Abnormal early warning and cause analysis method for health monitoring data of large bridge structure

Abstract

The invention discloses an abnormal early warning and cause analysis method for health monitoring data of a large-scale bridge structure, and aims to solve the problems of large environmental interference on data, low early warning reliability and inaccurate cause positioning in an existing monitoring method. The method sequentially comprises the three steps of data preprocessing and feature extraction, abnormal early warning and cause analysis. Firstly, through data cleaning, Z-score standardization, multi-domain feature extraction and PCA-multiple linear regression combined environmental factor separation, a net signal reflecting the state of a structure is obtained; triggering abnormal early warning in a mode of combining model driving and data driving; and finally, through data correlation analysis, multi-method damage identification and positioning and expert system multi-source information fusion, abnormal causes and positions are judged. According to the method, environmental interference can be effectively eliminated, the early warning false alarm and missing alarm rate is reduced, abnormal causes are accurately positioned, and reliable technical support is provided for safe operation and maintenance of a large bridge structure.

Description

Technical Field

[0001] This invention relates to the field of bridge engineering and structural health monitoring technology, specifically to a method for early warning and cause analysis of abnormalities in the structural health monitoring data of large bridges. Background Technology

[0002] As core hubs of transportation infrastructure, large bridges play a crucial role in transporting people and goods between regions. Their structural safety is directly related to traffic safety, protection of public property, and the stable operation of the social economy. With the rapid development of transportation construction in my country, a large number of long-span cable-stayed bridges, suspension bridges, and cross-sea bridges have been put into use. Some of the early-built bridges have entered the middle or late stages of their service life. Affected by long-term traffic loads, natural environmental erosion, and material aging, the structures are prone to the accumulation of hidden damage. If this damage is not detected and addressed in time, it may lead to structural performance degradation, reduced load-bearing capacity, or even major safety accidents such as collapse.

[0003] To ensure the safety of large bridges, the current mainstream approach is to establish a structural health monitoring system. This system utilizes equipment such as strain sensors, displacement sensors, vibration acceleration sensors, temperature and humidity sensors, and anemometers to collect real-time structural response data and environmental data, enabling continuous monitoring of the bridge's condition. However, existing methods for monitoring the health of large bridges still face three major technical bottlenecks in practical applications, making it difficult to meet the precise requirements of engineering operation and maintenance:

[0004] First, data interference was not effectively eliminated. Raw monitoring data is susceptible to multiple interferences. On the one hand, environmental factors can mask the damage signals of the structure itself, making the monitoring data unable to accurately reflect the actual state of the structure. On the other hand, during long-term operation, monitoring equipment may experience problems such as sensor drift, data transmission packet loss, and sudden electromagnetic interference, resulting in missing data, abnormal jumps, or high-frequency noise. If such data is used directly for analysis, it is very easy to misjudge the state of the structure, either misjudging normal deformation caused by the environment as damage, or missing early minor damage due to noise masking.

[0005] Secondly, the early warning methods are too simplistic and lack reliability. Traditional anomaly early warning systems often rely on a single technical approach: when using only model-driven methods, threshold settings are often based on theoretical design values ​​or empirical values, which cannot adapt to changes in the working conditions of bridges during service. This results in missed warnings if the warning threshold is too high, or false alarms if it is too low. When using only data-driven methods, model performance is highly dependent on the quality and quantity of samples. If historical anomaly samples are scarce, the model is prone to overfitting, has weak ability to identify unseen anomaly types, and is difficult to explain the early warning logic. This limits the trust and acceptance of early warning results by engineers.

[0006] Third, the accuracy of causal analysis is insufficient, and the location is ambiguous. When the monitoring system triggers an anomaly warning, existing methods are unable to accurately determine the cause of the anomaly and the location of the damage: on the one hand, there is a lack of deep integration of multi-source data, relying solely on sensor data for analysis without combining information such as video surveillance, manual inspection reports, and traffic management data, making it impossible to distinguish whether the anomaly is caused by structural damage or temporary external loads; on the other hand, the damage identification and location methods are singular, with traditional methods relying on finite element model correction or single dynamic characteristic analysis. The former requires high initial model accuracy, and the correction process is complex and computationally time-consuming, while the latter is difficult to accurately locate the specific location and extent of damage. For example, it can only determine that there is damage in a certain area of ​​the bridge, but cannot determine whether the problem is with a certain beam segment, a certain cable, or a certain support, resulting in a lack of precise direction for maintenance work and increasing operation and maintenance costs and time.

