data removal or baseline reset operation is performed.
[0030] Preferably, the step of constructing a dynamic weighted adjacency matrix based on risk homogeneity according to the numerical differences specifically includes:
[0031] For any two nodes in the sensor network, calculate the absolute value of the difference in the numerical value of the node comprehensive risk potential energy index of the two nodes at the current sampling time;
[0032] The absolute value of the numerical difference is mapped to the connection weight using an exponential decay function, wherein the smaller the absolute value of the numerical difference, the larger the connection weight.
[0033] A sparsity truncation threshold is set. When the absolute value of the numerical difference exceeds the sparsity truncation threshold, the corresponding connection weight is reset to zero, and only the node connections with similar risk characteristics are retained. A dynamic topology structure that evolves with the node comprehensive risk potential index is generated. This dynamic topology structure is the dynamic weighted adjacency matrix.
[0034] Preferably, in the step of using a graph convolutional neural network to perform feature aggregation on the effective strain dataset based on the dynamically weighted adjacency matrix, the specific process of constructing the input feature matrix of the graph convolutional neural network includes:
[0035] Set a preset time window length;
[0036] Extract the time series data of each node within the time window from the effective strain dataset, and use it as the effective strain time series vector;
[0037] Obtain the time series data of the effective structural temperature of each node within the time window, and use it as the time series vector of the effective structural temperature;
[0038] The effective strain time series vector of each node is concatenated with the effective temperature time series vector of the structure in the channel dimension to obtain the node feature vector of each node. The input feature matrix is composed of the node feature vectors of all nodes.
[0039] The graph convolutional neural network takes the input feature matrix as input, performs multi-layer graph convolution operations, and uses the dynamic weighted adjacency matrix to aggregate the features of neighboring nodes with similar risk potential, outputting the hidden layer feature vector containing structural spatiotemporal coupling information.
[0040] Preferably, the step of inputting the hidden layer feature vector and the effective temperature of the structure as a conditional constraint into a conditional variational autoencoder to generate the reconstruction residual specifically includes:
[0041] Construct the conditional variational autoencoder, which includes an encoder and a decoder, using the hidden layer feature vector as input data and the effective temperature of the structure as a conditional constraint;
[0042] The conditional variational autoencoder is trained by minimizing the reconstruction error and the potential distribution divergence.
[0043] The conditional variational autoencoder, after training, is used to reconstruct the hidden feature vector of the real-time input. The difference measure between the reconstructed features and the original input features is calculated, and the difference measure is used as the reconstruction residual.
[0044] Preferably, the step of adjusting the baseline alarm threshold based on the average level of the comprehensive risk potential energy index of all nodes of the bridge to generate a dynamic alarm threshold specifically includes:
[0045] Calculate the weighted average of the comprehensive risk potential energy index of all nodes of the bridge at the current sampling time, and use it as the average risk level of the entire bridge;
[0046] Based on the structural failure limit potential energy reference value, the benchmark alarm threshold is adjusted according to the nonlinear decay model and the average risk level of the whole bridge to obtain the dynamic alarm threshold. The dynamic alarm threshold decreases as the average risk level of the whole bridge increases. The nonlinear decay model controls the rate and magnitude of the decrease of the dynamic alarm threshold with the average risk level of the whole bridge through the adjustment coefficient and the sensitivity shape factor.
[0047] Preferably, the step of outputting damage identification and lifetime prediction results specifically includes:
[0048] When the reconstructed residual of any node exceeds the dynamic alarm threshold at the current sampling time, the node is determined to be abnormal, and the following processing is performed:
[0049] The abnormal physical coordinates are obtained by mapping the identifier of the node;
[0050] The dynamic risk level is determined based on the numerical range of the node's comprehensive risk potential index.
[0051] Based on the historical growth rate of the node's comprehensive risk potential energy index, the time span for the node's potential energy to reach the structural failure limit potential energy reference value is predicted using a time series extrapolation algorithm, and the remaining lifetime trend is output.
[0052] This invention provides a bridge damage identification and life prediction method integrating deep learning. It has the following beneficial effects:
[0053] 1. This invention establishes a hysteresis loop non-closure prediction model based on the node comprehensive risk potential energy index, thereby verifying the physical laws of the original monitoring data. Unlike traditional denoising methods that rely solely on numerical statistical features, this method utilizes thermodynamic hysteresis laws and the cumulative damage state of the structure to calculate the theoretical response. By comparing the deviation between the measured hysteresis characteristics and the theoretical prediction values, it can effectively distinguish between the true nonlinear response of the structure and the non-physical drift of the sensor itself. This mechanism solves the technical problem of baseline drift caused by sensor aging or environmental erosion in long-term monitoring systems, thus ensuring that the input data of the model reflects the true mechanical behavior of the structure.
[0054] 2. This invention constructs a dynamically weighted adjacency matrix based on the homogeneity of risk potential energy, and uses this matrix to drive a graph convolutional neural network for feature aggregation. This method overcomes the limitation of traditional graph neural networks that only construct topological structures based on the physical spatial distance of sensors, associating nodes that may be physically dispersed but have similar risk levels or damage evolution trends at the current moment. This graph construction method based on risk homogeneity rather than geographical proximity enables the network to capture the non-local correlation features of structures under complex stress states, improving the model's sensitivity to early minor damage to key stress components.
