Civil engineering structure safety risk intelligent early warning system

Abstract

The application relates to the technical field of structural health monitoring, and discloses a civil engineering structure safety risk intelligent early warning system, which comprises physical quantity collection, phase difference extraction, dissipation feature extraction, state evaluation, life prediction and model updating, early warning decision and visual display modules, the system acquires high-frequency acceleration, dynamic strain and environmental temperature sequences of monitoring nodes, extracts transient phase difference, combines strain sequences to construct a phase space trajectory to calculate a hysteresis area ratio, inputs a temperature decoupling feature into a neural network to acquire a real physical classification level and predict service days, updates a cloud twin model unit stiffness matrix by using Kalman filtering, and combines a maintenance index to render a damage thermodynamic diagram on a component surface. The application extracts dynamic dissipation features, performs temperature decoupling, removes environmental drift interference, realizes residual life estimation of a physical structure and synchronous correction of a cloud twin model, and improves risk determination accuracy.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring technology, specifically to an intelligent early warning system for structural safety risks in civil engineering. Background Technology

[0002] During long-term service, civil engineering structures are subjected to multiple effects such as material aging, fatigue damage, and external environmental vibration, resulting in a gradual decline in structural mechanical properties. To address this, the industry has widely deployed structural health monitoring systems, which use various sensors to collect physical parameters such as acceleration and strain to assess the physical service status of the structure. Existing civil engineering structural safety early warning systems rely on absolute threshold comparison of single-source physical parameters to determine the state. This single-source physical threshold comparison method remains at the stage of direct comparison of surface data, without involving joint analysis of underlying dynamic mechanisms such as internal stiffness degradation and energy dissipation. This results in a delay in risk warning. At the same time, the stiffness characteristics of structural materials and the output baseline of sensors are directly affected by the periodic fluctuations of external ambient temperature. Existing systems fail to decouple and eliminate temperature drift effects at the physical level, causing non-structural physical deformation data to be mixed into the extracted feature parameters, leading to high-frequency false alarms. Furthermore, the data processing architecture and underlying mechanism model of conventional monitoring systems operate in isolation. The structural finite element simulation analysis model is built on the initial health benchmark. After nonlinear damage and aging occur in the physical structure, the element stiffness matrix inside the model lacks a mechanism for adaptive closed-loop correction based on on-site measured feedback. The isolated state of the data model causes the simulation model to deviate from the evolution trajectory of the real structure, losing its long-term theoretical reference value. Based on the static evaluation framework, the existing early warning system only outputs the immediate status alarm after the fact, and cannot use the time-series cumulative characteristics of multi-dimensional parameters to perform continuous quantitative calculation of the remaining service life of the structure. As a result, the inspection scheduling and maintenance decisions on the engineering site lack coherent data support and forward-looking planning. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides an intelligent early warning system for safety risks in civil engineering structures. This system solves the problems of existing systems relying on a single physical quantity threshold and failing to eliminate temperature drift interference, as well as the inability to dynamically correct stiffness parameters and estimate remaining life due to isolated data models.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an intelligent early warning system for safety risks in civil engineering structures, comprising: The physical quantity acquisition module acquires the high-frequency acceleration sequence, dynamic strain sequence, and ambient temperature sequence of the structural monitoring nodes; The phase difference extraction module extracts the transient phase difference sequence between adjacent monitoring nodes based on the high-frequency acceleration sequence; The dissipation feature extraction module extracts a data window when the dynamic strain rate of the dynamic strain sequence exceeds the product threshold of the historical maximum strain rate and the safety tolerance coefficient corresponding to the random environmental excitation. The dynamic strain sequence and transient phase difference sequence within the data window are combined to construct a two-dimensional phase space discrete trajectory to calculate the hysteresis area ratio. The status assessment module combines the ambient temperature sequence to decouple the hysteresis area ratio into the temperature-decoupled hysteresis area ratio, inputs it into the multilayer perceptron classification model for decoding and mapping, and outputs the true health physical classification level. The life prediction and model update module feeds the temperature decoupling hysteresis area ratio into the long short-term memory neural network model to obtain the predicted remaining service days. When the predicted remaining service days reach the safe service days limit, the Kalman filter algorithm is used to iteratively correct the stiffness matrix parameters of the cloud digital twin model. The early warning decision-making and visualization module calculates the comprehensive maintenance priority index based on the real health physical classification level and the predicted remaining service days. When the comprehensive maintenance priority index falls into the highest response trigger range, the module renders the damage heat map texture on the surface of the corresponding physical component in the cloud digital twin model.

[0005] This invention simultaneously acquires multi-source heterogeneous sensor data, extracts phase space features reflecting local energy dissipation and stiffness changes in the structure, eliminates physical quantity drift errors caused by ambient temperature by constructing a temperature decoupling baseline, and establishes a data mapping path between structural state classification and remaining life prediction by combining a multilayer perceptron and a long short-term memory neural network. At the same time, it uses the Kalman filter algorithm to feed online real-time observation data back to the cloud-based digital twin model to correct stiffness matrix parameters, realizing dynamic interaction between structural monitoring data and the underlying physical mechanism model, and improving the accuracy of safety risk assessment.

[0006] Furthermore, the physical quantity acquisition module acquires the high-frequency acceleration sequence, dynamic strain sequence, and ambient temperature sequence of the structural monitoring node. The physical quantity acquisition module performs mean-reduction and de-trendization processing on the high-frequency acceleration sequence, dynamic strain sequence, and ambient temperature sequence, respectively deducting the arithmetic mean of the corresponding sequence and the linear trend term data obtained by least squares fitting, and outputs the preprocessed sequence. The preprocessing step performs baseline calibration on the acquired raw time-series signals to eliminate the inherent zero bias of the sensors and the low-frequency trend drift caused by non-load factors, so that the structural dynamic response analysis is based on a unified zero-mean benchmark, reducing the numerical calculation error in the feature extraction process.

[0007] Furthermore, the phase difference extraction module constructs a digital Butterworth filter by setting the lower and upper limits of the passband frequency based on the dominant natural frequency of the target structure. The phase difference extraction module inputs the high-frequency acceleration sequence into the digital Butterworth filter for forward and reverse filtering respectively. By time reversal reconstruction and superposition, the phase deviation between forward and reverse filtering is canceled to achieve zero-group delay alignment and output the dominant response sequence. The phase difference extraction module performs a discrete Hilbert transform on the dominant response sequence to obtain an orthogonal signal sequence. Based on the dominant response sequence and the orthogonal signal sequence, the four-quadrant arctangent function is substituted to calculate the wrapped phase sequence. The phase difference extraction module monitors the wrapped phase difference value of the wrapped phase sequence, controls the step compensation integer to perform a step compensation operation to obtain absolute phase data, and performs a subtraction operation on the absolute phase data of adjacent monitoring nodes at the same discrete time sampling point to output the transient phase difference sequence. To address the phase delay problem caused by traditional filtering, a digital Butterworth bidirectional filtering technique is used to offset the phase shift and isolate environmental noise interference outside the target frequency band. The transient phase of the analytical signal is extracted in the time domain through discrete Hilbert transform, and the phase difference time-series evolution data of elastic wave propagation between adjacent monitoring nodes under dynamic excitation is obtained, providing physical parameters for extracting energy dissipation characteristics.