[0007] Therefore, there is an urgent need to develop a health monitoring technology for large bridges that can effectively separate environmental interference, integrate multiple early warning methods, and accurately analyze the causes of anomalies. This technology would address the shortcomings of existing methods, improve the effectiveness of monitoring data, the reliability of early warning results, and the accuracy of cause location, thus providing scientific and efficient technical support for the safe operation and maintenance of large bridges. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for anomaly early warning and cause analysis of health monitoring data of large bridge structures. This method addresses how to effectively clean the original monitoring data and separate environmental interference to obtain a "clean signal" that reflects the structure's own state; how to integrate model-driven and data-driven methods to reduce the false alarm rate and false negative rate of anomaly early warning; and how to combine multi-source information and multi-method damage identification to accurately determine the cause of anomalies and the location of damage.

[0009] To achieve the above objectives, the present invention employs the following technical means:

[0010] A method for early warning and cause analysis of anomalies in health monitoring data of large bridge structures includes the following steps:

[0011] Step S1: Data preprocessing and feature extraction. The raw data collected by the bridge monitoring system is cleaned, standardized, multi-domain feature extracted, and environmental factors separated to obtain a "net signal" reflecting the state of the structure itself.

[0012] Step S2: Anomaly warning. Based on the "net signal" obtained in step S1, anomaly warning is triggered using a combination of model-driven and data-driven methods.

[0013] Step S3: Cause analysis. After the warning is triggered in step S2, the cause and location of the anomaly are determined through data correlation analysis, damage identification and localization, and multi-source information fusion.

[0014] Preferably, the "data cleaning" in step S1 specifically includes:

[0015] Short-term missing raw data are completed using linear interpolation, while long-term missing raw data are completed using the mean interpolation of adjacent data from the same period.

[0016] Abnormal jump data is identified using the 3σ criterion, and the jump data is replaced with the moving average of the five time points before and after the data point.

[0017] Low-pass filtering is used to remove high-frequency interference, and then moving average smoothing is used to reduce noise.

[0018] Preferably, the "data normalization" in step S1 adopts the Z-score normalization method, which calculates based on the mean and standard deviation of the original data to convert the original data into standardized data that eliminates differences in units and magnitudes, and the normalized data is mapped to the interval [-2,2].

[0019] Preferably, the "multi-domain feature extraction" in step S1 specifically includes:

[0020] Temporal characteristics: Calculate the mean, variance, peak value, kurtosis, and skewness using a sampling period of 1-5 minutes;

[0021] Frequency domain characteristics: Perform a fast Fourier transform on the vibration acceleration data to extract the dominant frequency and the proportion of the energy of the first three dominant frequencies to the total energy;

[0022] Time-frequency domain characteristics: For non-stationary signals, wavelet transform or Hilbert-Huang transform based on db4 wavelet basis is used to extract instantaneous frequency and instantaneous amplitude.

[0023] Preferably, the "environmental factor separation" in step S1 adopts a combination of principal component analysis and multiple linear regression:

[0024] A multiple linear regression model is constructed using temperature, humidity, wind speed, and wind direction as input variables and strain or displacement data as output variables.

[0025] By using PCA to reduce the dimensionality of environmental variables, the principal components with a cumulative contribution rate of ≥90% are retained and substituted into the regression model to obtain the predicted values ​​of environmental effects.

[0026] The "net signal" is obtained by subtracting the predicted environmental effects from the original monitoring data.

[0027] Preferably, the "model-driven early warning" in step S2 includes:

[0028] Threshold warning: Dual thresholds are set for stress and deflection indicators. The first-level warning threshold is 80% of the design value, and the second-level warning threshold is 95% of the design value.

[0029] Statistical process control: Shewhart control charts, CUSUM control charts, or EWMA control charts are used.

[0030] Dynamic characteristic warning: Modal parameters are identified through environmental vibration testing every 1-3 months. A warning is triggered when the dominant frequency of a certain order decreases by more than 5% relative to the initial value or the mode shape MAC value is less than 0.85.

[0031] Preferably, the "data-driven early warning" in step S2 includes:

[0032] Unsupervised learning: K-Means clustering, isolation forest, or 3-layer autoencoder are employed;

[0033] Supervised learning: Using "normal / abnormal" as labels, support vector machines, random forests, or backpropagation neural networks are employed, and the model achieves a test set accuracy of ≥90%.

