A structural health state intelligent evaluation method and system

Abstract

The application discloses a structural health state intelligent evaluation method and system, and the method comprises the following steps: S1, acquiring a multi-source time sequence signal and performing a multi-stage preprocessing operation on the signal; S2, constructing a structured loss function, separating a temperature environment reversible response from a structural irreversible deterioration trend through a constraint framework, and outputting a purified structural trend item; S3, constructing three types of core indexes of a temperature environment feature, a vibration modal feature and a multi-physical quantity coupling correlation feature, and forming a standardized multi-dimensional feature vector; S4, establishing an RBF-SVM multi-classification model, and acquiring an optimal hyperparameter; S5, inputting a new feature vector into the trained model, determining a grade through model voting, outputting a highest-vote winner, and making a decision according to a priority or a confidence in the case of a tie, and outputting a final evaluation result. The application further discloses a structural health state intelligent evaluation system. The application can realize full-dimension, high-precision and fine intelligent health evaluation of infrastructure structures.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring, and in particular to an intelligent method and system for assessing structural health status. Background Technology

[0002] With the large-scale and networked development of my country's transportation infrastructure, a large number of highway and railway bridges have been operating under high load and complex environmental conditions for a long time. Problems such as structural aging, component damage, and performance degradation are becoming increasingly prominent, and the pressure on bridge safety operation and maintenance continues to increase. Bridge structural health status assessment is a core technical means to ensure traffic operation safety, achieve refined bridge maintenance, and avoid the risk of overturning and collapse. The accuracy and refinement of the assessment results directly determine the scientific nature and safety of bridge operation and maintenance decisions.

[0003] Traditional bridge structural health assessment technologies suffer from numerous industry pain points, severely hindering the application of intelligent bridge monitoring technologies. Firstly, it's difficult to distinguish between environmental interference and structural deterioration characteristics. Bridge monitoring signals are greatly affected by diurnal temperature variations, seasonal changes, and external environmental disturbances. Traditional methods cannot effectively separate reversible environmental responses from irreversible structural deterioration trends, easily leading to misjudgments and omissions. Secondly, the feature dimensions are limited. Existing assessments often rely on single deformation or vibration characteristics, failing to cover multi-dimensional damage states such as hidden stiffness damage, localized coupled defects, and overall coordinated anomalies, resulting in insufficient feature representation capabilities. Thirdly, they have poor adaptability to small sample conditions. Bridge defect samples are characterized by small sample sizes, nonlinearity, and high dimensionality. Traditional assessment models have weak generalization capabilities, significantly reducing classification accuracy under small sample conditions. Fourthly, the risk assessment hierarchy is ambiguous. There is a lack of a penetrating hierarchical assessment system encompassing components, sections, and the entire structure, making it impossible to achieve precise defect localization and comprehensive risk assessment. Fifthly, there is a lack of specific risk quantification. Existing technologies have not established a dedicated quantitative assessment mechanism for the core risk of bridge overturning, resulting in strong subjectivity and low quantification accuracy in overturning risk assessment.

[0004] Existing bridge structural health monitoring and assessment methods mostly employ single physical quantity time-series comparisons, fixed threshold discrimination, and manual experience-based ratings. These methods have simple data preprocessing mechanisms and lack anti-interference mechanisms for multi-sensor lateral comparisons and longitudinal time-series baseline comparisons, failing to eliminate data anomalies caused by individual sensor biases and local environmental interference. Furthermore, they lack the ability to extract coupled features from multiple physical quantities, making it difficult to identify latent defects induced by temperature-vibration coupling and structural synergistic anomalies. Conventional machine learning assessment models often use fixed parameters and static feature inputs, failing to adapt to the nonlinear degradation patterns under complex bridge conditions, resulting in low classification accuracy and poor robustness in small sample scenarios. Moreover, existing technologies can only achieve a coarse overall bridge rating, lacking refined component-level discrimination, hierarchical risk aggregation, and a specific quantitative system for overturning, thus failing to meet the needs of modern bridge's refined, intelligent, and hierarchical safety operation and maintenance.

[0005] In summary, this invention urgently needs to develop a bridge intelligent health assessment method that integrates multiple features, has strong anti-interference capabilities, good small sample adaptability, hierarchical full coverage, and quantifiable overturning risk. Summary of the Invention

[0006] The purpose of this invention is to address the technical shortcomings of traditional bridge structural health monitoring, such as the inability to isolate environmental interference, single feature dimensions, low classification accuracy with small samples, ambiguous risk levels, and lack of quantification of overturning risk. This invention provides an intelligent assessment method and system for structural health status, which achieves comprehensive, high-precision, and refined intelligent health assessment of bridge structures through a complete technical process including multi-source signal preprocessing, structural trend term purification, multi-dimensional coupled feature construction, SVM multi-classification intelligent discrimination, and hierarchical risk assessment output.

[0007] The technical solution of this invention is: One of the present inventions provides an intelligent assessment method for structural health status, comprising the following steps: S1: Acquire multi-source time-series signals, and perform multi-level preprocessing operations on the multi-source time-series signals based on the spatial lateral comparison between the responses of similar sensors and the temporal longitudinal comparison of the responses of each measuring point relative to the historical baseline, and output the preprocessed multi-source time-series signal set. S2: For the preprocessed multi-source time series signal set, construct a structured loss function and adopt a four-dimensional composite constraint framework that includes data fitting constraints, spatial gradient consistency constraints, temperature change rate gating constraints, and weight regularization constraints. Through the constraint framework, separate the reversible response of the temperature environment from the irreversible degradation trend of the structure and output the purified structural trend term. S3: Based on the purified structural trend term and the preprocessed multi-source time series signal set, construct three core indicators: temperature environment characteristics, vibration mode characteristics, and multi-physical quantity coupling correlation characteristics, and form a standardized multi-dimensional feature vector. S4: Based on the standardized multidimensional feature vectors and combined with pre-labeled structural health level labels, construct a training set for model training; establish an RBF-SVM multi-classification model, construct sub-classifiers between any two health levels, and adopt a strategy combining dynamic sparse search with decreasing step size and cross-validation, with the average classification accuracy of the validation set as the optimization objective. Within the dynamically shrinking candidate subset, collaboratively optimize the kernel parameters and penalty parameters of each sub-classifier to obtain the optimal hyperparameters; based on the optimal hyperparameters, retrain the RBF-SVM multi-classification model using the training set. S5: The RBF-SVM multi-classification model trained in S4 is used to predict the new input feature vector. The feature vector is input into all sub-classifiers, and each sub-classifier classifies it into one of two health levels and casts a vote for the classified level. The cumulative number of votes for each health level is counted. If there is a unique highest vote, the health level corresponding to that vote is taken as the judgment result. If there is a tie, the decision is made according to the preset level priority or the sub-classifier confidence. The final evaluation result is output.

[0008] Furthermore, in S1, the multi-level preprocessing operation includes: By combining the results of horizontal comparison with similar measurement points and vertical comparison with historical baselines, abnormal data are removed from the original multi-source time series signals. For the high-frequency monitoring signals after removing anomalies, an adjustable window length moving average filtering algorithm is used to filter out high-frequency noise. The filtered monitoring signals are calibrated with a unified timestamp to achieve time synchronization of multi-source signals. For monitoring signals that have short-term data gaps after time synchronization, linear interpolation or nearest neighbor filling methods are used to repair the missing data.