[0055] 3. This invention utilizes a conditional variational autoencoder and a dynamic alarm threshold to achieve deep decoupling of environmental effects and adaptive assessment throughout the entire life cycle. By using the effective temperature of the structure as a conditional constraint input to the conditional variational autoencoder, strain fluctuations caused by temperature changes are effectively eliminated, reducing the false alarm rate caused by seasonal temperature differences. At the same time, the dynamic alarm threshold mechanism, which evolves with the average risk level of the entire bridge, enables the monitoring system to adapt to the inherent performance degradation patterns during bridge service, avoiding the problem of fixed alarm thresholds failing in the later stages of service due to the decline in structural baseline performance, and providing an accurate basis for the quantitative prediction of the remaining life of the bridge. Attached Figure Description
[0056] Figure 1 This is a schematic diagram of the life prediction system of the present invention;
[0057] Figure 2 This is a flowchart of a bridge damage identification and life prediction method integrating deep learning according to an embodiment of the present invention.
[0058] Figure 3 This is a schematic diagram of the signal decoupling verification process based on the physical hysteresis mechanism of the present invention.
[0059] Figure 4 This is a schematic diagram of the dynamic topology reconstruction process based on potential energy homogeneity according to the present invention.
[0060] Figure 5This is a schematic diagram of the whole-life adaptive decision-making and life prediction process of the present invention.
[0061] Among them, 10 is the parameter calculation module; 20 is the signal verification module; 30 is the network reconstruction module; and 40 is the state evaluation module. Detailed Implementation
[0062] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0063] This embodiment provides a bridge damage identification and life prediction method integrating deep learning. This method runs on a computing device containing a processor and memory. The computing device is connected to the bridge structure's sensor network via an industrial fieldbus. The sensor network consists of standardized monitoring components pre-embedded during the bridge construction phase, retained as permanent structural parts within the concrete and at key stress sections. The standardized monitoring components adopt a "one-tube, two-grating" packaging form, specifically including a fiber optic strain grating and a fiber optic temperature compensation grating integrated within the same package tube, and a resistance surface thermometer mounted on the concrete surface. The fiber optic strain grating is used to sense structural mechanical deformation and thermal expansion, while the fiber optic temperature compensation grating, in a stress-free state, is used to monitor the internal local temperature at the measuring point location in real time for wavelength drift correction. The sensor network includes fiber optic strain gauges and resistance thermometers. The fiber optic strain gauges are positioned along the longitudinal direction of the main beam and the vertical direction of the tower, distributed according to the stress key points in the structural design drawings. The resistance thermometers are arranged adjacent to the fiber optic strain gauges and located on the structural surface at the same cross-sectional coordinate, used to acquire the thermal input of the external environment on the structural surface.
[0064] See attached document Figure 1 , Figure 1 This is a schematic diagram of the life prediction system that implements the bridge damage identification and life prediction method integrating deep learning as described in this invention. The system includes:
[0065] The parameter calculation module 10 is used to perform basic calculations of multidimensional physical fields, read the original stress time history data and surface temperature data, and calculate the effective temperature of the structure and the comprehensive risk potential energy index of the nodes.
[0066] The signal verification module 20 is used to determine the validity of the original stress time history data based on the physical hysteresis law, and output the cleaned valid strain dataset.
[0067] The network reconstruction module 30 dynamically generates a dynamic weighted adjacency matrix of the graph neural network based on the distribution of the node comprehensive risk potential index, and extracts the hidden layer feature vectors.
[0068] The condition assessment module 40 receives the hidden layer feature vector and uses it as a conditional constraint to input the effective temperature of the structure to generate a reconstruction residual. By comparing the reconstruction residual with the dynamic alarm threshold, the module outputs the damage identification and life prediction results.
[0069] See attached document Figure 2 , Figure 2 This is a flowchart of a bridge damage identification and life prediction method integrating deep learning according to an embodiment of the present invention. The method includes the following steps:
[0070] S100 acquires the raw stress time history data and surface temperature data collected by the bridge's embedded sensor network, and performs physical field correction and feature construction on the raw stress time history data and surface temperature data.
[0071] The computing device reads sensor network data at a preset frequency; it uses the thermal conduction hysteresis formula to recursively compensate the surface temperature data to obtain the effective temperature of the structure inside the structure; it performs rainflow counting on the original stress time history data, counts the cumulative fatigue damage of each node, and retrieves the pre-stored structural design stress concentration factor; it combines the cumulative fatigue damage and the structural design stress concentration factor to generate a node comprehensive risk potential energy index that evolves with service time.
[0072] S200, based on the node comprehensive risk potential energy index and the effective temperature of the structure, performs hysteresis decoupling verification on the original stress time history data and outputs an effective strain dataset.
[0073] Effective temperature-strain hysteresis curves are constructed on a daily basis. The theoretical hysteresis loop non-closure degree is calculated based on the nodal comprehensive risk potential energy index. The measured hysteresis loop non-closure degree is calculated based on the original stress time history data. The absolute deviation between the measured hysteresis loop non-closure degree and the theoretical hysteresis loop non-closure degree is calculated. If the absolute deviation is less than the preset allowable error threshold, the current data segment corresponding to the original stress time history data is determined to be the true structural response, and the current data segment is retained and marked as effective strain data. If the absolute deviation is greater than or equal to the allowable error threshold, the current data segment is determined to contain sensor drift noise, and data removal or baseline reset operations are performed.
[0074] S300 dynamically reconstructs the topology of the sensor network based on the node comprehensive risk potential energy index, and uses graph convolutional neural networks to extract the spatiotemporal coupling features of the effective strain dataset to generate hidden layer feature vectors.
[0075] Calculate the numerical difference of the comprehensive risk potential index between any two nodes; construct a dynamic weighted adjacency matrix based on this numerical difference, assign higher connection weights to nodes with similar risk potential, and input the dynamic weighted adjacency matrix and the effective strain dataset into a graph convolutional neural network. The graph convolutional neural network outputs a hidden layer feature vector containing structural spatial correlation information through multi-layer aggregation operations.