[0034] Time series analysis: ARIMA model or bidirectional LSTM network is used.

[0035] Preferably, the "damage identification and localization" in step S3 includes:

[0036] Model correction method: Based on the bridge finite element model, sensitivity analysis or genetic algorithm is used to determine damage when the element stiffness correction amount is greater than 20%.

[0037] Damage fingerprint analysis: using modal strain energy method or curvature mode method;

[0038] Machine learning method: Use CNN or GNN, and the model's localization accuracy on the test set is ≥85%, and the damage level error is ≤5%.

[0039] Preferably, the "multi-source information fusion" in step S3 is implemented through an expert system:

[0040] A knowledge base is built by integrating sensor data, environmental data, video surveillance data, and manual inspection reports;

[0041] The "if-else" reasoning rule is used.

[0042] The present invention has the following beneficial effects:

[0043] 1. Higher data quality, accurately reflecting the true state of the structure.

[0044] Through a multi-stage data processing workflow, various interfering factors are effectively removed, ensuring that the data used for analysis accurately reflects the structural condition of the bridge itself.

[0045] The data cleaning process can eliminate equipment noise, missing data, and abnormal jumps, avoiding interference from flawed raw data in subsequent analysis.

[0046] The environmental factor separation process can eliminate the influence of external environmental variables such as temperature and wind speed, and obtain a "net signal" that is only related to the state of the structure itself, providing a reliable basis for subsequent early warning and cause judgment.

[0047] 2. The early warning performance is more reliable, significantly reducing the risk of false alarms and missed alarms.

[0048] By integrating model-driven and data-driven early warning methods, the limitations of single methods can be overcome, and the methods can be adapted to the complex service conditions of bridges.

[0049] The model-driven approach relies on structural theory and statistical laws to ensure the stability and theoretical rationality of the early warning results;

[0050] Data-driven approaches can capture hidden anomalies in data, improving the sensitivity of identifying early minor anomalies and anomalies under complex operating conditions.

[0051] By using a multi-method cross-validation logic, the false positives of a single model are further reduced, thus avoiding security risks caused by missed reports or increased operation and maintenance costs due to false reports.

[0052] 3. More accurate causal analysis, clearly identifying the root cause and location of the anomaly.

[0053] By combining multi-dimensional damage identification methods with multi-source information fusion logic, we can achieve accurate differentiation of abnormality types and clear location of damage.

[0054] Multi-method damage identification cross-validates different technical approaches to ensure a more comprehensive assessment of structural damage and avoid the limitations of a single method.

[0055] Multi-source information fusion integrates various monitoring and on-site recording data, which can effectively distinguish whether the anomaly is caused by structural damage or temporary external loads, providing clear guidance for subsequent maintenance and avoiding blind handling.

[0056] 4. It has stronger engineering adaptability and is easier to implement in various scenarios.

[0057] The method design has flexible adjustment space, which can be adapted to the monitoring needs of different types of large bridges and reduce the threshold for engineering application.

[0058] The core processing and analysis parameters can be flexibly adjusted according to the bridge characteristics and service environment to meet the personalized monitoring needs in different scenarios.

[0059] It does not require special equipment and can be implemented based on existing bridge health monitoring systems, reducing additional equipment investment and facilitating its application throughout the entire life cycle of bridge operation and maintenance. Detailed Implementation

[0060] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0061] A method for early warning and cause analysis of anomalies in health monitoring data of large bridge structures includes the following steps:

[0062] Step S1: Data Preprocessing and Feature Extraction

[0063] The raw data collected by the bridge health monitoring system (including structural response data such as strain, displacement, and vibration acceleration, as well as environmental data such as temperature, humidity, wind speed, and wind direction) are processed to obtain the "net signal," which specifically includes:

[0064] Data cleaning

[0065] For raw data with short-term missing periods (≤5 min), linear interpolation was used for completion; for raw data with long-term missing periods (>5 min), the mean interpolation of adjacent data from the same period was used for completion.

[0066] Abnormal jump data is identified using the 3σ criterion, and the jump data is replaced with the moving average of the five time points before and after the data point.

[0067] High-frequency interference is removed by using a low-pass filter with a cutoff frequency of 5-10Hz, and then the noise of the equipment is further reduced by smoothing through a moving average with a window size of 10-20 sampling points.