[0008] Furthermore, the dissipative feature extraction module performs a difference operation on the dynamic strain sequence after the moving average filtering to obtain the discrete dynamic strain rate as the dynamic strain rate. When the discrete dynamic strain rate exceeds the product threshold, the dissipative feature extraction module generates a trigger signal. The discrete time sampling point corresponding to the trigger signal is used as the starting point to extract a data window that is an integer multiple of the natural vibration period of the target structure foundation. In the two-dimensional Cartesian coordinate system, the dissipative feature extraction module uses the strain sequence in the data window as the abscissa and the corresponding transient phase difference sequence as the ordinate to map and generate a set of transient phase space trajectories. The dissipative feature extraction module converts the set of transient phase space trajectories into a set of forced closed trajectories by supplementing linear interpolation segments. The module performs a polygon cross multiplication and summation operation on the coordinate points contained in the set of forced closed trajectories to obtain the polygon envelope area. The polygon envelope area is calculated and divided by the area of ​​the outer envelope rectangle corresponding to the set of forced closed trajectories, and the hysteresis area ratio is output. Discrete dynamic strain rate is used as the trigger condition for data interception to filter out static load interference. The dynamic strain sequence and transient phase difference sequence are combined to construct a two-dimensional phase space trajectory. The forced closed trajectory is used to form a hysteresis space similar to the force-displacement relationship. The polygonal envelope area is calculated by cross-multiplication and summation. The hysteresis area ratio reflects the degree of local energy dissipation and stiffness degradation of the structure at the physical level and serves as an indicator for judging early minor damage to the structure.

[0009] Furthermore, the state assessment module extracts the ambient temperature sequence and calculates the arithmetic mean temperature value of discrete temperature data points within the data window. The state assessment module substitutes the arithmetic mean temperature value into the linear temperature mapping baseline function established using non-destructive initial service stage data combined with the least squares fitting algorithm to calculate the expected area ratio reference value under the corresponding temperature conditions. The state assessment module subtracts the expected area ratio reference value from the hysteresis area ratio and outputs the temperature decoupling hysteresis area ratio. Changes in ambient temperature cause temperature drift in the elastic modulus of structural materials and thermal expansion and contraction of sensors. The state judgment deviation is directly generated by using characteristic parameters. A temperature mapping baseline function is established using the initial non-destructive stage data. The expected area ratio reference value under the current temperature conditions is calculated and subtracted, thereby eliminating structural response changes caused by solar radiation and seasonal temperature differences and reducing the state misjudgment rate caused by temperature interference.

[0010] Furthermore, the state assessment module mixes historical measured abnormal data with finite element simulation analysis data to construct an offline training sample set containing one-hot encoded real labels. Based on the cross-entropy loss function value and the adaptive moment estimation optimization algorithm combined with the backpropagation mechanism, iteratively updates the network weight matrix and bias vector of the multilayer perceptron classification model. The state assessment module processes and reconstructs the temperature decoupling hysteresis area ratio using the maximum-minimum normalization algorithm into a single-dimensional input feature. In the multilayer perceptron classification model that has completed parameter iterative training, the output layer calculates and outputs a three-dimensional state probability vector. The state assessment module extracts the dimension space index corresponding to the probability element with the largest value in the three-dimensional state probability vector and decodes and maps the dimension space index to the real health physical classification level. Finite element simulation analysis data is introduced into the offline training set to make up for the lack of actual engineering fault samples. The input data dimensions are unified by the maximum and minimum normalization process. The normalized decoupled features are mapped into a three-dimensional state probability vector using a multilayer perceptron classification model. The index corresponding to the maximum probability is used to determine the current structural health level, thus solving the mapping problem of nonlinear structural degradation state classification.

[0011] Furthermore, the life prediction and model update module resamples the temperature decoupling hysteresis area ratio continuously output from the historical cycle into equidistant time-series feature data with fixed time intervals using a first-order linear interpolation algorithm. The equidistant time-series feature data is then truncated at a fixed time step to generate time-series sample data in three-dimensional tensor format. The life prediction and model update module mixes the collected accelerated fatigue test degradation data with finite element simulation data to construct an offline prediction sample set labeled with continuous numerical values ​​of real life. Based on the mean square error loss function combined with the backpropagation mechanism, the network weight matrix of the long short-term memory neural network model is iteratively updated. The life prediction and model update module inputs the three-dimensional tensor format time-series sample data into the trained long short-term memory neural network model to calculate and output the predicted remaining service days. Structural damage degradation has a time-cumulative effect. By resampling historical continuous feature data and constructing it into an equidistant three-dimensional tensor format sequence to meet the temporal input requirements of the long short-term memory neural network, and training the model with accelerated fatigue test degradation data, the internal memory units of the long short-term memory neural network are used to extract the historical damage evolution law, calculate and output the specific remaining service days, and provide continuous life prediction results.

[0012] Furthermore, when the predicted remaining service days are less than or equal to the safe service days limit, the life prediction and model update module uses the local finite element stiffness reduction coefficient corresponding to the damaged entity component in the cloud-based digital twin model as the state variable of the Kalman filter, and uses the online real-time acquired temperature decoupling hysteresis area ratio as the observation variable of the Kalman filter. The life prediction and model update module controls the cloud-based digital twin model to perform forward dynamic calculations to obtain the theoretical hysteresis area ratio. Based on the observation residuals of the online real-time acquired temperature decoupling hysteresis area ratio and the theoretical hysteresis area ratio, as well as the partial derivatives of the theoretical hysteresis area ratio, a sensitivity observation matrix is ​​constructed. The life prediction and model update module calculates the Kalman gain based on the sensitivity observation matrix and the prediction error covariance, and uses the Kalman gain multiplied by the observation residual to perform a posteriori iterative update of the local finite element stiffness reduction coefficient. The updated and converged local finite element stiffness reduction coefficient is multiplied into the initial element stiffness matrix of the cloud-based digital twin model to achieve the correction of the stiffness matrix parameters. When the structure reaches the preset safe service limit, the observation residual is constructed by using the measured temperature decoupling hysteresis area ratio and the theoretical area ratio calculated by the forward solution of the digital model. By calculating the sensitivity observation matrix and Kalman gain, the local finite element stiffness reduction coefficient of the cloud digital twin model is corrected online a posteriori. This allows the cloud digital model parameters to be updated synchronously with the actual degradation degree of the physical entity structure, thereby reducing the dynamic characteristic deviation between the virtual and real models.

[0013] Furthermore, the early warning decision and visualization module establishes corresponding discrete severity weight coefficients for different real health physical classification levels, extracts matching discrete severity weight coefficients based on the real health physical classification level, calculates the difference between the design baseline expected life and the predicted remaining service days and divides it by the design baseline expected life to obtain the life decay rate, and performs a weighted summation operation on the discrete severity weight coefficients and the life decay rate to output the comprehensive maintenance priority index. Structural maintenance decisions are influenced by both the current damage status and long-term service life. Severity weights are assigned based on the actual physical classification level, and the attenuation ratio of remaining life to design life is calculated. The discrete severity is weighted and summed with the continuous life attenuation rate to output a single quantitative comprehensive maintenance priority index, which guides the maintenance sequence and resource allocation of on-site components.