[0009] Furthermore, the comparison of similar horizontal measurement points is as follows: at the same sampling time t, the measurement points deployed in the infrastructure structure... The measured values ​​of all sensors of the same type are regarded as a spatial sample set X(t)={x1(t),x2(t),…,x n (t)}; where n is the total number of sensors of this type; calculate the median m(t) and the median absolute deviation MAD(t) of this sample set: ; ; Then the measuring point Lateral deviation at time t for: ; when If the measurement point is at that time, it is determined to be a lateral outlier; where, α is a very small positive number, and α is the lateral outlier threshold; The longitudinal historical baseline comparison refers to: for each independent measuring point Based on its infrastructure health status Establish a dynamic baseline based on historical data from the same period and calculate its mean. with standard deviation : ; ; Then the measuring point Longitudinal deviation at time t for: ; when When the measurement point is determined to be a longitudinal outlier at that moment; in, β represents the total number of valid samples during the healthy baseline period at this measurement point, and β is the longitudinal outlier threshold.

[0010] Furthermore, in S2, the structured loss function is obtained through the following formula: ; In the formula, λ1, λ2, and λ3 are regularization coefficients; The structured loss function consists of four terms: Data fitting terms: N is the total number of samples. For the sensor measured data, the first The actual monitoring value of each sample For the first The model predicted values ​​for each sample; where the actual monitored values ​​are... The values ​​are taken from the preprocessed multi-source time-series signal set in step S1 and are model predictions. The fitting target; Spatial gradient consistency constraint: ; For the residual vector, For spatial gradient operators; Temperature quiescent period time-series gradient penalty term: T represents the temperature field. The threshold value is used for gate control. This indicates that the model's predicted value is the partial derivative with respect to time. This represents the partial derivative of temperature with respect to time. This is an indicator function that takes the value 1 when the condition is true and 0 when the condition is false. Weight regularization term: ; These are all the trainable parameters of the model.

[0011] Furthermore, in S2, the purification of the structural trend term is obtained through the following formula: ; In the formula, Signal components in a preprocessed multi-source time-series signal set; reversible temperature response. For model predictions The component corresponding to the temperature-induced reversible structural response; in the structured loss function After optimization and convergence, the predicted values ​​output by the model It is broken down into two parts: ; in, It is a reversible structural response caused by changes in the temperature field, and it is periodic and recoverable; This indicates an irreversible trend of structural degradation.

[0012] Furthermore, in S3, the three types of core indicators specifically include: A. The temperature environment feature structure includes: A1. Temperature difference ΔT between structure and environment: ; A2. Lateral temperature non-uniformity coefficient δT cross : ; In the formula, T str The actual temperature of the object being measured, T env The ambient reference temperature; T L T R These are the temperatures at symmetrical measuring points on both sides of the object being measured. B. The vibration modal characteristics are constructed as follows: B1. Effective value of acceleration A rms : ; B2. Main frequency offset Δf: ; B3. Damping ratio change Δζ: ; B4. Vibrational energy spectral density change ΔPSD: ; In the formula, For discrete acceleration sample values, M is the number of sample points; f cur f is the first-order fundamental frequency of the object currently being tested. base ζ represents the reference frequency of the tested object under healthy conditions. cur ζ represents the current damping ratio of the object being measured. base The reference damping ratio of the measured object; PSD cur The vibration power spectral density (PSD) is the value of PSD under the current operating conditions. base The vibration power spectral density under healthy baseline operating conditions; C. The construction of multi-physical quantity coupling correlation features includes: C1. Temperature-vibration coupling coefficient : ; C2. Synchronization coefficient of lateral displacement of left and right carriages : ; In the formula, Let U be the temperature and vibration covariance, and D(T) and D(A) be the variances of the temperature and vibration data, respectively. L U R These are the average lateral displacements of the left and right sides of the measured object, respectively. The standardized multidimensional feature vector is: .

[0013] Furthermore, step S4 specifically includes: S41: Obtain the standardized multidimensional feature vector output from step S3 and the pre-labeled structural health level label corresponding to each sample. Associate each feature vector with its corresponding level label to form feature-label paired samples. Collect all paired samples to construct a training set. S42: Construct a one-to-one multi-class SVM model based on the RBF kernel, and build independent binary sub-classifiers between any two health levels; S43: A dynamic sparse search strategy with decreasing step size is adopted. In the first round, a coarse-grained traversal is performed in the global search space with a first preset step size to lock the high-potential region where the optimal parameters are located. In the second round, the high-potential region is used as the boundary, and a second preset step size smaller than the first preset step size is used for local fine-grained search. Combined with a cross-validation algorithm with a preset number of folds, the kernel parameters and penalty parameters of each sub-classifier are optimized in a coordinated manner with the average classification accuracy of the validation set as the optimization objective, and the optimal hyperparameters are output. S44: Based on the optimal hyperparameters, retrain all subclassifiers using the training set and save the model parameters.

[0014] Furthermore, the objective function used in the collaborative optimization is: ; ; In the formula, The search range for the penalty parameter C, The search range for the kernel function bandwidth parameter γ; Let V be the classification accuracy under the current parameter combination in the v-th fold cross-validation. is the complexity penalty coefficient; K is the cross-validation fold number.

[0015] Furthermore, the method for constructing the sub-classifier in step S4 and the voting mechanism in step S5 specifically include: The sub-classifiers are constructed as follows: For the four health levels—excellent, attentive, warning, and dangerous—corresponding to health level labels y∈{0,1,2,3}, an independent binary SVM sub-classifier is constructed between any two different health level categories p and q. A total of 6 sub-classifiers are constructed; each sub-classifier is trained using only training samples belonging to class p and class q, and its decision function is: ; in, For category With category The total number of training samples; These are the Lagrange multipliers obtained by optimizing the corresponding sub-classifiers; The labels are binary, with category p marked as +1 and category q marked as -1. For bias terms; For feature vectors With the sample to be predicted Kernel similarity between them; For a sign function, the value is +1 if the independent variable is greater than 0, and -1 if it is less than 0; Classified as ;like Classified as ; The voting method is as follows: For a newly input feature vector, it is sequentially input into all 6 sub-classifiers for voting. Each sub-classifier classifies it as category p or category q and casts a vote for the classified category. If there is a unique category with the highest vote, that category is taken as the health status judgment result. If there is a tie, the category with the higher risk level is selected as the judgment result according to the progressively higher risk level. If this rule still cannot distinguish, the category with the higher confidence level output by the corresponding sub-classifier decision function is selected as the judgment result.

[0016] An intelligent structural health status assessment system according to the present invention includes: The data acquisition and preprocessing module is used to acquire multi-source time-series signals and perform multi-level preprocessing operations on the multi-source time-series signals based on the spatial lateral comparison between the responses of similar sensors and the temporal longitudinal comparison of the responses of each measuring point relative to the historical baseline, and output the preprocessed multi-source time-series signal set. The structured loss function trend term extraction module is used to construct a structured loss function for the preprocessed multi-source time series signal set. It adopts a four-dimensional composite constraint framework that includes data fitting constraints, spatial gradient consistency constraints, temperature change rate gating constraints, and weight regularization constraints. Through the constraint framework, the reversible response of the temperature environment and the irreversible degradation trend of the structure are separated, and the purified structured trend term is output. The multidimensional coupling feature construction module is used to construct three core indicators—temperature environment features, vibration mode features, and multi-physical quantity coupling correlation features—based on the purified structural trend term and the preprocessed multi-source time series signal set, forming a standardized multidimensional feature vector. The SVM intelligent discrimination module is used to construct a training set for model training based on the standardized multidimensional feature vector and pre-labeled structural health level labels; establish an RBF-SVM multi-classification model, construct sub-classifiers between any two health levels, and adopt a strategy combining dynamic sparse search with decreasing step size and cross-validation, with the average classification accuracy of the validation set as the optimization objective. Within the dynamically shrinking candidate subset, the kernel parameters and penalty parameters of each sub-classifier are jointly optimized to obtain the optimal hyperparameters; based on the optimal hyperparameters, the RBF-SVM multi-classification model is retrained using the training set. The hierarchical risk assessment module is used to input the new input feature vector into all sub-classifiers of the trained RBF-SVM multi-classification model. Each sub-classifier uses the feature vector as input to determine the level, classifying it into one of two health levels and casting a vote for that level. The module counts the cumulative number of votes for each health level. If there is a unique highest vote, the health level corresponding to that vote is taken as the judgment result. If there is a tie, the decision is made according to the preset level priority or the sub-classifier confidence. The module then outputs the final assessment result.