[0068] Data normalization

[0069] The Z-score standardization method is used to convert the data into standardized data that eliminates differences in units and magnitudes based on the mean and standard deviation of the original data. The standardized data is mapped to the interval [-2,2], which facilitates subsequent cross-index analysis.

[0070] Multi-domain feature extraction

[0071] Time-domain characteristics: The mean, variance, peak value, kurtosis, and skewness are calculated according to a sampling period of 1-5 minutes to reflect the statistical distribution characteristics of the data;

[0072] Frequency domain characteristics: Fast Fourier Transform (FFT) is performed on vibration acceleration data to extract the dominant frequency and the proportion of the energy of the first three dominant frequencies to the total energy, reflecting the dynamic characteristics of the structure;

[0073] Time-frequency domain characteristics: For non-stationary vibration signals, wavelet transform or Hilbert-Huang transform (HHT) based on db4 wavelet basis is used to extract instantaneous frequency and instantaneous amplitude, capturing the time-varying pattern of the signal.

[0074] Environmental factors separation

[0075] A combination of principal component analysis (PCA) and multiple linear regression was used: temperature, humidity, wind speed, and wind direction were used as input variables, and strain or displacement data were used as output variables to construct a multiple linear regression model;

[0076] By using PCA to reduce the dimensionality of environmental variables, principal components with a cumulative contribution rate of ≥90% are retained and substituted into the regression model to obtain the predicted values ​​of environmental effects.

[0077] By subtracting the predicted environmental effects from the cleaned structural response data, a "net signal" is obtained that reflects only the state of the structure itself, eliminating environmental interference.

[0078] Step S2: Anomaly Warning

[0079] Based on the "net signal" obtained in step S1, a multi-level anomaly early warning system is implemented by integrating model-driven and data-driven methods, specifically including:

[0080] Model-driven early warning

[0081] Threshold warning: Two thresholds are set for key indicators such as stress and deflection. The first-level warning threshold is 80% of the structural design value (reminding attention), and the second-level warning threshold is 95% of the design value (triggering emergency response).

[0082] Statistical process control: Shewhart control charts (control limits are "mean ± 3 standard deviations"), cumulative sum control charts (CUSUM), or exponentially weighted moving average control charts (EWMA, weighting coefficients 0.1-0.3) are used to monitor whether the fluctuation of the "net signal" exceeds the normal statistical range.

[0083] Dynamic characteristic warning: Every 1-3 months, the structural modal parameters (dominant frequency, mode shape) are identified through environmental vibration testing. When the dominant frequency decreases by more than 5% relative to the initial service state, or the mode shape mode Assurance Criterion (MAC) value is less than 0.85, a dynamic characteristic abnormality warning is triggered.

[0084] Data-driven early warning

[0085] Unsupervised learning: K-Means clustering (3-5 clusters), isolated forest (100-200 isolated trees, warning when anomaly score ≥0.7) or 3-layer autoencoder (warning when reconstruction error >0.3) can identify anomalous data without labeled samples;

[0086] Supervised learning: Using "normal / abnormal" samples from historical monitoring data as labels, train the early warning model using support vector machine (RBF kernel function), random forest (100 decision trees), or BP neural network. The model's accuracy on the test set must be ≥90%.

[0087] Time series analysis: The ARIMA model (warning when the residual exceeds twice the standard deviation for 3 consecutive time steps) or the bidirectional LSTM network (warning when the input sequence length is 24, the number of hidden layer units is 64, and the relative error is >15%) is used to capture the time series trend anomalies of the "net signal".

[0088] When at least two of the model-driven and data-driven methods trigger an alert, it is determined to be a structural anomaly, and the process proceeds to step S3.

[0089] Step S3: Causal Analysis

[0090] After step S2 triggers an anomaly warning, the cause and location of the anomaly are determined through data correlation analysis, multi-method damage identification and localization, and multi-source information fusion. Specifically, this includes:

[0091] Data correlation analysis

[0092] Analyze the correlation between the abnormal "net signal" and concurrent environmental data (such as sudden changes in wind speed and temperature) and traffic data (such as the passage of overloaded vehicles) to preliminarily determine whether the anomaly is caused by temporary external loads.

[0093] The abnormal synchronization of "net signals" at different monitoring points is statistically analyzed. If multiple sensors in a certain area are synchronously abnormal, it is primarily determined to be a structural problem in that area.