[0014] Furthermore, when the comprehensive maintenance priority index falls into the highest response trigger range, the early warning decision and visualization module generates an emergency dispatch instruction and an abnormal alarm message. The spatial location node number specified in the abnormal alarm message is used as the central rendering anchor point. The early warning decision and visualization module extracts the local finite element mesh size of the cloud digital twin model and multiplies it by the mesh scaling factor to obtain the spatial influence attenuation parameter. The early warning decision and visualization module applies the Gaussian radial basis function containing the spatial influence attenuation parameter to calculate the stress concentration influence weight of the area surrounding the central rendering anchor point. Based on the stress concentration influence weight, a damage heat map texture with a nonlinear transition from red to yellow is generated and mapped to the surface of the corresponding solid component in the cloud digital twin model for rendering. By using the Gaussian radial basis function based on the distance decay principle to calculate the stress concentration weight, the mechanical influence range of local structural damage on the surrounding area is simulated. Combined with the local finite element mesh size, a corresponding mesh scaling factor is generated. The nonlinear transition damage heat map texture is calculated and generated. The damage heat map texture is directly rendered on the surface of the cloud digital twin model, and the numerical alarm is converted into a three-dimensional spatial distribution graphical display result, which facilitates intuitive location of the damaged location and the range of influence of the structure.

[0015] This invention provides an intelligent early warning system for structural safety risks in civil engineering. It has the following beneficial effects: 1. This invention extracts the transient phase difference sequence of the high-frequency acceleration sequence and constructs a two-dimensional phase space discrete trajectory by combining it with the dynamic strain sequence to calculate the hysteresis area ratio. At the same time, it establishes a temperature mapping baseline function by combining the initial non-destructive stage data to physically decouple the hysteresis area ratio. The phase space feature extraction and temperature decoupling mechanism eliminates the numerical fluctuations caused by environmental temperature drift, so that the physical characteristics of the final input evaluation model truly reflect the local stiffness degradation and energy dissipation state of the structure, and reduce the false alarm rate of warnings caused by non-load environmental factors.

[0016] 2. This invention resamples the temperature decoupling hysteresis area ratio output from historical cycles into time-series sample data in an equidistant three-dimensional tensor format, and inputs it into a long short-term memory neural network model for computation. The resampling and tensor input mechanism uses the continuous damage characteristics of the long-sequence time dimension as the basis for calculation, and combines actual degradation data and simulation data to optimize the network weights, thereby calculating a specific numerical prediction of the remaining service days, realizing the transformation from traditional static threshold judgment to dynamic remaining life prediction.

[0017] 3. This invention activates the Kalman filter algorithm when the predicted remaining service days reach the safety limit, calculates the Kalman gain using the observation residual between online measured features and digital twin model solution features, performs online posterior update of the local finite element stiffness reduction coefficient of the cloud mechanism model, and maps a damage heat map on the surface of the digital model when a high-priority response interval is reached. The Kalman filter closed-loop architecture enables the underlying mechanism model to synchronously correct the stiffness matrix according to the actual aging degree of the physical structure, reduces the dynamic characteristic deviation between the virtual and real models, and provides intuitive three-dimensional spatial damage location results. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the system timing interaction of the present invention; Figure 2 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the physical quantity acquisition and preprocessing process of the present invention; Figure 4 This is a schematic diagram of the dissipation feature extraction process of the present invention; Figure 5 This is a schematic diagram of the lifetime prediction and model update process of the present invention; Figure 6 This is a diagram illustrating the evolution of structural safety risks of the present invention over its service life. Detailed Implementation

[0019] 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.

[0020] See appendix Figure 1 This invention provides an intelligent early warning system for structural safety risks in civil engineering, the system comprising: The physical quantity acquisition module is configured to acquire the high-frequency acceleration sequence, dynamic strain sequence, and ambient temperature sequence of the structural monitoring nodes; The phase difference extraction module is configured to calculate the transient phase difference sequence between adjacent monitoring nodes based on zero-phase bandpass filtering and analytical transform. The dissipation feature extraction module is configured to extract a data window when the dynamic strain rate exceeds a set threshold. In this embodiment, the safety tolerance coefficient corresponding to the set threshold is in the range of 1.5 to 3.0. The module constructs a two-dimensional phase space discrete trajectory and calculates the hysteresis area ratio. The status assessment module is configured to calculate the temperature decoupling hysteresis area ratio based on the arithmetic mean temperature value and the linear temperature mapping baseline function, and input the temperature decoupling hysteresis area ratio into the multilayer perceptron classification model to perform forward inference calculation, decode the mapping and output the true health physical classification level of the target structure. The life prediction and model update module is configured to resample the temperature decoupling hysteresis area ratio within the historical cycle and extract it as time-series sample data. The time-series sample data is then fed into the long short-term memory neural network model to predict the remaining service days. When the remaining service days reach a set limit, the stiffness matrix parameters of the cloud-based digital twin model are iteratively corrected using the Kalman filter algorithm. The early warning decision and visualization module is configured to calculate the comprehensive maintenance priority index based on the weighted average of the actual health physical classification level and the predicted remaining service days. When the comprehensive maintenance priority index triggers the set risk range, it combines the Gaussian radial basis function to render the damage heat map texture on the surface of the digital twin model in the cloud and issues an emergency dispatch instruction.

[0021] See appendix Figure 2 This invention provides an intelligent early warning method for safety risks in civil engineering structures, the method comprising the following steps: S1. Establish a global synchronization clock, acquire the high-frequency acceleration sequence, dynamic strain sequence, and ambient temperature sequence of the structural monitoring node, and perform mean-reduction and de-trendization processing on the high-frequency acceleration sequence, dynamic strain sequence, and ambient temperature sequence. S2, perform zero-phase bandpass filtering on the high-frequency acceleration sequence to obtain the dominant response sequence, perform discrete Hilbert transform and phase dewinding on the dominant response sequence to obtain the absolute phase, and calculate the transient phase difference sequence between adjacent monitoring nodes; S3, obtain the dynamic strain rate based on the dynamic strain sequence, and when the dynamic strain rate exceeds the set threshold, cut off the data window, reconstruct the dynamic strain sequence and transient phase difference sequence in the data window into a forced closed phase space discrete trajectory, and calculate the hysteresis area ratio. S4. Extract the baseline hysteresis area ratio data corresponding to the historical lossless initial stage to establish a linear temperature mapping baseline function. Subtract the corresponding expected area ratio reference value from the calculated hysteresis area ratio to obtain the temperature decoupling hysteresis area ratio. Input the temperature decoupling hysteresis area ratio into the multilayer perceptron classification model that has completed parameter iteration training to perform forward inference. Decode the output vector of the multilayer perceptron classification model to obtain the true health physical classification level of the target structure. S5 resamples the continuously output temperature decoupling hysteresis area ratio sequence and reconstructs it into time-series sample data by truncating it at a fixed time step. The time-series sample data is fed into the long short-term memory neural network model to calculate the predicted remaining service days. When the predicted remaining service days are less than or equal to the safe service days limit, the parameters of the cloud digital twin model are iteratively corrected by fusing real-time feature data using the Kalman filter algorithm. S6 configures discrete severity weight coefficients for different real health physical classification levels, and calculates the comprehensive maintenance priority index by weighting the predicted remaining service days. When the comprehensive maintenance priority index falls into the highest response trigger range, it uses a Gaussian radial basis function containing spatial influence attenuation parameters to render a damage heat map texture on the surface of the corresponding entity component in the cloud digital twin model, and generates an emergency dispatch instruction.