[0017] The beneficial effects of this invention are: 1. Addressing the industry pain points of traditional bridge structural health assessment methods, such as the inability to isolate environmental interference, limited feature dimensions, low classification accuracy with small samples, ambiguous risk assessment levels, and lack of quantification of overturning risk, this invention achieves comprehensive purification of multi-source monitoring data through a four-level preprocessing mechanism combined with a dual anti-interference system of horizontal comparison with similar monitoring points and vertical comparison with historical time-series baselines; and constructs a structured loss function with L2 regularization constraints. A four-dimensional composite constraint framework of "data fitting, spatial gradient, temperature gating, and weight regularization" is adopted as the model optimization criterion, in which the spatial gradient consistency constraint term is used. Taking full account of the spatial mechanical transfer characteristics of the structural continuum along the bridge span direction, the temperature quiescent period time gradient penalty term is included. The temperature change rate gating mechanism accurately distinguishes between the reversible response of thermal expansion and contraction and the irreversible degradation trend of the structure itself, with weighted regularization terms. Effectively suppressing model overfitting, it ultimately achieves precise separation between the reversible response of the temperature environment and the irreversible degradation trend of the structure itself, purifying the true damage characteristics of the structure, and solving the technical problem of environmental interference masking structural degradation signals in traditional methods. By constructing three core indicators—temperature environment characteristics, vibration mode characteristics, and multi-physical quantity coupling correlation characteristics—it breaks through the limitations of single physical quantity assessment, forming a high-dimensional feature vector system with temperature and vibration monitoring as the main body and multi-parameter synergy, comprehensively characterizing the health status of bridge structures.

[0018] 2. To address the shortcomings of traditional SVM hyperparameter optimization methods, which rely solely on fixed-step grid traversal, suffer from high computational redundancy, and neglect the time-varying characteristics of bridge monitoring data, this invention proposes a dynamic sparse search strategy. After quickly identifying the optimal parameter region using a large-step coarse-grained traversal in the first round, a second round of small-step fine-grained search is performed only within the neighborhood of this region. This avoids ineffective traversal of the entire parameter space, reducing the computational cost of hyperparameter optimization by approximately 20% compared to traditional fixed-step grid search while maintaining global optimality search capability. Simultaneously, the objective function is optimized... A K-fold cross-validation mechanism is used to traverse the search space. and Obtain the globally optimal parameters When multiple sets of parameters achieve the same optimal accuracy, the combination with the smaller penalty parameter C value is preferred to suppress the risk of overfitting, which significantly improves the generalization stability and robustness of the model in scenarios with small sample high-dimensional data.

[0019] 3. The system of this invention uses temperature and vibration monitoring data as core inputs. It works collaboratively through five modules: data acquisition and preprocessing, structured loss function trend term extraction, multi-dimensional coupled feature construction, SVM intelligent discrimination, and hierarchical risk assessment. Ultimately, it outputs assessment results in the form of component-level health levels, confidence distributions, graded risk alarms, and penetrating assessment reports, achieving a precise, multi-dimensional characterization of the health status of bridge structures from micro-components to the macro-structure as a whole. This system ensures data input quality through multi-level preprocessing and a dual-contrast anti-interference mechanism; achieves precise separation of environmental interference and structural degradation through a structured loss function; and adapts to the intelligent discrimination requirements of small-sample, high-dimensional data using an RBF-SVM multi-classification model, significantly improving assessment accuracy compared to traditional single-threshold discrimination methods. Furthermore, it supports continuous access to monitoring data and dynamic model updates throughout the bridge's entire lifespan, providing efficient and reliable technical support for large-scale intelligent safety operation and maintenance of highway and railway bridges. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0022] This embodiment focuses on bridge vibration and temperature as the core monitoring subjects, supplemented by deformation and tilt monitoring data. It utilizes a dual anti-interference mechanism—comparing horizontally similar sensors and longitudinally historical time-series data—to complete standardized preprocessing of multi-source signals. Through a regularized constraint loss function, it accurately separates environmental response from structural degradation trends, refining effective damage features. It constructs a multi-dimensional feature vector coupling temperature, vibration, and multiple physical quantities. Based on an RBF kernel function SVM multi-classification model, it achieves intelligent four-level health status identification. Figure 1 As shown, the main steps include: S101: Acquire multi-source time-series signals, and perform four-level preprocessing operations on the multi-source time-series signals based on the spatial lateral comparison between the responses of similar sensors and the temporal longitudinal comparison of the responses of each measurement point relative to the historical baseline: abnormal data removal, high-frequency noise filtering, multi-source signal time synchronization, and missing data repair, and output the preprocessed multi-source time-series signal set.

[0023] Specifically, this embodiment conducts comprehensive health monitoring of the bridge structure. A triaxial accelerometer is used to collect real-time vibration data of the main beam, pier trunk, and cable anchorage zone, with a minimum sampling frequency of 50Hz. Simultaneously, high-precision temperature sensors with a sampling rate ≥1Hz are deployed at key bridge locations to obtain temperature field changes in the top and bottom slabs of the main beam and the surface of the piers. An open environmental reference point is also set up for temperature compensation. Furthermore, inclinometers and displacement gauges are used to assist in monitoring the displacement and tilt of the main beam at mid-span, beam ends, and pier top sections, thus forming a comprehensive data acquisition system for the static and dynamic response of the structure.

[0024] To address issues such as abnormal fluctuations, high-frequency noise, time misalignment, data gaps, and temperature-environment coupling interference in on-site monitoring data, and considering the dual-subject data characteristics of vibration and temperature, this embodiment employs a four-level preprocessing mechanism to comprehensively optimize data quality. The specific operations are as follows: (1) Abnormal data removal: By combining the results of horizontal comparison of similar measuring points and vertical comparison of historical baselines, we can accurately identify and remove abnormal jumps in vibration and temperature signals caused by communication interruption, sensor failure, or sudden external interference, so as to avoid erroneous data interfering with subsequent analysis.

[0025] In this embodiment, the comparison of similar measuring points in the lateral direction refers to: at the same sampling time t, considering the measured values ​​of all similar sensors (such as all triaxial accelerometers grouped together, all temperature sensors grouped together) deployed in the bridge structure as a spatial sample set X(t)={x1(t),x2(t),…,x n (t)}; where n is the total number of sensors of this type. Calculate the median m(t) and the median absolute deviation MAD(t) of this sample set: ; ; Then the measuring point Lateral deviation at time t for: ; In the formula, For a very small positive number, such as ;when When α is the lateral outlier threshold, it is usually taken as... The measurement point was determined to be a lateral outlier at that moment.

[0026] The longitudinal historical baseline comparison refers to: for each independent measuring point A dynamic baseline was established using historical data (same temperature range, same time period) on the bridge's health status, and its mean was calculated. with standard deviation : ; ; in, This represents the total number of valid samples during the health baseline period at this monitoring point.

[0027] Then the measuring point Longitudinal deviation at time t for: ; when When β is the longitudinal outlier threshold, it is usually taken as... The measurement point was determined to be a longitudinal outlier at that moment.

[0028] (2) High-frequency noise filtering: For high-frequency vibration signals, an adjustable window length moving average filtering algorithm is adopted, with window length W∈[5,10]min. The filtering calculation formula is as follows: ; in, This is the original vibration time series data. The filtered data effectively removes high-frequency random noise; for slowly changing temperature signals, steady-state mean filtering is used to preserve the true temperature field distribution and long-term trend of the structure.