[0094] Damage identification and localization

[0095] Model correction method: Based on the bridge finite element benchmark model, the model parameters are corrected by sensitivity analysis or genetic algorithm (population size 50, number of iterations 100). When the element stiffness correction amount is >20%, the element is determined to have damage.

[0096] Damage fingerprint analysis: The modal strain energy method (warning when the modal strain energy changes more than 25% relative to the normal state) or the curvature mode method (warning when the curvature mode mutation value exceeds twice that of the adjacent element) is used to locate damage by changes in structural dynamic characteristics;

[0097] Machine learning method: The damage localization model is trained using CNN (3×3 kernel size) or graph neural network (GNN). The model localization accuracy on the test set is ≥85%, and the damage degree error is ≤5%.

[0098] Multi-source information fusion

[0099] Multi-source information fusion is achieved through expert systems: integrating sensor "net signals", environmental data, video surveillance (such as crack propagation images) and manual inspection reports (such as support aging records) to build a knowledge base;

[0100] Using the "if-else" reasoning rule to determine the cause, examples include:

[0101] If the "dominant frequency decreases by more than 5%", the "strain increases by more than 30%", and the "crack width is greater than 0.3 mm", then it is judged as "concrete cracking damage".

[0102] If "frequency decreases but strain is not abnormal" and "wind speed > 15m / s", it is determined to be "temporary dynamic response abnormality caused by wind load".

[0103] If the response "returns to normal after a sudden increase" and "there is a record of overloaded vehicles passing through", it is determined to be "an abnormal instantaneous response caused by temporary overloading".

[0104] Example 1

[0105] Long-span cable-stayed bridge - Corrosion damage scenario of cable stays

[0106] 1.1 Bridge Types and Monitoring Background

[0107] This bridge is a key node on the inland mountainous expressway. It is a double-tower, double-cable-stayed bridge with a main span of 600m and has been in service for 10 years. The monitoring system focuses on collecting data on cable strain, bridge deck vibration acceleration, ambient temperature and humidity, and wind speed. Recently, abnormal strain fluctuations in some cables have been detected, requiring investigation into potential corrosion damage.

[0108] 1.2 Step S1: Data Preprocessing and Feature Extraction

[0109] Data cleaning: Cable strain data missing for a short period of time (≤5min) was completed by linear interpolation, and data missing for a long period of time (8min) was completed by the average of the same period of the previous two weeks; strain jump data were identified by the 3σ criterion and replaced with the moving average of the previous 5 time points; a 7Hz low-pass filter was used to remove high-frequency interference caused by mountain gusts, and then the data was smoothed by a moving average of 16 sampling points.

[0110] Data normalization: The strain and vibration acceleration data of the cables are normalized using Z-score and mapped to the [-2,2] interval.

[0111] Multi-domain feature extraction: Strain time-domain features (mean, variance, peak value) are calculated in 3-minute cycles; the first three dominant frequencies are extracted from the acceleration data using FFT; and the instantaneous frequency and amplitude of cable vibration are extracted using db4 wavelet transform.

[0112] Environmental factors separation: A multiple linear regression model is constructed with temperature, humidity, and wind speed as inputs and cable strain as output; after PCA dimensionality reduction, the two principal components with a cumulative contribution rate of 91% are retained and substituted into the model to obtain the predicted value of environmental effect, and the cable "net signal" is obtained after subtraction.

[0113] 1.3 Step S2: Anomaly Warning

[0114] Model-driven early warning: The first-level early warning threshold for cable stress is set at 80% of the design value, and the second-level threshold is 95%; EWMA control chart (weight 0.2) is used to monitor the "net signal"; environmental vibration tests are conducted every 2 months, and it was found that the main frequency corresponding to the No. 3 cable decreased by 6.2% compared with the initial value, triggering a dynamic characteristic early warning.

[0115] Data-driven early warning: 180 isolated trees (anomaly score ≥ 0.7) were used to analyze the "net signal". The anomaly score of cable #3 was 0.78. At the same time, bidirectional LSTM (input sequence length 24, number of hidden layer units 64) was used to predict strain with a relative error of 18.3%. The dual method triggered an early warning and determined it to be a structural anomaly.

[0116] 1.4 Step S3: Causal Analysis

[0117] Data correlation analysis: The wind speed during the same period was ≤6m / s, and there were no overloaded vehicles passing by, eliminating external load interference; two strain sensors near cable #3 showed synchronous abnormalities, locking down the area of ​​that cable.