[0022] See appendix Figure 3 In specific implementations, step S1 provided by the present invention may include the following steps: S11, the physical quantity acquisition module deploys monitoring nodes consisting of multiple sensors of different types in the key load-bearing areas of the civil engineering structure, and configures a distributed sensor network topology based on a distributed data acquisition hardware terminal that supports a precise time protocol. The physical quantity acquisition module sends clock synchronization messages to each distributed data acquisition hardware terminal through a fiber optic Ethernet bus to perform global clock synchronization. After completing global clock synchronization, each distributed data acquisition hardware terminal triggers an analog-to-digital conversion unit in conjunction with a set basic sampling frequency. In this embodiment, the value range of the basic sampling frequency is 100 to 500 Hz. The physical quantity acquisition module assigns a nanosecond-level timestamp with a unified time reference to the discrete data points collected by different sensors.

[0023] S12, the physical quantity acquisition module acquires the high-frequency acceleration sequence output by the piezoelectric accelerometer, the dynamic strain sequence output by the fiber optic strain gauge, and the ambient temperature sequence output by the temperature sensor of the monitoring node based on the nanosecond-level timestamp. The physical quantity acquisition module upsamples the ambient temperature sequence to the same discrete time node as the high-frequency acceleration sequence using zero-order hold interpolation. The physical quantity acquisition module outputs the high-frequency acceleration sequence, dynamic strain sequence, and ambient temperature sequence with the same sampling step size. The physical quantity acquisition module strictly aligns and combines the high-frequency acceleration sequence, dynamic strain sequence, and ambient temperature sequence in the data matrix according to the discrete time index to form a multi-dimensional heterogeneous data source.

[0024] S13, the physical quantity acquisition module extracts the original discrete sequence containing a set number of discrete sampling points. In this embodiment, the set number of discrete sampling points ranges from 1024 to 4096. The physical quantity acquisition module calculates the arithmetic mean of the original discrete sequence and subtracts the arithmetic mean from the original discrete sequence point by point. The physical quantity acquisition module performs linear fitting on the sequence after subtracting the arithmetic mean based on the least squares method to obtain the slope and intercept of the linear trend term. The physical quantity acquisition module simultaneously subtracts the arithmetic mean and the reconstructed linear trend data through a single algebraic operation. The physical quantity acquisition module outputs the preprocessed sequence after the mean and trend removal processing. The comprehensive data processing calculation formula for the preprocessed sequence is as follows: ; In the formula, The extracted preprocessed sequence values; This refers to the value of the discrete dynamic physical response sequence obtained from the original acquisition at the current discrete-time sampling point; This represents the total number of discrete-time sampling points covered by the physical response sequence. Summation traversal index for discrete-time sampling points with values ​​in the range of positive integers; This represents the value of the discrete dynamic physical response sequence obtained from the original acquisition under the summation traversal index; The slope coefficient of the linear trend term is obtained by using the least squares fitting algorithm; This is the index of the discrete-time sampling point sequence, with values ​​ranging from positive integers. This is the intercept constant of the linear trend term obtained by using the least squares fitting algorithm.

[0025] In specific implementations, step S2 provided by this invention may include the following steps: S21, the phase difference extraction module retrieves the design data or historical vibration test data of the civil engineering structure to obtain the dominant natural frequency of the target structure. The phase difference extraction module sets the lower limit and upper limit of the passband frequency of the digital bandpass filter with the dominant natural frequency as the center frequency. In this embodiment, the lower limit of the passband frequency is configured as the dominant natural frequency minus the set frequency tolerance margin, and the upper limit of the passband frequency is configured as the dominant natural frequency plus the set frequency tolerance margin. In this embodiment, the value range of the set frequency tolerance margin is 0.2 to 1.0 Hz. The phase difference extraction module constructs a digital Butterworth filter of a set order that covers the dominant response frequency band of the structure based on the lower limit and upper limit of the passband frequency. In this embodiment, the set order ranges from 2 to 8. The phase difference extraction module performs zero-padding extension processing at the beginning and end of the preprocessed sequence for a set length. In this embodiment, the set length ranges from 100 to 500. The phase difference extraction module inputs the extended preprocessed sequence forward into the digital Butterworth filter to obtain the forward filtered sequence. The phase difference extraction module reverses the forward filtered sequence in the time dimension and inputs it again into the digital Butterworth filter to obtain the reverse filtered sequence.

[0026] S22, the phase difference extraction module performs time reversal reconstruction on the inverse filtered sequence to output the dominant response sequence. By superimposing forward and backward filtering, the phase lag caused by forward filtering and the phase lead caused by inverse filtering are canceled, thereby achieving zero-group delay alignment of the dominant response sequence on the time axis.

[0027] S23, The phase difference extraction module performs a discrete Hilbert transform on the dominant response sequence to obtain the orthogonal signal sequence corresponding to the dominant response sequence; The phase difference extraction module takes the orthogonal signal sequence as the ordinate input parameter and the dominant response sequence as the abscissa input parameter, and substitutes them into the four-quadrant arctangent function to calculate the argument in the complex plane. The phase difference extraction module uses the argument as the wrapping phase sequence, and the calculation formula for wrapping the phase sequence is as follows: ; In the formula, For the first The package phase value calculated by each sensor node at the current discrete time sampling point; For bivariate arctangent operators covering all four quadrants; For the first The orthogonal imaginary component values ​​obtained by processing the preprocessed sequence corresponding to each sensor node through Hilbert transform; For the first The real component values ​​of the preprocessed sequence corresponding to each sensor node at the current discrete-time sampling point; This is a spatial index for sensor node numbers whose values ​​are positive integers. It is the index of the discrete-time sampling point permutation sequence with values ​​ranging from positive integers.

[0028] S24, The phase difference extraction module calculates the package phase difference between the current discrete time sampling point and the previous discrete time sampling point in the package phase sequence; The phase difference extraction module establishes a step compensation integer with an initial value of zero. When the phase difference of the packaged phase sequence is less than the negative pi constant, the phase difference extraction module determines that a positive step change has occurred in the packaged phase sequence and controls the step compensation integer to be incremented by one. When the phase difference of the packaged phase sequence is greater than the positive pi constant, the phase difference extraction module determines that a negative step change has occurred in the packaged phase sequence and controls the step compensation integer to be decremented by one.

[0029] S25, the phase difference extraction module extracts data from two adjacent monitoring nodes in the spatial topology under the same discrete-time sampling point index. The phase difference extraction module performs a point-to-point subtraction operation between the absolute phase data of the current monitoring node containing step compensation and the absolute phase data of the next adjacent monitoring node containing step compensation. The phase difference extraction module outputs a transient phase difference sequence. The comprehensive calculation formula for the transient phase difference sequence is as follows: ; In the formula, This represents the transient phase difference value between two adjacent sensor nodes at the current discrete-time sampling point; For the first The wrap-around phase value of each sensor node at the current discrete-time sampling point; For the first The step compensation integer corresponding to each sensor node at the current discrete-time sampling point; For the first The wrap-around phase value of each sensor node at the current discrete-time sampling point; For the first The step compensation integer corresponding to each sensor node at the current discrete-time sampling point; This is a spatial index for the basic sensor node number, with values ​​ranging from positive integers. Spatial index for the reference sensor node numbers adjacent to the base sensor node; It is the index of the discrete-time sampling point permutation sequence with values ​​ranging from positive integers.

