Special-shaped steel arch bridge health monitoring and early warning system based on multi-source data fusion

Abstract

The invention relates to the technical field of arch bridge monitoring, and particularly discloses a deformed steel arch bridge health monitoring and early warning system based on multi-source data fusion, which comprises a data acquisition module, a data preprocessing module, a multi-source data fusion module, an anomaly detection and mode recognition module and a health assessment and early warning module, a visual detection and image analysis module and a risk prediction and decision support module; by adopting the adaptive sampling method, the sampling frequency can be dynamically adjusted according to the dynamic change of the bridge, the load state and the importance of the sensor, so that the sampling frequency is reduced when the bridge is in a stable state, and the collection of redundant data is reduced; the redundant data generated when the data acquisition frequency is too high is reduced through self-adaptive sampling, so that the requirement for the storage space and the bandwidth requirement for data transmission are reduced, and the calculation efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of arch bridge monitoring, and in particular relates to a health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion. Background Art

[0002] In the field of bridge engineering, special-shaped steel arch bridges, as complex and critical transportation facilities, require the design and implementation of structural health monitoring and early warning systems. Traditional health monitoring systems typically rely on a single sensor data source, such as accelerometers, strain gauges, or displacement sensors, to monitor the bridge's dynamic response and structural condition. However, this single-source monitoring approach has significant limitations.

[0003] A single data source may not fully reflect the overall health of a bridge. Due to the complexity and diversity of bridge structures, different parts and components may exhibit different mechanical behaviors and response characteristics. Therefore, relying solely on a single type of sensor may not capture all critical structural changes, leading to inaccurate or missed health assessments.

[0004] Existing data collection methods often lack flexibility. Conventional sensor sampling frequencies are typically fixed and unable to adapt to the bridge's actual dynamic characteristics and load conditions. This results in the collection of large amounts of redundant data when the bridge is stable, increasing the consumption of storage and computing resources. However, when the bridge is heavily loaded or experiencing anomalies, sufficiently accurate and detailed real-time data may not be obtained in a timely manner, affecting the monitoring system's responsiveness and real-time performance.

[0005] To this end, the inventors proposed a health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion to solve the above problems. Summary of the Invention

[0006] The purpose of the present invention is to provide a health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions: A health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion, including: The data acquisition module is used to obtain various damage parameter data such as displacement, settlement, strain, and cracks caused by tension, compression, and bending during the use of the bridge through multiple sensors; A data preprocessing module is used to perform denoising, data cleaning and missing value filling on the disease parameter data to obtain denoised data; A multi-source data fusion module is used to fuse the denoised data and perform comprehensive processing on the data through weighted projection pursuit method and fuzzy inference system to obtain fused data; An anomaly detection and pattern recognition module is used to perform anomaly detection on the fused data using a deep generative model and a density estimation method, identify abnormal patterns of the bridge structure in combination with finite element calculation results, and obtain anomaly detection results; A health assessment and early warning module, configured to assess the health of the bridge based on the abnormality detection results and modal analysis, and generate early warning information; A visual inspection and image analysis module, which collects images of bridge surfaces and uses convolutional neural networks to detect cracks in the images, providing visual information about the structural health. The risk prediction and decision support module is used to combine the warning information and the visualization information to perform bridge risk prediction and obtain prediction results.

[0008] Preferably, the data acquisition module adopts an adaptive sensor sampling method to automatically adjust the sensor data acquisition frequency according to the dynamic characteristics of the bridge, thereby optimizing the timeliness and accuracy of the data; The adaptive weighted sampling method dynamically adjusts the sampling frequency of each sensor according to the actual needs of bridge health monitoring, and performs adaptive optimization based on the signal importance of each sensor and environmental changes. The formula of the adaptive weighted sampling method is: ; Where f(t): the final integrated sampling signal; wi(t): The adaptive weight of the i-th sensor, which is dynamically adjusted according to the signal’s rate of change, importance, and stability; xi(t): sampling data of the i-th sensor at time t; n: total number of sensors; t: current time.

[0009] Preferably, the data preprocessing module uses a hybrid denoising method of wavelet transform and adaptive filtering to extract the high-frequency components of the signal and remove the noise to obtain denoised data; The expression of the hybrid denoising method of wavelet transform and adaptive filtering is: ; in : denoised signal; ak: coefficient of wavelet transform, representing the weight of the signal in different frequency bands; ψk(t): wavelet basis function, which decomposes the signal at different scales; β: weight of the adaptive filter, used to adjust the strength of noise removal; filter(x(t)): The result of processing the original signal x(t) by the adaptive filter.