[0076] S400 performs adaptive anomaly detection and lifetime prediction throughout the entire lifecycle based on hidden layer feature vectors and effective structural temperature.
[0077] The hidden layer feature vector and the effective temperature of the structure as a conditional constraint are input into the conditional variational autoencoder. The conditional variational autoencoder outputs reconstructed features and calculates the reconstruction residual between the reconstructed features and the original input features. It then calculates the average risk level of the entire bridge based on the node comprehensive risk potential energy index of all nodes. Based on this average risk level, a dynamic alarm threshold that decreases nonlinearly over time is generated. The reconstruction residual is compared with the dynamic alarm threshold. When the reconstruction residual exceeds the dynamic alarm threshold, a monitoring result containing abnormal physical coordinates, remaining life trend, and dynamic risk level is generated.
[0078] To further clarify the implementation of each technical aspect of the present invention, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.
[0079] See attached document Figure 1 The execution of the method described in this embodiment depends on the sensor group and back-end computing terminal that are pre-embedded synchronously during the bridge construction stage.
[0080] The data acquisition is based on a pre-embedded sensor network, which is installed during the construction and pouring stage of the bridge's main structure. This sensor network primarily consists of fiber optic strain sensor arrays and temperature sensor arrays. Each fiber optic strain sensor array comprises several fiber optic nodes connected in series on a single-mode fiber. During the reinforcement binding process of the main beam, the fiber optic strain sensors are encapsulated in steel pipes for mechanical protection and spot-welded along the longitudinal direction of the main reinforcing steel bars before being poured into the concrete. The fiber optic strain sensors are positioned at the mid-span section, quarter-span section, and high-stress areas at the base of the pylons. For each measurement point where a fiber optic strain sensor is deployed, a temperature sensor is installed at the same cross-sectional coordinate on the corresponding concrete surface. The temperature sensor is used to acquire the thermal effect of the external environment on the structural surface, and it is strictly aligned spatially with the internal fiber optic strain sensor, forming a physical correspondence between "surface temperature and internal strain."
[0081] In terms of data transmission and processing, the sensor network is connected to the industrial control cabinet of the bridge maintenance center via armored optical cable; the data acquisition terminal adopts a multi-channel fiber Bragg grating demodulator, which is connected to the industrial control computer via an Ethernet interface; the fiber Bragg grating demodulator integrates a broadband light source and a spectral analysis module to detect the drift of the center wavelength of the fiber Bragg grating reflection.
[0082] The industrial control computer, serving as the execution entity of the method of this invention, is equipped with a processor and a large-capacity memory; the system sampling frequency is set to f. s The specific values are set to 50Hz to 100Hz to meet the sampling theorem requirements for dynamic vehicle loads; the data acquisition terminal records data in a continuous time stream and stores it according to a preset time window (e.g., 24 hours).
[0083] Define the basic data set for this embodiment: Assume the bridge is equipped with a total of... The sensor monitors several nodes; for any i-th node (where i = 1, 2, ..., N), the surface temperature data collected at time t is denoted as... The collected original stress time history data is denoted as .
[0084] Specifically, the original stress time history data The center wavelength output from the fiber optic grating demodulator is obtained by conversion according to the following formula:
[0085] ;
[0086] Among them, E c K represents the elastic modulus of concrete. ε K is the strain sensitivity coefficient of the fiber grating; T λ is the temperature sensitivity coefficient of the fiber Bragg grating; i (t) represents the measured center wavelength at time t; λ i,0 This is the initial reference wavelength; T0 is the local temperature measured by the temperature compensation grating integrated at the same location; T0 is the initial reference temperature. It characterizes the mechanical stress after deducting the effect of free thermal expansion.
[0087] The original stress time history data mentioned above is stored in the database in the form of time series vectors, serving as the input basis for subsequent multidimensional physical field correction and lifetime prediction algorithms; all nodes are assigned unique physical coordinate identifiers and channel numbers at this stage.
[0088] S110 converts surface temperature data into the effective temperature of the structure.
[0089] The concrete material of bridge structures has significant thermal inertia, which causes time lag and amplitude attenuation when changes in external ambient temperature are transmitted to the internal structure to generate stress response. Directly using surface temperature data cannot accurately describe the thermal expansion and contraction state inside the structure. This embodiment establishes a recursive mapping from surface temperature data to the effective temperature of the structure by using a simplified one-dimensional heat conduction model with discrete time.
[0090] For the Each sensor node, at sampling time t, calculates the effective structural temperature based on the previous sampling time. and surface temperature data at the current sampling time Calculate the effective temperature of the structure at the current sampling time. The specific recursive calculation formula is as follows:
[0091] ;
[0092] in, Let be the surface temperature data measured by the i-th sensor node at time t; The effective temperature of the structure is calculated at the previous sampling time; Δt is the sampling time interval; η is the dimensionless thermal hysteresis coefficient, with a value range of (0,1); the physical meaning of η is to characterize the rate at which the internal temperature response of the structure follows the change in surface temperature, and 1 / η approximately corresponds to the thermal time constant of the structure. In this implementation, η is determined through back-analysis of thermal conduction based on the thermal diffusivity of concrete, the sensor embedding depth, and the cross-sectional dimensions; this recursive mapping realizes low-pass filtering and phase delay compensation for high-frequency temperature fluctuations on the surface, thereby obtaining the true thermal field parameters that drive the thermal strain of the structure.