[0030] See appendix Figure 4 In specific implementations, step S3 provided by the present invention may include the following steps: S31, the dissipation feature extraction module performs moving average filtering on the dynamic strain sequence synchronously output by the physical quantity acquisition module. In this embodiment, the moving average filtering window length ranges from 5 to 15. The dissipative feature extraction module starts from the second discrete-time sampling point of the dynamic strain sequence after the moving average filtering process and performs a backward difference operation with the sampling point of the previous time step to obtain the discrete dynamic strain rate. The dissipation feature extraction module retrieves the historical maximum strain rate corresponding to random environmental excitation in the vehicle-free state and multiplies it by a set safety tolerance coefficient as the trigger threshold. In this embodiment, the set safety tolerance coefficient ranges from 1.5 to 3.0. The dissipation feature extraction module generates a trigger signal when the absolute value of the discrete dynamic strain rate first exceeds the trigger threshold. Starting from the discrete time sampling point corresponding to the trigger signal, a data window with a length that is an integer multiple of the natural vibration period of the target structure foundation is extracted. In this embodiment, the multiple of the natural vibration period of the foundation is in the range of 5 to 10. The dissipative feature extraction module extracts the truncated dynamic strain sequence and the time-synchronized truncated transient phase difference sequence within the data window, and locates the effective data segment containing the high signal-to-noise ratio structural physical response.

[0031] S32, the dissipative feature extraction module makes a one-to-one correspondence between the truncated dynamic strain sequence and the truncated transient phase difference sequence according to the discrete time sampling point index, and uses the values ​​in the truncated dynamic strain sequence as the abscissa of the two-dimensional Cartesian coordinate system and the values ​​in the truncated transient phase difference sequence as the ordinate of the two-dimensional Cartesian coordinate system. The dissipative feature extraction module maps and generates a set of transient phase space trajectories in a two-dimensional Cartesian coordinate system, which consists of multiple discrete coordinate points connected in chronological order. The dissipative feature extraction module extracts the first and last discrete coordinate points in the transient phase space trajectory set, and constructs a virtual linear interpolation line segment between the first and last discrete coordinate points. The dissipative feature extraction module supplements the interpolation coordinate points contained in the virtual linear interpolation line segment into the transient phase space trajectory set according to the spatial spacing, and outputs a forced closed trajectory set to eliminate the discontinuity caused by data window truncation.

[0032] S33, the dissipative feature extraction module performs polygon cross multiplication and summation operations on all coordinate points contained in the forced closed trajectory set according to the index order based on the discrete shoelace formula to obtain the polygon envelope area; The dissipative feature extraction module traverses the set of forced closed trajectories to extract the maximum and minimum values ​​of the truncated dynamic strain values ​​and the maximum and minimum values ​​of the truncated transient phase difference values ​​respectively; The dissipation feature extraction module calculates the difference between the maximum and minimum values ​​of the truncated dynamic strain values ​​to obtain the dynamic strain amplitude range, and calculates the difference between the maximum and minimum values ​​of the truncated transient phase difference fractions to obtain the phase amplitude range. The dissipation feature extraction module multiplies the strain amplitude range with the phase amplitude range to obtain the area of ​​the circumscribed rectangle, calculates the ratio of the polygonal envelope area to the area of ​​the circumscribed rectangle, and outputs the hysteresis area ratio. The dimensionless area mapping blocks the numerical interference of the absolute fluctuation of external excitation energy on the extraction of internal stiffness degradation features of the structure. The simplified comprehensive calculation formula for the hysteresis area ratio is as follows: ; In the formula, The hysteresis area ratio value is calculated for the output of the dissipation feature extraction module; Connect the polygons with coordinates that are positive integers and traverse the index; The total number of discrete coordinate points contained in the set of forced closed trajectories; For the set of forced closed trajectories, the first The truncated dynamic strain values ​​corresponding to each coordinate point; For the set of forced closed trajectories, the first The truncated transient phase difference values ​​corresponding to each coordinate point; For the set of forced closed trajectories, the first The truncated dynamic strain values ​​corresponding to each coordinate point; For the set of forced closed trajectories, the first The truncated transient phase difference values ​​corresponding to each coordinate point; The maximum value of the truncated dynamic strain value in the forced closed trajectory set; It is the minimum value of the truncated dynamic strain value in the set of forced closed trajectories; The maximum value of the truncated transient phase difference fraction in the set of forced closed trajectories; It is the minimum value of the truncated transient phase difference fraction in the set of forced closed trajectories.

[0033] In specific implementations, step S4 provided by this invention may include the following steps: S41, the state assessment module extracts the ambient temperature sequence output by the physical quantity acquisition module in time synchronization and calculates the arithmetic mean temperature value of the discrete temperature data points in the corresponding data window. The condition assessment module retrieves the baseline hysteresis area ratio data corresponding to different temperature conditions covering a complete natural year temperature cycle during the initial non-destructive service phase of the target structure, and uses the least squares fitting algorithm to establish a linear temperature mapping baseline function of the baseline hysteresis area ratio as a function of the arithmetic mean temperature value. The condition assessment module substitutes the arithmetic mean temperature value into the linear temperature mapping baseline function to calculate the expected area ratio reference value under the current temperature conditions. It subtracts the expected area ratio reference value from the hysteresis area ratio and outputs the temperature decoupling hysteresis area ratio. Through algebraic subtraction, it separates the characteristic data reflecting the internal physical damage of the structural material from the mixed response characteristics.

[0034] S42, the state evaluation module constructs a multilayer perceptron classification model containing an input layer, a first hidden layer, a second hidden layer and an output layer. It applies the maximum-minimum normalization algorithm to map the temperature decoupling hysteresis area ratio to the zero-to-one interval to reconstruct it into a single-dimensional input feature and feed it into the input layer of the multilayer perceptron classification model. The input layer passes the single-dimensional input features to the first hidden layer configured with a first set number of neuron nodes for feature mapping. In this embodiment, the first set number ranges from 16 to 64. A linear rectified activation function is applied inside the neuron nodes of the first hidden layer to output the first feature tensor. The first hidden layer passes the first feature tensor to the second hidden layer configured with a second set number of neuron nodes. In this embodiment, the value of the second set number ranges from 8 to 32. The second hidden layer applies a linear rectified activation function to generate the second feature tensor and passes it to the output layer configured with three neuron nodes. The state assessment module uses a normalized exponential function to calculate and output a three-dimensional state probability vector at the output layer. The three-dimensional elements of the three-dimensional state probability vector strictly correspond to the actual physical classification levels of the target structure's undamaged healthy state, slightly stiffness degradation state, and severely damaged state.

[0035] S43, the state assessment module retrieves historical measured abnormal data and finite element simulation analysis data of the target structure, and mixes them according to a set data mixing ratio to construct an offline training sample set. In this embodiment, the set data mixing ratio ranges from 1:2 to 1:5. The state assessment module strictly binds a uniquely encoded real label corresponding to the specific real health physical classification level to a single data point in the offline training sample set. The offline training sample set is then input into the multilayer perceptron classification model via forward propagation to obtain the predicted probability distribution vector. The state evaluation module calculates the loss gradient based on the adaptive moment estimation optimization algorithm with a set learning rate and the backpropagation mechanism, and iteratively updates the network weight matrix and bias vector of the multilayer perceptron classification model until the cross-entropy loss function value reaches the set convergence limit. In this embodiment, the learning rate is set to a range of 0.0001 to 0.005, and the convergence limit is set to a range of 0.001 to 0.01. The formula for calculating the cross-entropy loss function value is as follows: ; In the formula, The cross-entropy loss function value is calculated for the output of a single training batch. The total number of categories for the established true physical health classification levels; The index for traversing the true health physical classification level is a positive integer. In this embodiment, the total number of true health physical classification level categories is 3. The true probability of the one-hot encoded ground label of the offline training sample set under the current classification category traversal index; It is a logarithmic operator with the natural constant as its base; This represents the predicted probability distribution of the state probability vector output by the forward inference of the multilayer perceptron classification model under the current classification category traversal index.