[0010] Preferably, the weighted projection pursuit method and fuzzy reasoning perform dimensionality reduction on high-dimensional data through a projection pursuit algorithm, and fuzzy reasoning performs a comprehensive evaluation on the reduced-dimensional data. The expressions of the weighted projection pursuit method and fuzzy reasoning are: ; Where S: comprehensive assessment of health status indicators; wi: the weight of the i-th sensor, which is dynamically adjusted based on the importance of the sensor; proji(xi): projection function for reducing the dimension of the input data xi of the i-th sensor; n: number of sensors.

[0011] Preferably, the anomaly detection and pattern recognition module is based on a variational autoencoder model and density estimation method, which models normal data through a generative model and matches it with real-time data to identify potential abnormal patterns; The expression of the variational autoencoder model and density estimation method is: ; LVAE: loss function of variational autoencoder; X: reconstructed data; z: latent variable; E: represents the expectation of the latent variable z; Eq[log p(x|z)] represents the expected value of the log-likelihood of reconstructing the data x under the distribution of the latent variable z; q(z|x): latent variable distribution output by the encoder; p(x|z): the output distribution generated by the decoder; p(z): prior distribution, representing the prior distribution of the latent variable z; DKL: Kullback-Leibler divergence, which measures the difference between the distribution of the latent variable and the prior distribution.

[0012] Preferably, the health assessment and early warning module adopts a method combining integrated learning and fuzzy logic reasoning to integrate multiple health assessment models to generate a health early warning index for the bridge and issue early warning information based on the assessment results; The expression of the method combining ensemble learning with fuzzy logic reasoning is: ; ; Where y: comprehensive health assessment output; fe: ensemble learning model, combining the prediction results of multiple single models; x1,x2,...,xn: input data for each evaluation model; : Early warning index generated by fuzzy reasoning; μ: Fuzzy inference function used to convert health assessment results into warning levels.

[0013] Preferably, the risk prediction and decision support module uses a long short-term memory network combined with a Bayesian network to model time series data, predict future health change trends and perform risk reasoning to generate decision support information; The expression of the long short-term memory network combined with the Bayesian network is: ; ; Where ht: hidden state in LSTM, representing the memory information of time step t; h t-1 Represents the hidden state at the previous moment; σ: activation function, sigmoid function, used to map the input to the [0,1] interval; Wh, Wx: LSTM weight matrix, acting on hidden state and input data respectively; xt: input data of the current time step; n: number of sensors; p(r|x): risk probability given observation data x in the Bayesian network; p(x|r): likelihood of observation data under risk state; p(r): prior probability of risk; p(x): marginal probability of observing data x.

[0014] Preferably, the system further comprises a data visualization and reporting module for visually displaying the prediction results in the form of charts or reports; The system sets up differentiated sensor layout schemes and data processing algorithms based on the special geometric configuration and stress characteristics of special-shaped steel arch bridges, which is different from the monitoring methods of conventional bridges.

[0015] Preferably, the data visualization and reporting module adopts a dynamic interactive data visualization algorithm, combined with an incremental learning model to update charts and reports in real time, ensuring the real-time and accuracy of data display.

[0016] Preferably, the data processing algorithm includes a multimodal data fusion method, a cave volume calculation method and a permeability coefficient prediction method.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention adopts an adaptive sampling method, which can dynamically adjust the sampling frequency according to the dynamic changes of the bridge, the load status and the importance of the sensor, thereby reducing the sampling frequency when the bridge is in a stable state, reducing the collection of redundant data, and saving storage and computing resources: Adaptive sampling reduces the redundant data generated when the data collection frequency is too high, thereby reducing the demand for storage space and the bandwidth requirement for data transmission, and improving computing efficiency; when the bridge load is large, increasing the sampling frequency of key sensors helps to obtain more accurate real-time data, and improves the response speed and real-time performance of the monitoring system.