[0093] S120, based on the fusion of design parameters and measured data, constructs a node comprehensive risk potential energy index.
[0094] Node Comprehensive Risk Potential Index Ψ i (t) is a scalar field used to characterize the structural vulnerability of the region where the sensor node is located. This index is not a single stress amplitude, but a composite feature that integrates the static design properties of the structure (stress concentration) and the dynamic service history (fatigue cumulative damage).
[0095] First, calculate the cumulative fatigue damage at each node; based on the original stress-time history data... The stress cycles were extracted using the rainflow counting method, resulting in a series of stress cycle blocks with different amplitudes. The specific implementation algorithm of the rainflow counting method is a conventional technique well-known to those skilled in the art. Based on the extracted stress cycles, the cumulative fatigue damage up to time t was calculated using Miner's linear cumulative damage theory. :
[0096] ;
[0097] Where M is the number of stress cycles; n j N represents the number of measured cycles at the j-th stress amplitude; j The limit number of fatigue life cycles corresponding to this stress amplitude is determined by the material's SN curve (stress-life curve).
[0098] Secondly, retrieve the structural design stress concentration factor from the structural design data. This coefficient is calculated by the bridge finite element design model and represents the amplification factor of the local stress at the i-th node under standard load relative to the nominal stress. It reflects the inherent stress concentration characteristics (or the initial structural weakness) of this location in terms of geometry.
[0099] The node comprehensive risk potential energy index of the i-th node at time t ;
[0100] Where max(K) SCF ) represents the maximum value of the structural design stress concentration factor among all nodes of the bridge, used for normalization; D crit The preset critical fatigue damage threshold is set (e.g., 1.0); w(t) is a time weight function that decreases monotonically with service time t, used to achieve a smooth transition of risk assessment focus from "design-driven" to "damage-driven".
[0101] In this embodiment, w(t) specifically adopts the form of a logistic decay function:
[0102] ;
[0103] Among them, T design The design reference service life of the bridge; T half This refers to the half-life time point of the weight transition, i.e., when t=T. half At this time, the design factors and fatigue factors have equal weights; γ is the steepness coefficient controlling the conversion rate. Using this formula, the system dynamically updates the node comprehensive risk potential energy index of each node throughout the entire lifespan of the structure, providing a weighted basis with physical evolution laws for subsequent topology reconstruction based on graph neural networks.
[0104] See attached document Figure 3 , Figure 3 This is a schematic diagram of a signal decoupling verification process based on physical hysteresis mechanism according to an embodiment of the present invention.
[0105] S210, a theoretical hysteresis loop non-closure prediction model is established based on the node comprehensive risk potential energy index.
[0106] Under the alternating day and night temperature field, the strain response of a bridge structure does not follow a strict linear path with temperature change, but rather exhibits a hysteresis loop characteristic with a certain area. This hysteresis phenomenon originates from the viscoplastic damping of the material and the opening and closing effect of microcracks. In particular, for areas in high-risk conditions (i.e., areas with high comprehensive risk potential energy index of nodes), the irreversible deformation (residual deformation) after a single temperature cycle will increase significantly due to stiffness degradation caused by damage accumulation.
[0107] This embodiment uses the nodal comprehensive risk potential energy index calculated in step S120 as physical prior knowledge to quantitatively predict the theoretical residual deformation under a single-day temperature cycle; a complete temperature cycle is set as 24 hours, and the effective temperature amplitude ΔT within this cycle is extracted. cyc Based on the structural damage evolution mechanism, the theoretical hysteresis loop non-closure degree of the i-th node in the current period. Calculated based on the following nonlinear relationship:
[0108] ;
[0109] Among them, Ψ i (t) represents the node comprehensive risk potential index of the i-th node at the current sampling time t; ΔT cyc α is the peak-to-valley difference of the effective temperature of the structure within the calculation period; k is the baseline deformation coefficient related to the rheological properties of concrete; k is the risk sensitivity index, used to describe the nonlinear amplification effect of damage on residual deformation, usually with a value greater than 1; β is the temperature activation factor. It is a hyperbolic sine function used to characterize the nonlinear excitation effect of thermal driving force on residual deformation.
[0110] The specific values of the above parameters α, k, and β are determined during the system initialization phase by selecting a segment of health monitoring data from the early stage of bridge opening (at which point the comprehensive risk potential energy index of the node is considered to be at a low level and there is no substantial damage), and using the nonlinear least squares method to perform regression fitting on the above formula; or by calibrating based on laboratory creep and thermal expansion test data of similar grade concrete materials.
[0111] This formula suggests that, theoretically, observable residual deformation will only occur when the structure has a high risk potential energy and is significantly driven by temperature.
[0112] S220 performs signal drift detection and data cleaning based on the comparison of residuals between measured data and theoretical predictions.
[0113] Extract the initial and final data of the raw stress time history data collected by the sensor within the same temperature cycle, and calculate the measured hysteresis loop non-closure degree. The measured hysteresis loop non-closure degree is composed of the cumulative deviation of the original stress time history data itself. The measured hysteresis loop non-closure degree includes the actual physical deformation of the structure and the zero-point drift error of the sensor itself.
[0114] To eliminate spurious signals caused by sensor malfunctions, a verification criterion based on physical consistency is constructed; the absolute deviation between the measured hysteresis loop non-closure degree and the theoretical hysteresis loop non-closure degree is calculated and compared with a preset tolerance error threshold.
[0115] ;
[0116] Where, δ th The basic tolerance threshold set for the system is based on the sensor's factory accuracy specifications; σ noise γ is the standard deviation of the background noise in the current data segment; γ is the confidence interval coefficient (e.g., take 3).