[0036] S44, the state assessment module will input the temperature decoupling hysteresis area ratio, which is obtained online in real time and processed by the maximum-minimum normalization algorithm, into the multilayer perceptron classification model that has completed parameter iterative training and perform a single forward inference calculation. The state assessment module extracts the dimension space index corresponding to the largest probability element in the three-dimensional state probability vector output by the multilayer perceptron classification model, and decodes and maps the dimension space index to the current true physical health classification level of the target structure. When the actual health physical classification level indicates a state of slight stiffness degradation or severe structural damage, the status assessment module generates an abnormal alarm message with a specific discrete timestamp and spatial location node number, and pushes the abnormal alarm message to the remote monitoring terminal device held by the operation and maintenance personnel through the wireless communication bus.

[0037] See appendix Figure 5 In specific implementations, step S5 provided by the present invention may include the following steps: S51, the life prediction and model update module extracts the temperature decoupling hysteresis area ratio sequence continuously output by the condition assessment module during the historical service cycle; The temperature decoupling hysteresis area ratio sequence with irregular acquisition time is resampled into equidistant time-series feature data with a fixed time interval and a set step size by applying a first-order linear interpolation algorithm. In this embodiment, the set step size ranges from 1 to 6 hours. The lifespan prediction and model update module performs sliding truncation on equidistant time series feature data at a fixed daily time step to generate time series sample data in three-dimensional tensor format. The length of the sliding truncation data window is configured to cover the number of days of at least one complete natural year cycle. In this embodiment, the length of the sliding truncation data window ranges from 365 to 1095 days. By eliminating the impact of irregular timestamps on underlying tensor operations through resampling and fixed-length truncation, time-series sample data can have a temporal receptive field for extracting long-term degradation features.

[0038] S52, the lifetime prediction and model update module constructs a long short-term memory neural network model containing an input layer, a hidden layer and a fully connected output layer, and feeds time series sample data in three-dimensional tensor format into the input layer of the long short-term memory neural network model; The input layer passes the time-series sample data to the hidden layer of the Long Short-Term Memory (LSTM) with a configured number of units. In this embodiment, the value of the configured number of units ranges from 32 to 128. The forget gate inside the hidden layer determines the retention ratio of historical cell state information through the logistic activation function, the input gate generates the candidate cell state at the current time through the hyperbolic tangent activation function and the logistic activation function, and the output gate generates the hidden layer feature tensor based on the updated cell state. The hidden layer feature tensor is passed to the fully connected output layer and mapped to the predicted remaining service days, which represents the remaining healthy service days of the target structure.

[0039] S53, the life prediction and model update module collects accelerated fatigue test degradation data and finite element simulation data to construct an offline prediction sample set; The offline prediction sample set is labeled with continuous values ​​that decrease linearly from the initial total service life to zero as the true service life label. The offline prediction sample set is then input into the long short-term memory neural network model in batches to calculate the predicted remaining service life. The gradient is calculated based on the mean squared error loss function combined with the backpropagation mechanism with a set learning rate. The network weight matrix is ​​iteratively updated until the mean squared error loss function value reaches the set convergence limit. In this embodiment, the learning rate is set to a range of 0.0001 to 0.005, and the convergence limit is set to a range of 0.0001 to 0.001. The comprehensive calculation formula for the mean squared error loss function is as follows: ; In the formula, The mean squared error loss function value is calculated within a single training batch. The total number of time-series samples included in a single iteration training batch is 32 to 128 in this embodiment. Summing and traversing the index for time-series samples with values ​​in the range of positive integers; For the offline prediction sample set, the first The actual lifetime label value corresponding to each sample; The output of the forward inference calculation for the Long Short-Term Memory Neural Network Model The predicted remaining service days for each sample.

[0040] S54, the life prediction and model update module compares the predicted remaining service days obtained by online real-time calculation with the safe service days limit. In this embodiment, the safe service days limit ranges from 30 to 90 days. When the life prediction and model update module predicts that the remaining service days are less than or equal to the safe service days limit, it issues a mandatory maintenance command to the remote monitoring terminal and activates the cloud-based digital twin model update mechanism. The life prediction and model update module uses the Kalman filter algorithm to fuse real-time feature data to iteratively correct the stiffness matrix parameters of the cloud-based digital twin model; The local finite element stiffness reduction coefficients corresponding to the damaged entity components in the cloud-based digital twin model are used as the state variables of the Kalman filter, and the state transition matrix is ​​configured as the identity matrix to construct the random walk state equation without external input control. The temperature decoupling hysteresis area ratio continuously output by the state assessment module is used as the observation variable for Kalman filtering; The cloud-based digital twin model is controlled to perform forward dynamic calculations under the current parameters to obtain the theoretical hysteresis area ratio. The sensitivity observation matrix is ​​constructed based on the partial derivatives of the theoretical hysteresis area ratio with respect to the stiffness reduction coefficients of each local finite element. The observation residuals of the actual temperature decoupling hysteresis area ratio and the theoretical hysteresis area ratio are calculated online. The Kalman gain is calculated based on the sensitivity observation matrix and the prediction error covariance. The local finite element stiffness reduction coefficient is updated a posteriori by multiplying the Kalman gain by the observation residual. The updated and converged local finite element stiffness reduction coefficients are multiplied into the initial element stiffness matrix of the cloud-based digital twin model to achieve twin synchronization between the model's physical parameters and the actual state of the structure.

[0041] In specific implementations, step S6 provided by the present invention may include the following steps: S61, the early warning decision and visualization module extracts the real health physical classification level and predicts the remaining service days, and configures the corresponding discrete severity weight coefficients for the undamaged health state, the slight stiffness degradation state and the severe structural damage state respectively. In this embodiment, the discrete severity weight coefficient corresponding to the undamaged health state is 0.1, the discrete severity weight coefficient corresponding to the slight stiffness degradation state is 0.5, and the discrete severity weight coefficient corresponding to the severe structural damage state is 0.9. The early warning decision and visualization module calculates the difference between the set baseline expected lifespan and the predicted remaining service days and divides it by the set baseline expected lifespan to obtain the lifespan decay rate. In this embodiment, the set baseline expected lifespan ranges from 18,250 to 36,500 days. The comprehensive maintenance priority index is output by weighting and summing the discrete severity weighting coefficient and the lifespan decay rate. The specific calculation formula for the comprehensive maintenance priority index is as follows: ; In the formula, The comprehensive maintenance priority index is calculated and output by the early warning decision-making and visualization module; In this embodiment, the dimensionless adjustment factor, which is set to balance the weight of physical classification level and the proportion of lifetime decay, ranges from 0.4 to 0.6. The severity weight coefficients are obtained by discretizing and mapping the actual health physical classification levels. The baseline expected lifespan in days set for the target structure in the design documents; The predicted remaining service days are calculated and output by the forward inference of the long short-term memory neural network model.