[0018] (2) The multi-source data fusion in the present invention enhances accuracy and robustness. The complementary characteristics of different sensors enhance the robustness of the system, enabling the system to more accurately evaluate abnormal or uncertain data, reducing the misjudgment that may be caused by single sensor data; data fusion helps to reduce noise interference and avoid distortion of a single data source in a high-noise environment, thereby improving the accuracy of the health assessment model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a block diagram of the composition of a health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion of the present invention. DETAILED DESCRIPTION

[0020] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0021] Example 1: Please refer to Figure 1 As shown in the figure, a health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion includes:

[0022] A data acquisition module is used to obtain data on various damage parameters such as displacement, settlement, strain, and cracks caused by tension, compression, and bending during the use of the bridge through multiple sensors, including vibration sensors, strain gauges, accelerometers, and temperature sensors; The data acquisition module uses an adaptive sensor sampling method to automatically adjust the sensor data acquisition frequency according to the dynamic characteristics of the bridge, thereby optimizing the timeliness and accuracy of the data; The adaptive weighted sampling method dynamically adjusts the sampling frequency of each sensor according to the actual needs of bridge health monitoring, and performs adaptive optimization based on the signal importance of each sensor and environmental changes. The formula of the adaptive weighted sampling method is: ; Where f(t): the final integrated sampling signal; wi(t): The adaptive weight of the i-th sensor, which is dynamically adjusted according to the signal’s rate of change, importance, and stability; xi(t): sampling data of the i-th sensor at time t; n: total number of sensors; t: current time.

[0023] This method dynamically adjusts the sampling frequency based on the quality and importance of each sensor's real-time data, enabling the system to perform high-frequency sampling of critical areas while reducing the sampling frequency of stable areas, thereby optimizing data collection efficiency and conserving resources. Through dynamic weighting and adaptive sampling strategies, it avoids data redundancy in traditional methods and improves the bridge health monitoring system's ability to respond to emergencies or extreme situations. A data preprocessing module is used to perform denoising, data cleaning and missing value filling on the disease parameter data to obtain denoised data; The data preprocessing module uses a hybrid denoising method of wavelet transform and adaptive filtering to extract the high-frequency components of the signal and remove the noise to obtain denoised data; The expression of the hybrid denoising method of wavelet transform and adaptive filtering is: ; in : denoised signal; ak: coefficient of wavelet transform, representing the weight of the signal in different frequency bands; ψk(t): wavelet basis function, which decomposes the signal at different scales; β: weight of the adaptive filter, used to adjust the strength of noise removal; filter(x(t)): The result of processing the original signal x(t) by the adaptive filter.

[0024] This method uses wavelet transform to decompose the signal into different frequency bands and effectively separate low-frequency useful signals from high-frequency noise. An adaptive filter further filters the noise, enhancing the signal's usability. This method significantly improves signal quality and reduces errors introduced by noise, making it particularly suitable for processing vibrations and weak signals commonly found in bridge health monitoring. A multi-source data fusion module is used to fuse the denoised data and perform comprehensive processing on the data through weighted projection pursuit method and fuzzy inference system to obtain fused data; The weighted projection pursuit method and fuzzy reasoning reduce the dimensionality of high-dimensional data through the projection pursuit algorithm, and the fuzzy reasoning performs a comprehensive evaluation on the reduced dimensionality data. The expression of the weighted projection pursuit method and fuzzy reasoning is: ; Where S: comprehensive assessment of health status indicators; wi: the weight of the i-th sensor, which is dynamically adjusted based on the importance of the sensor; proji(xi): projection function for reducing the dimension of the input data xi of the i-th sensor; n: number of sensors.

[0025] The WPP method reduces high-dimensional data to a low-dimensional space, reducing data processing complexity and extracting useful pattern features. The fuzzy inference system makes comprehensive decisions based on the fuzziness of the data and provides health status assessment results. This method can maintain high accuracy and robustness in the fusion of multi-source heterogeneous data, especially when the data uncertainty is high. Fuzzy inference can effectively process various uncertain and ambiguous data. An anomaly detection and pattern recognition module is used to perform anomaly detection on the fused data through a deep generative model and a density estimation method, identify abnormal patterns of the bridge structure, and obtain anomaly detection results; The anomaly detection and pattern recognition module is based on the variational autoencoder model and density estimation method. It models normal data through a generative model and matches it with real-time data to identify potential abnormal patterns. The expression of the variational autoencoder model and density estimation method is: ; LVAE: loss function of variational autoencoder; X: reconstructed data; z: latent variable; E: represents the expectation of the latent variable z; Eq[log p(x|z)] represents the expected value of the log-likelihood of reconstructing the data x under the distribution of the latent variable z; q(z|x): latent variable distribution output by the encoder; p(x|z): the output distribution generated by the decoder; p(z): prior distribution, representing the prior distribution of the latent variable z; DKL: Kullback-Leibler divergence, which measures the difference between the distribution of the latent variable and the prior distribution.