[0117] When the above inequality is satisfied, the non-closed characteristics of the current data segment are determined to conform to the physical damage law of the structure, and the data is valid. The current data segment is then marked as valid strain data and retained in the valid strain dataset. At this time, the non-closed quantity is considered to be the true cumulative damage manifestation of the structure.
[0118] When the above inequality is not satisfied, since the calculation model of the theoretical hysteresis loop non-closure degree already includes the nodal comprehensive risk potential energy index, that is, it has already considered the real physical hysteresis increment that the structure may generate due to cumulative damage or stress concentration, when the absolute deviation significantly exceeds the physical scope explained by risk potential energy, it is determined that the data segment has non-physical drift or abnormal jump of the sensor itself; the system performs a rejection operation on the current data segment and resets and corrects the sensor baseline to prevent false trends from misleading the subsequent life prediction model; through this step, it is ensured that the data input to the neural network only contains information reflecting the real mechanical behavior of the structure.
[0119] See attached document Figure 4 , Figure 4 This is a schematic diagram of a dynamic topology reconstruction process based on potential energy homogeneity according to an embodiment of the present invention.
[0120] S310, construct a dynamic weighted adjacency matrix based on risk homogeneity.
[0121] Traditional graph neural networks, when processing bridge monitoring data, typically construct the topology based on the Euclidean distance between sensors in three-dimensional space, assuming that nodes with closer physical locations have higher correlation. However, in structural life-cycle assessment, physical proximity does not necessarily indicate consistent damage evolution patterns. Nodes at different spans but with similar stress states (e.g., both at mid-span) or similar damage levels often exhibit higher time-series correlation. This embodiment abandons the fixed geometric distance mapping method and constructs a dynamically changing risk homogeneity topology based on the node comprehensive risk potential energy index calculated in step S120.
[0122] For any two nodes i and j in the sensor network, the system calculates the connection weight A between them at time t. ij (t); the magnitude of this weight directly depends on the degree of difference in the comprehensive risk potential index of the current nodes of the two nodes; the specific formula for calculating the elements of the adjacency matrix is as follows:
[0123] ;
[0124] Among them, Ψ i (t) and Ψ j (t) represents the node comprehensive risk potential energy index of node i and node j at time t; μ is the potential energy similarity decay coefficient, which is used to control the rate at which the connection weight decays as the potential energy difference increases. The smaller the value of μ, the more likely it is that only nodes with highly similar potential energy are allowed to establish strong connections. The sparsity truncation threshold is used to forcibly disconnect the connection (reset the connection weight to 0) when the potential energy difference between two nodes exceeds this threshold, in order to eliminate noise interference between heterogeneous nodes and reduce computational complexity.
[0125] Using the above formula, the system logically clusters and connects all nodes of the bridge with similar risk characteristics (such as being high stress concentration areas or high fatigue damage areas) to form several "risk subgraphs," regardless of how far apart these nodes are in physical space. As time goes by and structural damage accumulates, the node comprehensive risk potential energy index of each node dynamically evolves, and the dynamic weighted adjacency matrix is also updated adaptively to reflect the current damage distribution topology of the structure in real time.
[0126] S320 performs feature propagation and aggregation based on Graph Convolutional Networks (GCN).
[0127] Using the constructed dynamic weighted adjacency matrix, features are extracted from the effective strain dataset output in step S200 in the graph convolutional network (GCN) layer. Through graph convolution operations, each node not only utilizes its own time series information, but also aggregates the feature information of its neighboring nodes with similar risk, thereby enhancing the ability to capture minor structural anomalies.
[0128] Constructing the input feature matrix H of a graph convolutional neural network (0) For time t, a time window of length L is extracted (e.g., L=50 sampling points), and the effective strain data and effective temperature of the structure of N nodes of the bridge are used as node features.
[0129] Specifically, the input feature matrix H (0) The dimension is (N, F), where N is the number of nodes and F is the node feature dimension. In this embodiment, in order to preserve the complete temporal fluctuation information, F is defined as 2×L, that is, it is formed by splicing the effective strain time series vector of length L and the structural effective temperature time series vector of length L in the channel dimension.
[0130] For the i-th node, its feature vector is represented as: ;
[0131] Where L is the length of the time window; The effective strain data of the i-th node at time t; Let be the effective structural temperature of the i-th node at time t; through this concatenation method, the input feature matrix simultaneously includes the mechanical response history and the thermodynamic environment history; in the l-th layer graph convolutional network, the propagation and updating of features follow the following formula:
[0132] Among them, H (l) H is the input feature matrix of the l-th layer; (l+1) W is the output feature matrix of the l-th layer; (l) is the learnable weight matrix to be trained in the l-th layer; f() is a non-linear activation function (such as ReLU or TanH function); To add self-loops to the adjacency matrix, ensuring that nodes retain their own characteristics during aggregation, I N It is the identity matrix. For connection weight A ij The dynamic weighted adjacency matrix at time t is constructed using (t), where N×N indicates that the dynamic weighted adjacency matrix is an N-row, N-column matrix, and N is the total number of nodes in the sensor network. for The degree matrix is a diagonal matrix with all off-diagonal elements being 0. The diagonal element in the i-th row and i-th column of this degree matrix is... , Representation matrix The element in the i-th row and j-th column, item Symmetric normalization of the adjacency matrix is implemented to prevent the feature values from exploding or disappearing during multi-layer propagation.