[0029] (3) Multi-source signal time synchronization: The vibration, temperature, displacement, tilt angle and other multi-dimensional signals are uniformly timestamped to eliminate the sampling time difference of multiple devices and ensure the accuracy of temperature-vibration coupling correlation analysis and time series trend analysis.

[0030] (4) Missing data repair: To address the problem of short-term missing data, linear interpolation or neighboring value filling method is used to complete the main data of vibration and temperature, ensuring the integrity and continuity of monitoring time series data, and providing complete data support for the dual comparison analysis mechanism.

[0031] Finally, the preprocessed multi-source time-series signal set is output. : ; in, , , , These are the pre-processed vibration, temperature, displacement, and tilt timing signals, respectively.

[0032] This embodiment uses a long-span cable-stayed bridge in operation in China as the application object. A complete multi-source monitoring data acquisition system for the bridge structure was constructed by deploying 36 triaxial accelerometers, 24 temperature sensors, and multiple inclinometers and displacement gauges at the mid-span, 1 / 4-span, 3 / 4-span, and key sections of the bridge towers. Assuming that 9 acceleration measurement points were deployed at the same time across the bridge span, the measured acceleration values ​​(unit: m / s²) were [0.85, 0.92, 0.88, 0.95, 5.2, 0.90, 0.87, 0.93, 0.89]. The median m(t) of this spatial sample set was calculated to be 0.89, and the median absolute deviation MAD(t) = median{|0.85-0.89|, |0.92-0.89|, ..., |0.89-0.89|} = 0.04. The lateral deviation of measurement point 5 was: ; Let α=3, because d p (5,t)=107.75 > 3, therefore the measuring point is determined to be a lateral outlier.

[0033] Simultaneously, a longitudinal historical baseline was established for this measuring point. Its historical average acceleration under healthy conditions was μ=0.88, and its standard deviation was σ=0.06. Therefore, the longitudinal deviation is: ; Take β=4, because d t (5,t)=72.00 > 4, which is also determined to be a vertical outlier.

[0034] This data point was simultaneously identified as an outlier by both horizontal and vertical mechanisms, confirming it as a genuine anomalous jump point caused by sensor malfunction or sudden external interference. It was then removed and replaced with a moving average filtered value. Taking a certain evaluation period at the mid-span section of the main span as an example, the acceleration measurement point at the mid-span of a bridge's main span was selected, with a sampling frequency of 60Hz and a window length W=10min (36000 points). The moving average filter was: ; After being removed, the point is replaced by a moving average filtered value, and the filtered amplitude is approximately 0.9 m / s².

[0035] S102: For the preprocessed multi-source time series signal set, construct a structured loss function with L2 regularization constraints. Use a four-dimensional composite constraint framework that includes data fitting constraints, spatial gradient consistency constraints, temperature change rate gating constraints, and weight regularization constraints. Separate the reversible response of the temperature environment from the irreversible degradation trend of the structure through the constraint framework, and output the purified structural trend term.

[0036] Specifically, this embodiment constructs a dedicated loss function with regularization constraints and relies on bidirectional comparative baselines (horizontal and vertical) to achieve accurate separation of temperature environmental effects, structural degradation trends, and noise signals. By refining the characteristic information that truly reflects bridge structural damage, it provides clean and effective input data for subsequent intelligent assessment models.

[0037] The specific steps for constructing the loss function and extracting the structural trend term include: A01: Construction of the Structured Loss Function. Based on the temporal characteristics and environmental coupling patterns of bridge monitoring signals, a dedicated loss function with L2 regularization constraints is constructed. This function uses both lateral deviations from similar measuring points and longitudinal historical baseline deviations as dual constraints to suppress model overfitting and accurately separate the reversible temperature response from the irreversible degradation trend of the structure itself. The structured loss function... The expression is as follows: ; Where λ1, λ2, and λ3 are regularization coefficients, typically ranging from 0 to 1; the structured loss function consists of four terms, including the data fitting term: N is the total number of samples. For the sensor measured data, the first The actual monitoring value of each sample For the first The model predicted values ​​for each sample, of which the actual monitored values ​​are... Taken from the preprocessed multi-source time-series signal set in step S101 , is the model prediction value The fitting objective; spatial gradient consistency constraint: ; For the residual vector, Spatial gradient operator; Temporal gradient penalty term during temperature quiescence period: T represents the temperature field. The threshold value is used for gate control. This indicates that the model's predicted value is the partial derivative with respect to time. Represents the partial derivative of temperature with respect to time; indicator function The value is 1 if the condition is true, and 0 if the condition is false; weight regularization term: ; These are all the trainable parameters of the model.

[0038] In this embodiment, the above-mentioned structured loss function The precise separation of temperature environment effects and structural degradation trends is achieved through the synergistic effect of four constraints. The functions and effective conditions of each constraint are as follows: (1) Data fitting term Constrained model predictions Approximating the preprocessed true monitoring values ​​at all sample points It is the fundamental fidelity term of the loss function, ensuring that the prediction results do not deviate from the observed data.

[0039] (2) Spatial gradient consistency constraint term Constrained residuals The spatial distribution along the longitudinal direction of the bridge should be continuous, meaning that the residual changes between adjacent measuring points should not exhibit drastic abrupt changes. This constraint is based on the principle of continuity in structural mechanics—the difference in response between adjacent sections should transition smoothly to avoid the model misfitting local sensor noise or single-point anomalies as structural damage signals.

[0040] (3) Temperature quiescent period time gradient penalty term Constraints are applied when the rate of temperature change is less than a gating threshold. The "temperature quiescent period" (i.e.) Within ) the model predicted value It should not change significantly over time. The physical basis for this is that when the ambient temperature remains almost constant, the structure does not have a reversible driving force for thermal expansion and contraction. If the trend continues, then this change must originate from irreversible degradation of the structure itself, rather than from temperature effects. Therefore, this constraint blocks the temporal drift of model predictions during the temperature quiescent period, forcing the irreversible trend term to... It must be expressed through data fitting terms rather than time-series gradient penalty terms, thereby forcibly separating the irreversible degradation component from the temperature-reversible component.

[0041] (4) Weight regularization term Constraint model parameters The L2 norm is used to suppress overfitting of the model to noise and outliers in the training data, and to ensure the generalization stability of the separation results in small sample scenarios.

[0042] Traditional bridge monitoring methods only consider data fitting accuracy or temperature compensation effect. This embodiment constructs a structured loss function... A four-dimensional composite constraint framework of "data fitting - spatial gradient - temperature gating - weight regularization" is adopted as the model optimization criterion. This framework ensures that the optimization objective takes into account both the basic approximation accuracy of the sensor's measured and predicted values, as well as the spatial mechanical transmission characteristics of the structural continuum along the bridge span direction. Furthermore, a temperature change rate gating mechanism is introduced to distinguish between the reversible response of thermal expansion and contraction and the irreversible degradation trend of the structure itself. At the same time, parameter decay constraints are used to suppress model overfitting. This framework can guide the model to more comprehensively and physically consistently separate the temperature effect and the structural health degradation component from the monitoring signal coupled with environmental noise.

[0043] A02: Precise Extraction of Structural Trend Term. Based on the constructed structured loss function, the preprocessed multi-source time-series signals such as temperature, vibration, and displacement are iteratively solved to remove the diurnal and seasonal periodic reversible fluctuation components of temperature and random noise components. Finally, irreversible structural trend terms that only reflect the stiffness degradation, component damage, and deformation accumulation of the bridge structure are extracted. This completely solves the problems of environmental interference superposition and blurred deterioration characteristics in traditional monitoring signals, providing accurate basic data support for the construction of multi-dimensional features.

[0044] Trend extraction formula as follows: ; This sequence represents the core features that reflect only the irreversible deterioration trend of the bridge structure itself, and is directly used for subsequent damage early warning and remaining life assessment. In this formula, The preprocessed multi-source time-series signal set Signal components in; reversible temperature response For model predictions This corresponds to the temperature-induced reversible structural response component. In the structured loss function... After optimization and convergence, the predicted values ​​output by the model It is broken down into two parts: ; in, The reversible structural response caused by temperature field changes is periodic and recoverable; This reflects the irreversible deterioration trend of structural stiffness degradation and damage accumulation.