[0042] S62, the early warning decision and visualization module establishes a three-level response trigger interval corresponding to the comprehensive maintenance priority index. In this embodiment, the three-level response trigger interval includes a low-risk monitoring interval with a value range of 0.0 to 0.4, a medium-risk early warning interval with a value range of 0.4 to 0.7, and a highest response trigger interval with a value range of 0.7 to 1.0. The early warning decision and visualization module generates an emergency dispatch instruction when the comprehensive maintenance priority index falls into the highest response trigger range, and uses the spatial location node number specified in the abnormal alarm message as the central rendering anchor point. The local finite element mesh size of the cloud-based digital twin model is extracted and multiplied by a set scaling factor as a spatial influence attenuation parameter. In this embodiment, the value range of the scaling factor is 1.5 to 3.0. In a three-dimensional spatial coordinate system, a Gaussian radial basis function containing spatial influence attenuation parameters is applied to calculate the stress concentration influence weight of the area surrounding the central rendering anchor point. Based on the stress concentration influence weight, a damage heat map texture with a nonlinear color transition from red to yellow is generated and mapped to the surface of the corresponding solid component in the cloud digital twin model. The global spatial correlation of discrete node alarms is established through three-dimensional spatial Gaussian rendering.

[0043] S63, the early warning decision and visualization module will push the cloud-based digital twin model, which integrates damage heat map texture, to the remote monitoring terminal through the graphics library interface; The early warning decision and visualization module automatically fills in the specific discrete timestamp and spatial location node number contained in the abnormal alarm message in the emergency dispatch instruction, and sends the emergency dispatch instruction to the mobile smart terminal bound to the operation and maintenance personnel through the application layer communication protocol.

[0044] See appendix Figure 6 To aid in understanding the technical solution of this invention, an application embodiment of an intelligent early warning system for structural safety risks in civil engineering is provided below: For long-span highway steel box girder continuous bridges under the dual effects of high temperature and heavy traffic, a sensor network containing accelerometers, strain gauges and temperature sensors is deployed in the mid-span and quarter-span load-bearing areas of the bridge to perform monitoring operations. The system's underlying layer is based on a fiber optic network and a precise time protocol to control all sensors to add fully synchronized nanosecond-level timestamps when trucks pass over the bridge deck. The system performs mean and trend removal processing to remove static offsets and long-term drift trends from the data and extract dynamic fluctuation data.

[0045] The system filters out the dominant response sequence corresponding to the dominant natural frequency of the target structure through a bandpass filter. The system performs discrete Hilbert transform and phase dewinding processing on the dominant response sequence. The system calculates the transient phase difference sequence between the mid-span monitoring node and the adjacent monitoring nodes to characterize the phase hysteresis feature of vibration transmission.

[0046] When the system detects that the dynamic strain rate of the mid-span node exceeds the trigger threshold corresponding to the set safety tolerance coefficient, it extracts a data window containing integer multiples of the natural vibration period of the target structure foundation. The system uses the truncated dynamic strain sequence within the data window as the abscissa of the two-dimensional Cartesian coordinate system and the truncated transient phase difference sequence as the ordinate. The system generates a set of transient phase space trajectories in the two-dimensional Cartesian coordinate system and constructs a set of forced closed trajectories through virtual linear interpolation. The system calculates the polygonal envelope area of ​​the forced closed trajectory set based on the discrete shoelace formula and divides it by the area of ​​the circumscribed rectangle to output the hysteresis area ratio.

[0047] The system substitutes the arithmetic mean temperature value of discrete temperature data points within the current data window into the linear temperature mapping baseline function established using historical lossless period data to calculate the expected area ratio reference value. The system subtracts the expected area ratio reference value from the hysteresis area ratio to obtain the temperature decoupling hysteresis area ratio. The system inputs the temperature decoupling hysteresis area ratio processed by the max-min normalization algorithm into the multilayer perceptron classification model that has completed parameter iteration training to perform forward inference calculation. The system extracts the dimension space index corresponding to the largest probability element in the state probability vector output by the multilayer perceptron classification model and decodes and maps it to output the current true health physical classification level of the target structure.

[0048] The system resamples the temperature decoupling hysteresis area ratio sequence continuously output during the historical service cycle and reconstructs it into time-series sample data in three-dimensional tensor format by truncating it at a fixed time step. The system feeds the time-series sample data into a long short-term memory neural network model to calculate and obtain the predicted remaining service days. When the predicted remaining service days are less than or equal to the safe service days limit, the system issues a mandatory maintenance command to the remote monitoring terminal. The system uses a Kalman filter algorithm to fuse the observed residuals of the temperature decoupling hysteresis area ratio obtained online and the theoretical hysteresis area ratio obtained by forward dynamic calculation to perform a posteriori iterative update of the local finite element stiffness reduction coefficient of the cloud digital twin model.

[0049] The system extracts the actual health physical classification level and the predicted remaining service days, and calculates the comprehensive maintenance priority index by combining it with the set discrete severity weight coefficient. When the comprehensive maintenance priority index falls into the highest response trigger range, the system uses the spatial location node number specified in the abnormal alarm message as the central rendering anchor point. The system applies a Gaussian radial basis function containing spatial impact attenuation parameters to render damage heat map textures on the surface of the corresponding entity component in the cloud digital twin model. The system pushes the cloud digital twin model with the integrated damage heat map textures to the remote monitoring terminal and automatically fills in the specific discrete timestamp and spatial location node number in the emergency dispatch instruction and sends it to the mobile smart terminal.

[0050] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result using substantially the same method falls within the scope of the present invention.

Claims

1. A smart early warning system for structural safety risks in civil engineering, characterized in that, The system includes: The physical quantity acquisition module acquires the high-frequency acceleration sequence, dynamic strain sequence, and ambient temperature sequence of the structural monitoring nodes; The phase difference extraction module extracts the transient phase difference sequence between adjacent monitoring nodes based on the high-frequency acceleration sequence; The dissipation feature extraction module extracts a data window when the dynamic strain rate of the dynamic strain sequence exceeds the product threshold of the historical maximum strain rate and the safety tolerance coefficient corresponding to the random environmental excitation. It then combines the dynamic strain sequence within the data window with the transient phase difference sequence to construct a two-dimensional phase space discrete trajectory and calculate the hysteresis area ratio. The status assessment module decouples the hysteresis area ratio into a temperature-decoupled hysteresis area ratio by combining the ambient temperature sequence, inputs it into the multilayer perceptron classification model for decoding and mapping, and outputs the true health physical classification level. The life prediction and model update module feeds the temperature decoupling hysteresis area ratio into the long short-term memory neural network model to obtain the predicted remaining service days. When the predicted remaining service days reach the safe service days limit, the Kalman filter algorithm is used to iteratively correct the stiffness matrix parameters of the cloud digital twin model. The early warning decision and visualization module calculates a comprehensive maintenance priority index based on the actual health physical classification level and the predicted remaining service days. When the comprehensive maintenance priority index falls into the highest response trigger range, it renders a damage heat map texture on the surface of the corresponding physical component in the cloud digital twin model.