[0133] After multi-layer graph convolution operations, the feature vectors of each sensor node are fused with the structural response patterns of all nodes of the bridge with similar risk potential energy. This feature aggregation mechanism based on potential energy homogeneity enables the system to use the collective behavior of similar nodes to verify the anomaly of a single node, which has a significant signal-to-noise ratio gain for identifying early local damage. The resulting high-dimensional node features (hidden layer feature vectors) are then fed into the subsequent prediction module.
[0134] See attached document Figure 5 , Figure 5 This is a schematic diagram of a life-cycle adaptive decision-making and life-cycle prediction process according to an embodiment of the present invention.
[0135] S410 is based on Conditional Variational Autoencoder (C-VAE) for unsupervised feature reconstruction and residual generation.
[0136] To extract anomalous changes caused by structural damage from complex nonlinear features, this embodiment constructs a variational autoencoder conditioned on environmental parameters. The high-dimensional node features output from the graph convolutional network in step S300 are used as input data X, along with the corresponding effective structural temperature T. eff The encoder and decoder network (conditional variational autoencoder) is introduced as a conditional variable c; this conditional variational autoencoder aims to learn the joint distribution law of structural response and ambient temperature under normal conditions.
[0137] During training, the encoder maps the input data to a latent space distribution, while the decoder reconstructs the original input based on the sampled latent vector and condition variables. The optimization objective of the conditional variational autoencoder is to minimize the difference between the reconstruction loss and the latent distribution. The total loss function L... total The definition is as follows:
[0138] ;
[0139] Among them, L total This represents the overall objective loss function, used to measure the combined performance of model reconstruction quality and potential spatial distribution characteristics; θ represents the set of weights and bias parameters of the encoder neural network; θ represents the set of weights and bias parameters of the decoder neural network; X represents the input data, which includes the spatiotemporal characteristics of effective strain data and effective structural temperature; c represents the condition variable, specifically the effective structural temperature, used to constrain the generated context; z represents the latent variable vector in the latent space, used to capture the low-dimensional core manifold features of the data. For parameters The defined probabilistic encoder distribution (approximate posterior distribution) is used to map the input data to the latent space; p θ(X|z,c) is the probability decoder distribution (likelihood distribution) defined by parameter θ, used to reconstruct the original input data from the latent variables; E is the mathematical expectation operator, representing the expectation of the latent variable vector z from the distribution. The average reconstructed log-likelihood obtained from mid-sample; logp θ (X|z,c) is the log-likelihood function, and its negative value represents the difference (reconstruction error) between the reconstructed data and the original input data X; D KL (·||·) is the Kullback-Leibler divergence operator, used to measure the difference between two probability distributions; p(z|c) is the prior distribution of the latent variable vector, pre-defined as the standard normal distribution N(0,I).
[0140] The first term is the reconstruction error, which characterizes the decoder output. The ability to recover the original input X; the second term is the Kullback-Leibler (KL) divergence, which constrains the distribution of the latent variable to approximate the prior normal distribution.
[0141] After training, the conditional variational autoencoder is used to infer the real-time monitored input data. For each sensor node, the reconstruction residual R(t) between its reconstructed features and the original input features is calculated. This reconstruction residual eliminates interpretable fluctuations caused by normal temperature changes and retains uninterpretable components caused by structural material degradation or damage, serving as the quantitative basis for subsequent anomaly detection.
[0142] S420 is based on dynamic alarm threshold adaptive control driven by the comprehensive risk potential energy of the full bridge.
[0143] Traditional structural health monitoring typically uses fixed alarm thresholds, which ignores the physical fact that a structure's ability to resist environmental loads gradually declines with the length of its service life. This embodiment proposes an adaptive threshold mechanism that evolves with the structure's state.
[0144] The weighted average of the comprehensive risk potential energy index of all nodes of the bridge is calculated and denoted as the average risk level of the entire bridge. Set the initial baseline alarm threshold to τ. base This value is determined based on the statistical distribution of data from bridge completion load tests (e.g., taking 3 times the standard deviation); as the service time increases, the dynamic alarm threshold τ... alert (t) Automatically adjusted based on the following nonlinear decay model:
[0145] ;
[0146] in, Ψ represents the average risk level of the entire bridge at the current moment. limitξ is the defined reference value for the structural failure limit potential energy; ξ is the adjustment coefficient, with a value range of (0,1), used to set the maximum reduction of the threshold under the extreme risk state; m is the sensitivity shape factor.
[0147] This formula clarifies the negative correlation between the dynamic alarm threshold and the risk potential energy: when a bridge is in its early stages of service, the risk potential energy is low, and the system maintains a high dynamic alarm threshold to reduce the false alarm rate; as the bridge experiences fatigue accumulation or stress concentration, the average risk level of the entire bridge decreases. Increase, dynamic alarm threshold τ alert (t) automatically decreases, thereby increasing the system's sensitivity to minute damage signals and realizing an adaptive strategy of "the more fragile the structure, the more stringent the monitoring".
[0148] S430 outputs anomaly location, risk classification, and remaining life trend prediction results.
[0149] The system combines the reconstruction residual R(t) calculated by S410 with the dynamic alarm threshold τ calculated by S420. alert (t) Perform real-time comparison; when the reconstruction residual of a node exceeds the dynamic alarm threshold at the current moment, the system determines that the node has an anomaly.
[0150] The output module specifically includes the following information:
[0151] Abnormal physical coordinates: Based on the sensor node ID that exceeded the limit, it is mapped back to the specific geometric location of the bridge's 3D model (e.g., "left main span L / 4 section bottom plate").
[0152] Dynamic risk level: based on the node's comprehensive risk potential index Ψ i The numerical range of (t) divides the structural state into four levels: "healthy", "observation", "early warning" and "dangerous".