[0118] Damage identification and localization: Based on the finite element model, the stiffness of cable element #3 was corrected by a genetic algorithm (population size 50, iteration number 100) with a correction amount of 26.5%; the modal strain energy method was adopted, and the modal strain energy of the cable body changed by 31% compared with the normal state; the localization result of the CNN model pointed to the middle and lower part of the cable body.

[0119] Multi-source information fusion: By integrating the abnormal results of the "net signal", the manual inspection record (a small amount of rust on the surface of the No. 3 cable) and the internal inspection images of the cable (local corrosion of the steel wire in the middle and lower part), the expert system determined that it was "corrosion damage in the middle and lower part of the cable".

[0120] 1.5 Implementation Results

[0121] It accurately locates the damaged cable and its specific position, guiding maintenance units to replace the damaged cable in a targeted manner, avoiding the need to replace the entire cable and reducing maintenance costs; the early warning misjudgment rate is 0, and the damage location deviation is ≤1.5m.

[0122] Example 2

[0123] Cross-sea suspension bridge - temporary dynamic response scenario caused by strong winds

[0124] 2.1 Bridge Types and Monitoring Background

[0125] This bridge is a core project of the cross-sea passage, a single-span suspension bridge with a main span of 1200m, and has been in service for 5 years. Coastal areas are often affected by the outer bands of typhoons, and the monitoring system focuses on collecting data on bridge deck displacement, tower and beam vibration acceleration, real-time wind speed and direction, and ocean humidity. After a strong wind, the system triggers an early warning, and it is necessary to determine whether the anomaly is due to structural damage.

[0126] 2.2 Step S1: Data Preprocessing and Feature Extraction

[0127] Data cleaning: Short-term missing (3min) displacement data were completed using linear interpolation, and long-term missing (12min) data were completed using the average of the same period of the previous week; displacement jump data were removed using the 3σ criterion and replaced with the moving average of the previous 5 time points; a 6Hz low-pass filter was used to remove high-frequency noise caused by wave impact, and then a moving average of 14 sampling points was used for smoothing.

[0128] Data normalization: Displacement and acceleration data are normalized using Z-score and mapped to the [-2,2] interval.

[0129] Multi-domain feature extraction: Calculate displacement time-domain features (mean, skewness, kurtosis) in 4-minute cycles; extract the first three dominant frequencies from acceleration data using FFT; extract the instantaneous frequency and amplitude of tower-beam vibration using HHT transform.

[0130] Environmental factors separation: A multiple linear regression model is constructed with wind speed, wind direction, and ocean humidity as inputs and bridge deck displacement as output; after PCA dimensionality reduction, the two principal components with a cumulative contribution rate of 93% are retained and substituted into the model to obtain the predicted value of environmental effect, and the "net signal" is obtained after subtraction.

[0131] 2.3 Step S2: Anomaly Warning

[0132] Model-driven early warning: The first-level early warning threshold for bridge deck deflection is set at 80% of the design value, and the second-level threshold is 95%; the CUSUM control chart is used to monitor the "net signal" and it is found that the displacement fluctuation exceeds the control limit; the dynamic characteristic test shows that the first-order main frequency drops by 4.8%, which is close to the early warning threshold.

[0133] Data-driven early warning: K-Means clustering (4 clusters) was used to analyze the "net signal". Some data points deviated from the normal clusters. The ARIMA model predicted the displacement. The residual exceeded twice the standard deviation for three consecutive time points. The two methods triggered an early warning and were judged as abnormal.

[0134] 2.4 Step S3: Causal Analysis

[0135] Data correlation analysis: The synchronous wind speed during the abnormal period reached 18 m / s (more than 3 times the daily average wind speed), and the wind direction made an angle of 30° with the bridge axis, which meets the conditions for triggering wind-induced vibration; there were no other synchronous abnormalities in the sensors, ruling out the preliminary suspicion of structural damage.

[0136] Damage identification and localization: The finite element model correction shows that the stiffness correction of each element is less than 10%; the modal strain energy change is less than 15%, and there is no obvious abrupt change in the curvature mode; the CNN model did not identify the damage area.

[0137] Multi-source information fusion: By integrating "net signal" fluctuation data, meteorological station wind records (the outer influence of the typhoon), and bridge surface video monitoring (no obvious structural deformation), the expert system determined it to be "temporary dynamic response anomaly caused by strong wind load".