[0045] Specific example: Taking regularization parameters Gating threshold ε = 0.02°C / min, structured loss function : ; Where N = 720 (12 hours × 60 minutes), the rate of temperature change The indicator function is now active.

[0046] Trend extraction formula : ; S103: Based on the purified structural trend term and the preprocessed multi-source time series signal set, construct three core indicators: temperature environment characteristics, vibration mode characteristics, and multi-physical quantity coupling correlation characteristics, and form a standardized eight-dimensional feature vector.

[0047] Specifically, based on the purified structural trend terms and the pre-processed vibration, temperature, and auxiliary deformation signals, this method breaks through the limitations of traditional single physical quantity features and constructs three core indicators: temperature environment characteristics, vibration modal characteristics, and multi-physical quantity coupling correlation characteristics. This forms a high-dimensional, high-discrimination feature vector system with temperature and vibration monitoring as the main body and multi-parameter synergy. The specific steps are as follows: B01: Temperature Environment Characteristics Construction. Based on the comparison of temperature at similar measuring points in the lateral direction and the comparison of historical temperature baselines in the longitudinal direction, the temperature field distribution and time-varying characteristics of the bridge structure are extracted. This is used to isolate the reversible structural response caused by environmental temperature and accurately distinguish between environmental disturbances and structural degradation.

[0048] The structural-environmental temperature difference ΔT characterizes the difference between the temperature of the bridge component itself and the external environment temperature, reflecting the structural temperature stress level. Its calculation formula is as follows: ; Among them, T str For the measured temperature of bridge components, T env This is the ambient reference temperature.

[0049] Lateral temperature non-uniformity coefficient δT cross The temperature difference between symmetrical members on both sides of the same span is used to identify structural anomalies such as unilateral heating and local constraint anomalies. The calculation formula is as follows: ; Among them, T L T R The temperatures were measured at symmetrical points on the left and right sides of the bridge, respectively.

[0050] B02: Vibration Modal Characteristic Construction. Based on structural vibration response changes and combined with the temperature-vibration coupling relationship to isolate environmental interference, this study characterizes latent defects such as bridge stiffness degradation, component loosening, and internal damage, selecting four types of modal characteristics: (1) Effective value of acceleration A rms The formula for calculating the overall vibration response intensity of a structure is as follows: ; in, M represents the discrete acceleration sample values, where M is the number of sampling points.

[0051] (2) Dominant frequency offset Δf: The difference between the current structural dominant frequency and the dominant frequency during the healthy reference period, representing the change in structural stiffness. Its calculation formula is: ; Among them, f cur f is the first-order main frequency of the current structure. base This is the reference frequency for a bridge in a healthy state.

[0052] (3) Damping ratio change Δζ: The change in structural damping characteristics, reflecting the connection state and internal damage degree of the components. Its calculation formula is: ; Where, ζ cur Let ζ be the current damping ratio of the structure. base The structural reference damping ratio is given.

[0053] (4) Vibration energy spectral density change ΔPSD: Characterizes the abnormal shift in the distribution of structural vibration energy, and its calculation formula is: ; Among them, PSD cur The vibration power spectral density (PSD) is the value of PSD under the current operating conditions. base The vibration power spectral density is the value under healthy baseline operating conditions.

[0054] B03: Construction of Multi-Physical Quantity Coupling Correlation Features. Breaking through the limitations of single-physical-quantity assessment, this method uses temperature and vibration signals as the core, and through multi-dimensional signal coupling relationships, identifies complex structural damage and the overall collaborative working state. Temperature-vibration coupling coefficient To quantify the coupling relationship between temperature change and structural vibration response, eliminate temperature-induced spurious vibration anomalies, and calculate using the Pearson correlation coefficient: ; in, The value is usually taken as [-1, 1]. Let T be the temperature and vibration covariance; D(T) and D(A) be the variances of the temperature and vibration data, respectively. The closer the value is to 1, the higher the probability that the vibration abnormality is induced by the temperature environment.

[0055] Synchronization coefficient of lateral displacement of left and right sides This characterizes the coordinated deformation capacity of the left and right spans of a bridge, reflecting the overall stress balance. Its calculation formula is as follows: ; Among them, U L U R These are the average lateral displacements of the left and right sides, respectively. The closer the value is to 1, the better the performance of the structure in collaborative work.

[0056] In this embodiment, the data source mapping relationships for the aforementioned features are as follows: Temperature environment characteristics (ΔT, δT) cross The preprocessed raw temperature signal The input is used to quantify the temperature-dependent boundary conditions of the bridge structure.

[0057] Vibration modal characteristics (A) rms Δf, Δζ, ΔPSD) are acceleration components in the trend term. Replace the preprocessed original vibration signal As input, it is ensured that the extracted modal features are completely stripped of temperature reversible effects and only reflect the structural stiffness degradation and damage accumulation state.

[0058] Among the characteristics of multi-physical quantity coupling correlation, the temperature-vibration coupling coefficient The preprocessed raw temperature signal Compared with the preprocessed original vibration signal As input, it is used to quantify the degree to which temperature changes drive the vibration response, and to help determine whether vibration anomalies are induced by changes in the temperature environment; the lateral displacement synchronization coefficient of the left and right axes. displacement components in the trend term Replace the preprocessed original displacement signal As input, it is used to evaluate the synergy of irreversible deformation of the structural body.

[0059] In the above mapping relationship, Using the original temperature and vibration signals as inputs to preserve the discriminative nature of the physical coupling between temperature and vibration, other structural response characteristics (vibration modal characteristics and displacement coordination) are all... The data source is used to strip away temperature reversible effects.

[0060] B04: Standardized Feature Vector Construction. Integrating the above-mentioned temperature environment features, vibration mode features, and multi-physical quantity coupling correlation features, a standardized feature vector with uniform dimensions and complete features is constructed for each component and each measuring point. ; Specific example: Temperature measurement point value T str =28.5°C, environmental reference point T env =25.2°C. Temperature difference between structure and environment ΔT: ; Lateral temperature non-uniformity coefficient δT cross : ; In the vibration modal characteristics, M=36000 sampling points are taken, and the effective value of acceleration A is... rms : ; Main frequency offset Δf: ; Damping ratio change Δζ: ; Vibrational energy spectral density change ΔPSD: ; Temperature-vibration coupling coefficient : ; Synchronization coefficient of lateral displacement of left and right sides : ; Standardized feature vector F: ; S104: Based on the standardized eight-dimensional feature vector and combined with the pre-labeled structural health level labels, construct a training set for model training; establish an RBF-SVM multi-classification model, constructing sub-classifiers between any two health levels, and adopting a strategy combining dynamic sparse search with decreasing step size and cross-validation, with the average classification accuracy of the validation set as the optimization objective. Within the dynamically shrinking candidate subset, collaboratively optimize the kernel parameters and penalty parameters of each sub-classifier to obtain the optimal hyperparameters; based on the optimal hyperparameters, retrain the RBF-SVM multi-classification model using the training set; for a new input feature vector, input it into all sub-classifiers of the trained RBF-SVM multi-classification model, with each sub-classifier using the feature vector as input for level determination, classifying it as one of two health levels and casting a vote for that level; count the cumulative votes for each health level; if there is a unique highest vote, the health level corresponding to that vote is taken as the determination result; if there is a tie, the decision is made according to the preset level priority or the sub-classifier confidence; output the final evaluation result.