[0153] Remaining lifetime trend: Based on the node comprehensive risk potential index Ψ i The historical growth rate of (t) is used to predict the potential energy at this node reaching the destruction limit Ψ using a time series extrapolation algorithm. limit The remaining time span is used as a reference value for the remaining life of the component.
[0154] See attached document Figure 1 The system mainly includes:
[0155] The parameter calculation module 10 is configured to perform basic calculations and parameter transformations of multidimensional physical fields; specifically, the parameter calculation module 10 is used to read the raw stress time history data and surface temperature data collected by the sensor network.
[0156] On the one hand, through the preset thermal inertia recursive correction logic, the unsteady surface temperature data is mapped to the effective temperature of the structure that characterizes the real thermal state inside the structure;
[0157] On the other hand, the parameter calculation module 10 integrates the structural design stress concentration factor in the structural design data with the cumulative fatigue damage degree calculated based on historical stress cycles, and quantitatively characterizes the current node comprehensive risk potential energy index of each node, providing physical prior parameters for subsequent data cleaning and topology construction.
[0158] The signal verification module 20 is configured to perform data validity determination and cleaning based on physical hysteresis laws. This module utilizes the node comprehensive risk potential energy index and temperature cycle amplitude to predict the theoretical hysteresis loop non-closure degree under a single temperature cycle. By calculating the absolute deviation between the measured and theoretical hysteresis loop non-closure degrees and comparing it with an allowable error threshold based on sensor accuracy, it identifies and eliminates false signals containing non-physical drift or abnormal jumps, outputting a valid strain dataset reflecting the true mechanical behavior of the structure.
[0159] The network reconstruction module 30 is configured to dynamically generate a graph neural network topology based on potential homogeneity. This module 30 abandons the traditional fixed geometric distance graph construction method, calculates connection weights based on the numerical differences in the comprehensive risk potential index between nodes, and clusters and connects nodes with similar risk characteristics across the entire bridge. Based on this, graph convolution operations are used to perform feature propagation and aggregation on the generated risk homogeneous topology (dynamically weighted adjacency matrix), extracting high-dimensional features of nodes that incorporate group behavior patterns.
[0160] The condition assessment module 40 is configured to perform anomaly detection and lifespan prediction decisions throughout the entire lifespan. This module 40 utilizes a variational autoencoder conditioned on the effective temperature of the structure to reconstruct the high-dimensional features of the input nodes, generating a reconstruction residual that eliminates environmental temperature variations. Simultaneously, based on the evolution of the bridge's overall average risk level, the module adaptively adjusts the dynamic alarm threshold, establishing a dynamic alarm threshold curve that decreases with structural aging. By comparing the reconstruction residual with the dynamic alarm threshold in real time, the module 40 outputs specific anomaly physical coordinates, the current dynamic risk level, and the remaining lifespan trend of the components derived from the potential energy growth trend.
[0161] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A bridge damage identification and life prediction method integrating deep learning, characterized in that, Includes the following steps: The original stress time history data and surface temperature data collected by the bridge pre-embedded sensor network are obtained. Based on the heat conduction hysteresis law, the surface temperature data is converted into the effective temperature of the structure. The node comprehensive risk potential energy index that evolves with service time is calculated by combining the structural design parameters and the cumulative fatigue damage degree. A theoretical hysteresis loop non-closure prediction model is established using the node comprehensive risk potential energy index. The measured hysteresis loop non-closure of the original stress time history data is compared with the output value of the theoretical hysteresis loop non-closure prediction model. Sensor drift noise is removed and the cleaned effective strain dataset is output. The numerical differences of the node comprehensive risk potential index among the nodes in the sensor network are calculated. Based on the numerical differences, a dynamic weighted adjacency matrix based on risk homogeneity is constructed. A graph convolutional neural network is used to aggregate features of the effective strain dataset based on the dynamic weighted adjacency matrix, and the hidden layer feature vector is output. The hidden layer feature vector and the effective temperature of the structure as a conditional constraint are input into a conditional variational autoencoder to generate a reconstruction residual. At the same time, the baseline alarm threshold is adjusted according to the average level of the node comprehensive risk potential energy index of all nodes of the bridge to generate a dynamic alarm threshold. By comparing the reconstruction residual with the dynamic alarm threshold, the damage identification and life prediction results are output.
2. The bridge damage identification and life prediction method integrating deep learning according to claim 1, characterized in that, The step of converting the surface temperature data into the effective temperature of the structure based on the thermal conduction hysteresis law specifically includes: A discrete-time recursive correction model is adopted to calculate the effective temperature of the structure at the current sampling time using the effective temperature of the structure at the previous sampling time and the surface temperature data at the current sampling time. The discrete-time recursive correction model defines a dimensionless thermal hysteresis coefficient, which is used to perform low-pass filtering and phase delay compensation on the high-frequency fluctuations of the surface temperature data.
3. The bridge damage identification and life prediction method integrating deep learning according to claim 1, characterized in that, The steps for calculating the nodal comprehensive risk potential energy index evolving over service time by combining structural design parameters and cumulative fatigue damage include: Rainflow counting is performed on the original stress time history data to extract stress cycles, and the cumulative fatigue damage of each node is calculated by combining the material fatigue life curve. Obtain the structural design parameters corresponding to each node, including the structural design stress concentration factor; The stress concentration factor and cumulative fatigue damage of the structure are weighted and summed using a time-weighted function to obtain the comprehensive risk potential energy index of the node; the time-weighted function is configured to decrease monotonically with service time.