[0138] 2.5 Implementation Results

[0139] Accurately distinguish between temporary loads and structural damage, avoid unnecessary bridge closures for maintenance, and ensure the passage of the cross-sea channel; the accuracy rate of anomaly cause judgment is 100%, and the response time is less than 1 hour.

[0140] Example 3

[0141] Urban long-span continuous beam bridges – scenarios of abnormal structural response due to bearing aging

[0142] 3.1 Bridge Types and Monitoring Background

[0143] This bridge is a river-crossing bridge on a main urban road, consisting of a 5-span continuous beam (main span 80m) and has been in service for 8 years. It experiences high daily traffic volume, including heavy-load vehicles. The monitoring system primarily collects data on beam strain, support displacement, traffic flow, and ambient temperature. Recently, an abnormal strain was detected in one span of the beam, and the cause needs to be investigated.

[0144] 3.2 Step S1: Data Preprocessing and Feature Extraction

[0145] Data cleaning: Short-term missing (2min) strain data were completed using linear interpolation, and long-term missing (6min) data were completed using the average of the previous 3 days; instantaneous strain jumps (non-abnormal jumps) caused by heavy vehicle traffic were removed using the 3σ criterion and replaced with the moving average of the previous 5 time points; traffic noise was removed using an 8Hz low-pass filter and then smoothed using a moving average of 18 sampling points.

[0146] Data normalization: The beam strain and support displacement data are normalized using Z-score and mapped to the [-2,2] interval.

[0147] Multi-domain feature extraction: Calculate the time-domain features of strain (mean, variance, peak value) in 1-minute cycles; extract the first three dominant frequencies from the beam vibration acceleration data using FFT; extract the instantaneous amplitude of strain using db4 wavelet transform.

[0148] Environmental factors separation: A multiple linear regression model is constructed with ambient temperature and traffic flow as inputs and beam strain as output; after PCA dimensionality reduction, one principal component with a cumulative contribution rate of 90% is retained and substituted into the model to obtain the predicted value of environmental effect, and the "net signal" is obtained after subtracting it.

[0149] 3.3 Step S2: Anomaly Warning

[0150] Model-driven early warning: The first-level early warning threshold for beam stress is set at 80% of the design value, and the second-level threshold is 95%; Shewhart control charts (control limits "mean ± 3 standard deviations") are used to monitor the "net signal". The strain of a certain span of beam exceeds the control limit multiple times; dynamic characteristic tests show that the main frequency drops by 5.2%, triggering an early warning.

[0151] Data-driven early warning: A random forest model (100 decision trees) was used with historical "normal / abnormal" samples as labels. The test set accuracy was 92%, and the prediction result for this span beam was "abnormal". The relative error of the bidirectional LSTM strain prediction was 16.7%. The dual-method early warning was triggered and the structure was judged to be abnormal.

[0152] 3.4 Step S3: Causal Analysis

[0153] Data correlation analysis: During the same period, there was no concentrated passage of heavy-load vehicles, and the ambient temperature fluctuation was <5℃, excluding external load and temperature interference; abnormal strain was concentrated at both ends of a certain span of the beam, corresponding to the support positions.

[0154] Damage identification and localization: The finite element model correction shows that the stiffness of the beam element corresponding to the supports at both ends of the span is corrected by 18%; using the curvature mode method, the curvature mode mutation value at both ends of the beam span is 2.3 times that of the adjacent element; the CNN model points to the beam area above the support.

[0155] Multi-source information fusion: By integrating the abnormal results of the "net signal", the manual inspection record (the seal of the span support is aging and there is slight displacement) and the support pressure monitoring data (uneven distribution of bearing capacity), the expert system determined that "the beam stress response is abnormal due to the aging of the support".

[0156] 3.5 Implementation Results

[0157] It accurately locates structural anomalies caused by aging bearings, guiding maintenance units to replace aging bearings in a timely manner to prevent further deformation of the beam; the accuracy rate of damage correlation cause judgment is 100%, and maintenance costs are reduced by 30%.

[0158] The examples provided in this invention are not intended to limit the implementation. Those skilled in the art will recognize that various variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations, and any obvious variations or modifications derived therefrom are still within the scope of this invention.

Claims

1. A method for early warning and cause analysis of abnormalities in health monitoring data of large bridge structures, characterized in that, Includes the following steps: Step S1: Data preprocessing and feature extraction. The raw data collected by the bridge monitoring system is cleaned, standardized, multi-domain feature extracted, and environmental factors separated to obtain a "net signal" reflecting the state of the structure itself. Step S2: Anomaly warning. Based on the "net signal" obtained in step S1, anomaly warning is triggered using a combination of model-driven and data-driven methods. Step S3: Cause analysis. After the warning is triggered in step S2, the cause and location of the anomaly are determined through data correlation analysis, damage identification and localization, and multi-source information fusion.