[0061] Specifically, considering the characteristics of small-sample, nonlinear, and high-dimensional datasets in bridge structural health monitoring, an SVM multi-classification intelligent assessment model is established using standardized feature vectors purified from temperature and vibration data as input. This model employs an SVM multi-classification algorithm based on the RBF kernel function (hereinafter referred to as the RBF-SVM multi-classification model), combined with sample labeling, parameter optimization, and a voting classification mechanism, to achieve accurate classification and discrimination of the health status at the bridge component level, thus solving the problems of poor generalization ability and low classification accuracy of traditional machine learning models with small samples. The specific steps are as follows: C01: Training Sample Hierarchical Labeling. Based on monitoring samples divided into weekly time windows, and combined with industry safety standards such as the "Highway Bridge Maintenance Specification" and the "General Specification for Highway Bridge and Culvert Design," as well as the gradual evolution of structural defects from microscopic damage to macroscopic deterioration, the bridge structural health status is divided into four levels by comparing the response deviations of similar transverse measuring points (such as the left, middle, and right webs of the same bridge span or adjacent transverse sections) and analyzing the longitudinal historical time-series deterioration trends. Corresponding numerical labels are matched for each level, and the standardized feature vectors of the samples are associated with the health labels to complete the labeling, thus constructing a standardized training dataset.

[0062] Excellent (Label 0): Temperature environment characteristic indicators and vibration modal characteristic indicators are all within the historical benchmark range and within the statistical range of historical data of the same period under healthy conditions. The response deviation between transverse similar measuring points maintains reasonable consistency. There is no statistically significant continuous deterioration trend. The overall working condition of the structure is normal and meets the design expectations.

[0063] Note (Label 1): If a single temperature or vibration characteristic index exceeds the normal historical benchmark range, or there is a slight deviation at the same type of measurement point in the lateral direction, but none of the indicators have reached the warning threshold defined by the standard or experience, and the time series analysis has not found a continuous deterioration trend, the structure has slight anomalies but is still within the acceptable range. It is recommended to increase the frequency of observation.

[0064] Warning (Label 2): ​​Multiple temperature-vibration coupling correlation characteristic indicators simultaneously exceed the normal range, or a certain core load-bearing characteristic indicator shows a continuous deterioration trend in the time dimension, while the deviation of the same type of measuring points in the lateral direction increases significantly, indicating that the structural disease has entered a clear development state, and detailed testing and special assessment need to be initiated.

[0065] Hazard (Label 3): Significant nonlinear coupling anomaly exists between temperature and vibration characteristics. The deviation of the same type of transverse measuring points is seriously exceeded and shows a divergent trend. At the same time, abnormal structural deformation is also present, indicating that the structural deterioration has entered a serious stage. The overall bearing capacity has deteriorated significantly, and the risk of overturning or instability has increased significantly. Traffic control and emergency reinforcement measures must be taken immediately.

[0066] A single training sample consists of a standardized feature vector F and a health label y∈{0,1,2,3}, forming a complete supervised learning standardized training dataset.

[0067] C02: Kernel Function Selection. The Radial Basis Function (RBF) is chosen as the kernel mapping function for SVM. The RBF kernel function can map nonlinearly inseparable problems in the original low-dimensional feature space to a high-dimensional feature space, making them linearly separable in that space. Its mathematical expression is: ; in, For any two standardized input feature vectors; The square of the Euclidean distance between the two vectors; For kernel function bandwidth parameters ( ), used to control the radial range of the Gaussian kernel; The kernel similarity between any two vectors is calculated using the RBF kernel function, which is a general similarity measurement rule.

[0068] C03: One-to-One Multi-Classification Strategy. For the four-level classification task of bridge structural health status (Excellent, Attention, Warning, Hazardous), a one-to-one multi-classification strategy is adopted, constructing an independent binary SVM sub-classifier between any two health level categories. Assuming the total number of categories is K=4, the number of sub-classifiers required is... Each subclassifier SVM p,q (in The decision function is to train using only training samples belonging to class p and class q. ; in, For category With category The total number of training samples; These are the Lagrange multipliers obtained by optimizing the corresponding sub-classifiers; The labels are binary (category p is denoted as +1, and category q is denoted as -1). For bias terms; To predict the samples under the general similarity measurement rule The vector substituted into the position of the second parameter for actual prediction calculation With the sample to be predicted Kernel similarity between them; For a sign function, the value is +1 if the independent variable is greater than 0, and -1 if it is less than 0; Classified as ;like Classified as .

[0069] For the new input feature vector F new When making predictions, a "voting" mechanism is used: the feature vector is sequentially input into all 6 sub-classifiers, each of which is an SVM. p,q It is classified into category p or category q, and a vote is cast for the classified category. If there is a tie, the category with higher confidence or the superior category is chosen. The superior categories are ordered in ascending order of risk: Excellent < Caution < Warning < Danger. The cumulative number of votes received for each numerical label (0, 1, 2, 3) is recorded. ( Finally, the category with the highest number of votes will be used as the health status determination result for that sample, that is: ; C04: Global Optimization of Model Hyperparameters. Addressing the high sensitivity of the penalty parameter C and kernel function bandwidth parameter γ to classification performance in the RBF-SVM multi-class classification model, an optimization strategy combining grid search and K-fold cross-validation is employed. The optimal hyperparameters are selected from a dynamically shrinking candidate subset, with the average classification accuracy on the validation set as the optimization objective.

[0070] The objective function to be optimized is: ; ; in, The search range for the penalty parameter C, The search range for the kernel function bandwidth parameter γ; Let V be the classification accuracy under the current parameter combination in the v-th fold cross-validation. This is a complexity penalty coefficient used to suppress the risk of overfitting.

[0071] After iterating through all parameter combinations, select The largest one is used as the globally optimal parameter. This penalty term automatically tends to select a smaller C value when the accuracy is similar, thus effectively suppressing the risk of overfitting.

[0072] Traditional grid search methods traverse the entire parameter space using fixed, equal-interval step sizes, resulting in high computational redundancy and a tendency to get trapped in local optima. This embodiment proposes a dynamic sparse search strategy. In the first round, a large-step, coarse-grained traversal is used to identify high-potential regions containing the optimal parameters at low computational cost. Then, a second round of fine-grained, small-step search is performed only within the neighborhood of this region. Simultaneously, a complexity penalty term is introduced into the objective function, imposing a soft constraint on excessively large penalty parameters C. This dynamically suppresses overfitting during the search process, significantly reducing ineffective traversal of the entire parameter space while maintaining search accuracy, thus improving optimization efficiency.

[0073] C05: Model output results. Based on optimal hyperparameters. The final SVM multi-class classification model is retrained using all labeled training samples. For any sample to be evaluated, after processing in steps S101 to S103, it is input into the model, and a health status level label is output through a one-to-one voting mechanism. These correspond to four status levels: Excellent, Caution, Warning, and Danger. The output also shows the vote vector (Vote) for each level. ; Normalized confidence level is used for discrimination, and the normalized confidence level P is... conf : ; Normalized confidence level When the confidence level is high, it indicates that the judgments of the six sub-classifiers are highly consistent, and the evaluation conclusion can be directly adopted; when At a medium confidence level, it indicates that most subclassifiers agree but there are minor disagreements; the evaluation results are still acceptable, and it is recommended to consider historical trends for a comprehensive assessment. At times, the confidence level is low, and the voting results are scattered, requiring manual verification or supplementary monitoring data; when If the confidence level is extremely low, the evaluation result is invalid, and data needs to be collected again or the model status needs to be checked. In practical engineering, this is generally... As the minimum standard for the normal adoption of assessment results, for safety-sensitive scenarios (such as important bridges and high-speed rail lines), it is recommended to raise the threshold to above 75% to further enhance the reliability of the assessment conclusions.

[0074] when When the system automatically triggers a graded risk alert: a yellow alert is activated in the warning state (label 2), and it is recommended to initiate detailed testing and special assessment; a red alert is activated in the danger state (label 3), and it is recommended to immediately take traffic control and emergency reinforcement measures. At the same time, the assessment results and alert information are pushed to the bridge operation and maintenance management platform.