2. The intelligent early warning system for structural safety risks in civil engineering according to claim 1, characterized in that, The physical quantity acquisition module acquires the high-frequency acceleration sequence, dynamic strain sequence, and ambient temperature sequence of the structural monitoring node; The physical quantity acquisition module performs mean and trend removal processing on the high-frequency acceleration sequence, the dynamic strain sequence, and the ambient temperature sequence, respectively deducting the arithmetic mean of the corresponding sequence and the linear trend term data obtained by least squares fitting, and outputting the preprocessed sequence.

3. The intelligent early warning system for structural safety risks in civil engineering according to claim 1, characterized in that, The phase difference extraction module constructs a digital Butterworth filter by setting the lower and upper limits of the passband frequency based on the dominant natural frequency of the target structure. The phase difference extraction module inputs the high-frequency acceleration sequence into the digital Butterworth filter for forward filtering and reverse filtering respectively. The phase deviation between forward filtering and reverse filtering is canceled by time reversal reconstruction and superposition to achieve zero group delay alignment and output the dominant response sequence. The phase difference extraction module performs a discrete Hilbert transform on the dominant response sequence to obtain an orthogonal signal sequence, and calculates the wrapping phase sequence by substituting the dominant response sequence and the orthogonal signal sequence into the four-quadrant arctangent function. The phase difference extraction module monitors the wrapped phase difference value of the wrapped phase sequence, controls the step compensation integer to perform a step compensation operation to obtain absolute phase data, and performs a subtraction operation on the absolute phase data of adjacent monitoring nodes at the same discrete time sampling point to output the transient phase difference sequence.

4. The intelligent early warning system for structural safety risks in civil engineering according to claim 1, characterized in that, The dissipative feature extraction module performs a difference operation on the dynamic strain sequence after the moving average filtering process to obtain the discrete dynamic strain rate as the dynamic strain rate. When the discrete dynamic strain rate exceeds the product threshold, the dissipation feature extraction module generates a trigger signal and uses the discrete time sampling point corresponding to the trigger signal as the starting point to extract the data window that is an integer multiple of the natural vibration period of the target structure foundation. The dissipative feature extraction module uses the dynamic strain sequence within the data window as the abscissa and the corresponding transient phase difference sequence as the ordinate in a two-dimensional Cartesian coordinate system to generate a set of transient phase space trajectories. The dissipative feature extraction module converts the transient phase space trajectory set into a forced closed trajectory set by supplementing linear interpolation line segments. It performs polygon cross multiplication and summation operations on the coordinate points contained in the forced closed trajectory set to obtain the polygon envelope area. It calculates the polygon envelope area and divides it by the area of ​​the circumscribed rectangle corresponding to the forced closed trajectory set, and outputs the hysteresis area ratio.

5. The intelligent early warning system for structural safety risks in civil engineering according to claim 1, characterized in that, The state assessment module extracts the ambient temperature sequence and calculates the arithmetic mean temperature value of discrete temperature data points within the data window. The condition assessment module substitutes the arithmetic mean temperature value into the linear temperature mapping baseline function established using non-destructive initial service phase data combined with the least squares fitting algorithm to calculate the expected area ratio reference value under the corresponding temperature conditions. The state assessment module subtracts the expected area ratio reference value from the hysteresis area ratio and outputs the temperature decoupling hysteresis area ratio.

6. The intelligent early warning system for structural safety risks in civil engineering according to claim 1, characterized in that, The state assessment module mixes historical measured abnormal data with finite element simulation analysis data to construct an offline training sample set containing one-hot encoded real labels. Based on the cross-entropy loss function value and the adaptive moment estimation optimization algorithm combined with the backpropagation mechanism, iteratively updates the network weight matrix and bias vector of the multilayer perceptron classification model. The state evaluation module processes and reconstructs the temperature decoupling hysteresis area ratio using the maximum-minimum normalization algorithm into a single-dimensional input feature input in the multilayer perceptron classification model that completes parameter iterative training. In the output layer, it calculates and outputs a three-dimensional state probability vector. The state assessment module extracts the dimensional space index corresponding to the probability element with the largest value in the three-dimensional state probability vector, and decodes and maps the dimensional space index to the real health physical classification level.

7. The intelligent early warning system for structural safety risks in civil engineering according to claim 1, characterized in that, The lifetime prediction and model update module resamples the temperature decoupling hysteresis area ratio continuously output from the historical cycle into equidistant time series feature data with fixed time intervals using a first-order linear interpolation algorithm. The equidistant time series feature data is then truncated by sliding at a fixed time step to generate time series sample data in three-dimensional tensor format. The life prediction and model update module mixes the collected accelerated fatigue test degradation data with finite element simulation data to construct an offline prediction sample set labeled with continuous numerical values ​​of real life, and iteratively updates the network weight matrix of the long short-term memory neural network model based on the mean square error loss function combined with the backpropagation mechanism. The lifespan prediction and model update module inputs the time-series sample data in the three-dimensional tensor format into the trained long short-term memory neural network model to calculate and output the predicted remaining service days.

8. The intelligent early warning system for structural safety risks in civil engineering according to claim 1, characterized in that, When the predicted remaining service days are less than or equal to the safe service days limit, the life prediction and model update module uses the local finite element stiffness reduction coefficient corresponding to the target damaged entity component in the cloud digital twin model as the state variable of the Kalman filter, and uses the temperature decoupling hysteresis area ratio obtained online in real time as the observation variable of the Kalman filter. The lifetime prediction and model update module controls the cloud-based digital twin model to perform forward dynamics calculations to obtain the theoretical hysteresis area ratio. Based on the observation residuals of the online real-time temperature decoupling hysteresis area ratio and the theoretical hysteresis area ratio, as well as the partial derivative of the theoretical hysteresis area ratio, a sensitivity observation matrix is ​​constructed. The lifetime prediction and model update module calculates the Kalman gain based on the sensitivity observation matrix and the prediction error covariance, and uses the Kalman gain multiplied by the observation residual to perform a posteriori iterative update of the local finite element stiffness reduction coefficient. The updated and converged local finite element stiffness reduction coefficient is multiplied into the initial element stiffness matrix of the cloud digital twin model to correct the stiffness matrix parameters.

9. The intelligent early warning system for structural safety risks in civil engineering according to claim 1, characterized in that, The early warning decision and visualization module establishes corresponding discrete severity weight coefficients for different real health physical classification levels, and extracts the matching discrete severity weight coefficients based on the real health physical classification levels. The early warning decision and visualization module calculates the difference between the design baseline expected life and the predicted remaining service days and divides it by the design baseline expected life to obtain the life decay rate. The early warning decision and visualization module performs a weighted summation operation on the discrete severity weight coefficient and the life decay rate to output the comprehensive maintenance priority index.

10. The intelligent early warning system for structural safety risks in civil engineering according to claim 1, characterized in that, When the comprehensive maintenance priority index falls into the highest response trigger range, the early warning decision and visualization module generates an emergency dispatch instruction and an abnormal alarm message, and uses the spatial location node number specified in the abnormal alarm message as the central rendering anchor point. The early warning decision and visualization module extracts the local finite element mesh size of the cloud-based digital twin model and multiplies it by a mesh scaling factor to obtain the spatial influence attenuation parameter. The early warning decision and visualization module uses a Gaussian radial basis function containing the spatial influence attenuation parameter to calculate the stress concentration influence weight of the area surrounding the central rendering anchor point. Based on the stress concentration influence weight, it generates the damage heat map texture that transitions non-linearly from red to yellow and maps it to the surface of the corresponding physical component of the cloud digital twin model for rendering.

Images