2. The method for abnormal early warning and cause analysis of health monitoring data of large bridge structures according to claim 1, characterized in that, The "data cleaning" in step S1 specifically includes: Short-term missing raw data are completed using linear interpolation, while long-term missing raw data are completed using the mean interpolation of adjacent data from the same period. Abnormal jump data is identified using the 3σ criterion, and the jump data is replaced with the moving average of the five time points before and after the data point. Low-pass filtering is used to remove high-frequency interference, and then moving average smoothing is used to reduce noise.

3. The method for abnormal early warning and cause analysis of health monitoring data of large bridge structures according to claim 1, characterized in that, The "data normalization" in step S1 adopts the Z-score normalization method, which calculates the original data based on the mean and standard deviation of the original data, converting the original data into standardized data that eliminates differences in units and magnitudes, and the normalized data is mapped to the interval [-2,2].

4. The method for abnormal early warning and cause analysis of health monitoring data of large bridge structures according to claim 1, characterized in that, The "multi-domain feature extraction" in step S1 specifically includes: Temporal characteristics: Calculate the mean, variance, peak value, kurtosis, and skewness using a sampling period of 1-5 minutes; Frequency domain characteristics: Perform a fast Fourier transform on the vibration acceleration data to extract the dominant frequency and the proportion of the energy of the first three dominant frequencies to the total energy; Time-frequency domain characteristics: For non-stationary signals, wavelet transform or Hilbert-Huang transform based on db4 wavelet basis is used to extract instantaneous frequency and instantaneous amplitude.

5. The method for abnormal early warning and cause analysis of health monitoring data of large bridge structures according to claim 1, characterized in that, The "environmental factor separation" in step S1 adopts a combination of principal component analysis and multiple linear regression: A multiple linear regression model is constructed using temperature, humidity, wind speed, and wind direction as input variables and strain or displacement data as output variables. By using PCA to reduce the dimensionality of environmental variables, the principal components with a cumulative contribution rate of ≥90% are retained and substituted into the regression model to obtain the predicted values ​​of environmental effects. The "net signal" is obtained by subtracting the predicted environmental effects from the original monitoring data.

6. The method for abnormal early warning and cause analysis of health monitoring data of large bridge structures according to claim 1, characterized in that, The "model-driven early warning" in step S2 includes: Threshold warning: Dual thresholds are set for stress and deflection indicators. The first-level warning threshold is 80% of the design value, and the second-level warning threshold is 95% of the design value. Statistical process control: Shewhart control charts, CUSUM control charts, or EWMA control charts are used. Dynamic characteristic warning: Modal parameters are identified through environmental vibration testing every 1-3 months. A warning is triggered when the dominant frequency of a certain order decreases by more than 5% relative to the initial value or the mode shape MAC value is less than 0.

85.

7. The method for abnormal early warning and cause analysis of health monitoring data of large bridge structures according to claim 1, characterized in that, The "data-driven early warning" in step S2 includes: Unsupervised learning: K-Means clustering, isolation forest, or 3-layer autoencoder are employed; Supervised learning: Using "normal / abnormal" as labels, support vector machines, random forests, or backpropagation neural networks are employed, and the model achieves a test set accuracy of ≥90%. Time series analysis: ARIMA model or bidirectional LSTM network is used.

8. The method for abnormal early warning and cause analysis of health monitoring data of large bridge structures according to claim 1, characterized in that, The "damage identification and localization" in step S3 includes: Model correction method: Based on the bridge finite element model, sensitivity analysis or genetic algorithm is used to determine damage when the element stiffness correction amount is greater than 20%. Damage fingerprint analysis: using modal strain energy method or curvature mode method; Machine learning method: Use CNN or GNN, and the model's localization accuracy on the test set is ≥85%, and the damage level error is ≤5%.

9. The method for abnormal early warning and cause analysis of health monitoring data of large bridge structures according to claim 1, characterized in that, The "multi-source information fusion" in step S3 is achieved through an expert system: A knowledge base is built by integrating sensor data, environmental data, video surveillance data, and manual inspection reports; The "if-else" reasoning rule is used.