[0075] Specific example: using two sample feature vectors and For example, let γ = 0.5, the radial basis function (RBF): ; One-to-one SVM subclassifier (total) (N), taking the sub-classifiers for categories 0 and 3 as an example, let N be... 0.3 For 150 training samples, the decision function is: ; If it is classified as category 0, which is excellent (label 0), then vote for category 0.

[0076] Hyperparameter grid search range: , After two rounds of dynamic sparse search, the first round used a large step size (step size factor of 2). 2 A coarse-grained traversal is performed to initially pinpoint the optimal region in C≈2. 7 =128, γ≈2 -7 A high-potential area near 0.0078; the second round focuses on this area, using a small step size (step size factor of 2). 0.5 Perform a fine-grained search. Let η = 0.05, K = 5, then the objective function for 5-fold cross-validation is: ; After 5-fold cross-validation calculation, when At that time, the globally optimal parameters were obtained after two rounds of dynamic sparse search optimization; the average accuracy was... Penalty items The overall score is It is the highest among all candidate combinations. Therefore, the globally optimal parameter is... That is, (64, 0.005): Voting output: The voting results of the 6 sub-classifiers are as follows: Class 0 received 4 votes, Class 1 received 1 vote, Class 2 received 1 vote, and Class 3 received 0 votes. ; Health status assessment results: ; Normalized confidence level P conf : ; in, The number of votes to be obtained for the predicted label (label=0).

[0077] This embodiment uses an eight-dimensional feature vector as input and an RBF-SVM multi-classification model optimized by dynamic sparse search. The optimal hyperparameters are C*=64 and γ*=0.005. The average accuracy of 5-fold cross-validation is 96.8%. After voting by 6 one-to-one sub-classifiers, the final result is that the health status of the bridge structure during the evaluation period is "excellent" (label 0), with a normalized confidence level of 66.7%. This indicates that all monitoring indicators of the bridge structure during this period are within the health benchmark range, the overall working status is normal, and it meets the design expectations. No additional inspection or maintenance measures are required. This evaluation result is consistent with the manual inspection records during the same period.

[0078] This embodiment also provides an intelligent structural health status assessment system, including: The data acquisition and preprocessing module is used to acquire multi-source time-series signals and perform multi-level preprocessing operations on the multi-source time-series signals based on the spatial lateral comparison between the responses of similar sensors and the temporal longitudinal comparison of the responses of each measuring point relative to the historical baseline, and output the preprocessed multi-source time-series signal set. The structured loss function trend term extraction module is used to construct a structured loss function for the preprocessed multi-source time series signal set. It adopts a four-dimensional composite constraint framework that includes data fitting constraints, spatial gradient consistency constraints, temperature change rate gating constraints, and weight regularization constraints. Through the constraint framework, the reversible response of the temperature environment and the irreversible degradation trend of the structure are separated, and the purified structured trend term is output. The multidimensional coupling feature construction module is used to construct three core indicators—temperature environment features, vibration mode features, and multi-physical quantity coupling correlation features—based on the purified structural trend term and the preprocessed multi-source time series signal set, forming a standardized multidimensional feature vector. The SVM intelligent discrimination module is used to construct a training set for model training based on the standardized multidimensional feature vector and pre-labeled structural health level labels; establish an RBF-SVM multi-classification model, construct sub-classifiers between any two health levels, and adopt a strategy combining dynamic sparse search with decreasing step size and cross-validation, with the average classification accuracy of the validation set as the optimization objective. Within the dynamically shrinking candidate subset, the kernel parameters and penalty parameters of each sub-classifier are jointly optimized to obtain the optimal hyperparameters; based on the optimal hyperparameters, the RBF-SVM multi-classification model is retrained using the training set. The hierarchical risk assessment module is used to input the new input feature vector into all sub-classifiers of the trained RBF-SVM multi-classification model. Each sub-classifier uses the feature vector as input to determine the level, classifying it into one of two health levels and casting a vote for that level. The module counts the cumulative number of votes for each health level. If there is a unique highest vote, the health level corresponding to that vote is taken as the judgment result. If there is a tie, the decision is made according to the preset level priority or the sub-classifier confidence. The module then outputs the final assessment result.

Claims

1. A method for intelligent assessment of structural health status, characterized in that, Includes the following steps: S1: Acquire multi-source time-series signals, and perform multi-level preprocessing operations on the multi-source time-series signals based on the spatial lateral comparison between the responses of similar sensors and the temporal longitudinal comparison of the responses of each measuring point relative to the historical baseline, and output the preprocessed multi-source time-series signal set. S2: For the preprocessed multi-source time series signal set, construct a structured loss function and adopt a four-dimensional composite constraint framework that includes data fitting constraints, spatial gradient consistency constraints, temperature change rate gating constraints, and weight regularization constraints. Through the constraint framework, separate the reversible response of the temperature environment from the irreversible degradation trend of the structure and output the purified structural trend term. S3: Based on the purified structural trend term and the preprocessed multi-source time series signal set, construct three core indicators: temperature environment characteristics, vibration mode characteristics, and multi-physical quantity coupling correlation characteristics, and form a standardized multi-dimensional feature vector. S4: Based on the standardized multidimensional feature vectors and combined with pre-labeled structural health level labels, construct a training set for model training; establish an RBF-SVM multi-classification model, construct sub-classifiers between any two health levels, and adopt a strategy combining dynamic sparse search with decreasing step size and cross-validation, with the average classification accuracy of the validation set as the optimization objective. Within the dynamically shrinking candidate subset, collaboratively optimize the kernel parameters and penalty parameters of each sub-classifier to obtain the optimal hyperparameters; based on the optimal hyperparameters, retrain the RBF-SVM multi-classification model using the training set. S5: The RBF-SVM multi-classification model trained in S4 is used to predict the new input feature vector. The feature vector is then input into all sub-classifiers. Each sub-classifier determines the feature vector as one of two health levels and casts a vote for the determined level. Count the cumulative number of votes obtained for each health level; if there is a unique highest number of votes, the health level corresponding to that number of votes will be used as the judgment result; If a tie occurs, the decision is made according to the preset priority level or the confidence of the subclassifier; the final evaluation result is then output.

2. The intelligent structural health status assessment method according to claim 1, characterized in that, In S1, the multi-level preprocessing operation includes: By combining the results of horizontal comparison with similar measurement points and vertical comparison with historical baselines, abnormal data are removed from the original multi-source time series signals. For the high-frequency monitoring signals after removing anomalies, an adjustable window length moving average filtering algorithm is used to filter out high-frequency noise. The filtered monitoring signals are calibrated with a unified timestamp to achieve time synchronization of multi-source signals. For monitoring signals that have short-term data gaps after time synchronization, linear interpolation or nearest neighbor filling methods are used to repair the missing data.

3. The intelligent structural health status assessment method according to claim 2, characterized in that, The horizontal comparison of similar measurement points is as follows: at the same sampling time t, the measured values ​​of all similar sensors deployed in the infrastructure structure are regarded as a spatial sample set X(t)={x1(t),x2(t),…,x n (t)}; where n is the total number of sensors of this type; calculate the median m(t) and the median absolute deviation MAD(t) of this sample set: ； ； Then the measuring point Lateral deviation at time t for: ； when If the measurement point is at that time, it is determined to be a lateral outlier; where, α is a very small positive number, and α is the lateral outlier threshold; The longitudinal historical baseline comparison refers to: for each independent measuring point A dynamic baseline is established using historical data from the same period when the infrastructure structure is in a healthy state, and its mean is calculated. with standard deviation : ； ； Then the measuring point Longitudinal deviation at time t for: ； when When the measurement point is determined to be a longitudinal outlier at that moment; in, β represents the total number of valid samples during the healthy baseline period at this measurement point, and β is the longitudinal outlier threshold.