4. The bridge damage identification and life prediction method integrating deep learning according to claim 1, characterized in that, The steps for establishing a theoretical hysteresis loop non-closure prediction model using the aforementioned node comprehensive risk potential energy index specifically include: The difference between the peak and valley values of the effective temperature of the structure is extracted as the effective temperature amplitude on a daily basis. Obtain the nonlinear relationship between the baseline deformation coefficient, risk sensitivity index, and temperature activation factor; Using the node comprehensive risk potential energy index at the current sampling time, the effective temperature amplitude, and the nonlinear relationship as inputs, and the theoretical hysteresis loop non-closure degree under a single temperature cycle as output, a theoretical hysteresis loop non-closure prediction model is constructed. The theoretical hysteresis loop non-closure degree is the theoretical residual deformation of the structure due to viscoplastic damping or microcrack effects under the current risk potential energy level and temperature drive.
5. The bridge damage identification and life prediction method integrating deep learning according to claim 4, characterized in that, The step of comparing the measured hysteresis loop non-closure degree of the original stress time history data with the output value of the theoretical hysteresis loop non-closure prediction model specifically includes: The difference between the first and last data of the original stress time history data within the same temperature cycle period is calculated as the measured hysteresis loop non-closure degree. Calculate the absolute deviation between the measured hysteresis loop non-closure degree and the theoretical hysteresis loop non-closure degree; When the absolute deviation is less than the preset allowable error threshold, the current data segment corresponding to the original stress time history data is determined to be the true structural response, the current data segment is retained and marked as valid strain data; When the absolute deviation is greater than or equal to the allowable error threshold, it is determined that the current data segment contains sensor drift noise, and data removal or baseline reset operation is performed.
6. The bridge damage identification and life prediction method integrating deep learning according to claim 1, characterized in that, The steps for constructing a dynamic weighted adjacency matrix based on risk homogeneity according to the numerical differences specifically include: For any two nodes in the sensor network, calculate the absolute value of the difference in the numerical value of the node comprehensive risk potential energy index of the two nodes at the current sampling time; The absolute value of the numerical difference is mapped to the connection weight using an exponential decay function, wherein the smaller the absolute value of the numerical difference, the larger the connection weight. A sparsity truncation threshold is set. When the absolute value of the numerical difference exceeds the sparsity truncation threshold, the corresponding connection weight is reset to zero, and only the node connections with similar risk characteristics are retained. A dynamic topology structure that evolves with the node comprehensive risk potential index is generated. This dynamic topology structure is the dynamic weighted adjacency matrix.
7. The bridge damage identification and life prediction method integrating deep learning according to claim 1, characterized in that, In the step of using a graph convolutional neural network to aggregate features of the effective strain dataset based on the dynamically weighted adjacency matrix, the specific process of constructing the input feature matrix of the graph convolutional neural network includes: Set a preset time window length; Extract the time series data of each node within the time window from the effective strain dataset, and use it as the effective strain time series vector; Obtain the time series data of the effective structural temperature of each node within the time window, and use it as the time series vector of the effective structural temperature; The effective strain time series vector of each node is concatenated with the effective temperature time series vector of the structure in the channel dimension to obtain the node feature vector of each node. The input feature matrix is composed of the node feature vectors of all nodes. The graph convolutional neural network takes the input feature matrix as input, performs multi-layer graph convolution operations, and uses the dynamic weighted adjacency matrix to aggregate the features of neighboring nodes with similar risk potential, outputting the hidden layer feature vector containing structural spatiotemporal coupling information.
8. The bridge damage identification and life prediction method integrating deep learning according to claim 1, characterized in that, The step of inputting the hidden layer feature vector and the effective temperature of the structure as a conditional constraint into a conditional variational autoencoder to generate the reconstruction residual specifically includes: Construct the conditional variational autoencoder, which includes an encoder and a decoder, using the hidden layer feature vector as input data and the effective temperature of the structure as a conditional constraint; The conditional variational autoencoder is trained by minimizing the reconstruction error and the potential distribution divergence. The conditional variational autoencoder, after training, is used to reconstruct the hidden feature vector of the real-time input. The difference measure between the reconstructed features and the original input features is calculated, and the difference measure is used as the reconstruction residual.
9. The bridge damage identification and life prediction method integrating deep learning according to claim 1, characterized in that, The steps for generating a dynamic alarm threshold by adjusting the baseline alarm threshold based on the average level of the comprehensive risk potential energy index of all nodes of the bridge include: Calculate the weighted average of the comprehensive risk potential energy index of all nodes of the bridge at the current sampling time, and use it as the average risk level of the entire bridge; Based on the structural failure limit potential energy reference value, the benchmark alarm threshold is adjusted according to the nonlinear decay model and the average risk level of the whole bridge to obtain the dynamic alarm threshold. The dynamic alarm threshold decreases as the average risk level of the whole bridge increases. The nonlinear decay model controls the rate and magnitude of the decrease of the dynamic alarm threshold with the average risk level of the whole bridge through the adjustment coefficient and the sensitivity shape factor.
10. The bridge damage identification and life prediction method integrating deep learning according to claim 1, characterized in that, The steps for outputting damage identification and lifetime prediction results specifically include: When the reconstructed residual of any node exceeds the dynamic alarm threshold at the current sampling time, the node is determined to be abnormal, and the following processing is performed: The abnormal physical coordinates are obtained by mapping the identifier of the node; The dynamic risk level is determined based on the numerical range of the node's comprehensive risk potential index. Based on the historical growth rate of the node's comprehensive risk potential energy index, the time span for the node's potential energy to reach the structural failure limit potential energy reference value is predicted using a time series extrapolation algorithm, and the remaining lifetime trend is output.