4. The intelligent structural health status assessment method according to claim 1, characterized in that, In S2, the structured loss function is obtained through the following formula: ； In the formula, λ1, λ2, and λ3 are regularization coefficients; The structured loss function consists of four terms: Data fitting terms: N is the total number of samples. For the sensor measured data, the first The actual monitoring value of each sample For the first The model predicted values ​​for each sample; where the actual monitored values ​​are... The values ​​are taken from the preprocessed multi-source time-series signal set in step S1 and are model predictions. The fitting target; Spatial gradient consistency constraint: ; For the residual vector, For spatial gradient operators; Temperature quiescent period time-series gradient penalty term: T represents the temperature field. The threshold value is used for gate control. This indicates that the model's predicted value is the partial derivative with respect to time. This represents the partial derivative of temperature with respect to time. This is an indicator function that takes the value 1 when the condition is true and 0 when the condition is false. Weight regularization term: ; These are all the trainable parameters of the model.

5. The intelligent structural health status assessment method according to claim 4, characterized in that, In S2, the purification of the structural trend term is obtained through the following formula: ； In the formula, Signal components in a preprocessed multi-source time-series signal set; reversible temperature response. For model predictions The component corresponding to the temperature-induced reversible structural response; in the structured loss function After optimization and convergence, the predicted values ​​output by the model It is broken down into two parts: ; in, It is a reversible structural response caused by changes in the temperature field, and it is periodic and recoverable; This indicates an irreversible trend of structural degradation.

6. The intelligent structural health status assessment method according to claim 1, characterized in that, In S3, the three types of core indicators specifically include: A. The temperature environment feature structure includes: A1. Temperature difference ΔT between structure and environment: ; A2. Lateral temperature non-uniformity coefficient δT cross : ; In the formula, T str The actual temperature of the object being measured, T env The ambient reference temperature; T L T R These are the temperatures at symmetrical measuring points on both sides of the object being measured. B. The vibration modal characteristics are constructed as follows: B1. Effective value of acceleration A rms : ; B2. Main frequency offset Δf: ; B3. Damping ratio change Δζ: ; B4. Vibrational energy spectral density change ΔPSD: ; In the formula, For discrete acceleration sample values, M is the number of sample points; f cur f is the first-order fundamental frequency of the object currently being tested. base ζ represents the reference frequency of the tested object under healthy conditions. cur ζ represents the current damping ratio of the object being measured. base The reference damping ratio of the measured object; PSD cur The vibration power spectral density (PSD) under the current operating conditions. base The vibration power spectral density under healthy baseline operating conditions; C. The construction of multi-physical quantity coupling correlation features includes: C1. Temperature-vibration coupling coefficient : ; C2. Synchronization coefficient of lateral displacement of left and right carriages : ; In the formula, Let U be the temperature and vibration covariance, and D(T) and D(A) be the variances of the temperature and vibration data, respectively. L U R These are the average lateral displacements of the left and right sides of the measured object, respectively. The standardized multidimensional feature vector is: .

7. The intelligent structural health status assessment method according to claim 1, 2, or 3, characterized in that, Step S4 specifically includes: S41: Obtain the standardized multidimensional feature vector output from step S3 and the pre-labeled structural health level label corresponding to each sample. Associate each feature vector with its corresponding level label to form feature-label paired samples. Collect all paired samples to construct a training set. S42: Construct a one-to-one multi-class SVM model based on the RBF kernel, and build independent binary sub-classifiers between any two health levels; S43: A dynamic sparse search strategy with decreasing step size is adopted. In the first round, a coarse-grained traversal is performed in the global search space with a first preset step size to lock the high-potential region where the optimal parameters are located. In the second round, the high-potential region is used as the boundary, and a second preset step size smaller than the first preset step size is used for local fine-grained search. Combined with a cross-validation algorithm with a preset number of folds, the kernel parameters and penalty parameters of each sub-classifier are optimized in a coordinated manner with the average classification accuracy of the validation set as the optimization objective, and the optimal hyperparameters are output. S44: Based on the optimal hyperparameters, retrain all subclassifiers using the training set and save the model parameters.

8. The intelligent structural health status assessment method according to claim 1, characterized in that, The objective function used in the collaborative optimization is: ； ； In the formula, The search range for the penalty parameter C, The search range for the kernel function bandwidth parameter γ; Let V be the classification accuracy under the current parameter combination in the v-th fold cross-validation. is the complexity penalty coefficient; K is the cross-validation fold number.

9. The intelligent structural health status assessment method according to claim 1, characterized in that, The construction method of the sub-classifier in step S4 and the voting mechanism in step S5 specifically include: The sub-classifiers are constructed as follows: For the four health levels—excellent, attentive, warning, and dangerous—corresponding to health level labels y∈{0,1,2,3}, an independent binary SVM sub-classifier is constructed between any two different health level categories p and q. A total of 6 sub-classifiers are constructed; each sub-classifier is trained using only training samples belonging to class p and class q, and its decision function is: ； in, For category With category The total number of training samples; These are the Lagrange multipliers obtained by optimizing the corresponding sub-classifiers; The labels are binary, with category p marked as +1 and category q marked as -1. For bias terms; For feature vectors With the sample to be predicted Kernel similarity between them; For a sign function, the value is +1 if the independent variable is greater than 0, and -1 if it is less than 0; Classified as ;like Classified as ; The voting method is as follows: For a newly input feature vector, it is sequentially input into all 6 sub-classifiers for voting. Each sub-classifier classifies it as category p or category q and casts a vote for the classified category. If there is a unique category with the highest vote, that category is taken as the health status judgment result. If there is a tie, the category with the higher risk level is selected as the judgment result according to the progressively higher risk level. If this rule still cannot distinguish, the category with the higher confidence level output by the corresponding sub-classifier decision function is selected as the judgment result.

10. An intelligent assessment system for structural health status, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source time-series signals and perform multi-level preprocessing operations on the multi-source time-series signals based on the spatial lateral comparison between the responses of similar sensors and the temporal longitudinal comparison of the responses of each measuring point relative to the historical baseline, and output the preprocessed multi-source time-series signal set. The structured loss function trend term extraction module is used to construct a structured loss function for the preprocessed multi-source time series signal set. It adopts a four-dimensional composite constraint framework that includes data fitting constraints, spatial gradient consistency constraints, temperature change rate gating constraints, and weight regularization constraints. Through the constraint framework, the reversible response of the temperature environment and the irreversible degradation trend of the structure are separated, and the purified structured trend term is output. The multidimensional coupling feature construction module is used to construct three core indicators—temperature environment features, vibration mode features, and multi-physical quantity coupling correlation features—based on the purified structural trend term and the preprocessed multi-source time series signal set, forming a standardized multidimensional feature vector. The SVM intelligent discrimination module is used to construct a training set for model training based on the standardized multidimensional feature vector and pre-labeled structural health level labels; establish an RBF-SVM multi-classification model, construct sub-classifiers between any two health levels, and adopt a strategy combining dynamic sparse search with decreasing step size and cross-validation, with the average classification accuracy of the validation set as the optimization objective. Within the dynamically shrinking candidate subset, the kernel parameters and penalty parameters of each sub-classifier are jointly optimized to obtain the optimal hyperparameters; based on the optimal hyperparameters, the RBF-SVM multi-classification model is retrained using the training set. The hierarchical risk assessment module is used to input the feature vector of the new input into all sub-classifiers of the trained RBF-SVM multi-classification model. Each sub-classifier uses the feature vector as input to make a level determination. Each sub-classifier determines it as one of two health levels and casts a vote for the determined level. Count the cumulative number of votes received for each health level; If there is a single highest vote, the health level corresponding to that vote will be used as the judgment result; If a tie occurs, the decision is made according to the preset priority level or the confidence of the subclassifier; the final evaluation result is then output.

Images