[0026] A health assessment and early warning module is used to assess the health status of the bridge based on the abnormality detection results and generate early warning information; The health assessment and early warning module uses a method that combines ensemble learning with fuzzy logic reasoning to integrate multiple health assessment models to generate a health early warning index for the bridge and issue early warning information based on the assessment results. The expression of the method combining ensemble learning with fuzzy logic reasoning is: ; ; Where y: comprehensive health assessment output; fe: ensemble learning model, combining the prediction results of multiple single models; x1,x2,...,xn: input data for each evaluation model; : Early warning index generated by fuzzy reasoning; μ: Fuzzy inference function used to convert health assessment results into warning levels.

[0027] By integrating multiple models and combining them with a fuzzy logic system, the system comprehensively considers multiple factors to determine the health status of the bridge, thereby generating reliable early warning information. Ensemble learning can improve the accuracy of health assessments, while fuzzy reasoning enhances the system's adaptability to uncertainty and complexity. This combination enables the system to provide accurate and effective health assessments even in complex environments. A visual inspection and image analysis module, which collects images of bridge surfaces and uses convolutional neural networks to detect cracks in the images, providing visual information about the structural health. A risk prediction and decision support module is used to combine the warning information and visualization information to perform bridge risk prediction and obtain prediction results; The risk prediction and decision support module uses a long short-term memory network (LSTM) combined with a Bayesian network to model time series data, predict future health trends and perform risk reasoning to generate decision support information; The expression of the long short-term memory network combined with the Bayesian network is: ; Where ht: hidden state in LSTM, representing the memory information of time step t; h t-1 Represents the hidden state at the previous moment; σ: activation function, sigmoid function, used to map the input to the [0,1] interval; Wh, Wx: LSTM weight matrix, acting on hidden state and input data respectively; xt: input data of the current time step; n: number of sensors; p(r|x): risk probability given observation data x in the Bayesian network; p(x|r): likelihood of observation data under risk state; p(r): prior probability of risk; p(x): marginal probability of observing data x.

[0028] LSTM can effectively capture the time series dependencies of bridge health changes, thereby predicting future changes in health status; Bayesian networks combine health data for risk reasoning and generate risk prediction information; this method can combine the characteristics of time series to make more accurate health trend predictions, and provide risk warnings and decision support based on the prediction results, enhancing the system's preventive function.

[0029] Specifically, the system further includes a data visualization and reporting module for visually displaying the prediction results in the form of charts or reports; The data visualization and reporting module adopts a dynamic interactive data visualization algorithm combined with an incremental learning model to update charts and reports in real time, ensuring the real-time and accuracy of data display.

[0030] The data processing algorithm includes a multimodal data fusion method, a cave volume calculation method and a permeability coefficient prediction method; 1) During shield tunnel construction, the data collected by various sensors (geology, seepage, surrounding rock deformation, etc.) is heterogeneous and multi-source. Directly using this data to identify anomalies may be subject to noise interference. Therefore, data fusion technology and machine learning models are needed to extract key features and identify anomalies. Using multimodal data fusion and anomaly detection methods, we assume that different types of sensor data sets are: X={Xg,Xh,Xs,Xm} Where: Xg: geological structure data (obtained in step S1) Xh: Groundwater seepage path data (obtained in step S3) Xs: surrounding rock deformation and stress data (obtained in step S4) Xm: Mechanical parameters of the shield machine, such as thrust, torque, and earth pressure (obtained in step S2) Use a time series data fusion model based on LSTM (Long Short-Term Memory Network) or Transformer: ; Where ht: hidden state in LSTM, representing the memory information of time step t; h t-1 Represents the hidden state at the previous moment; σ: activation function, sigmoid function, used to map the input to the [0,1] interval; Wh, Wx: LSTM weight matrix, acting on hidden state and input data respectively; xt: input data of the current time step; n: number of sensors; Then, the fused features are fed into an anomaly detection model, such as an anomaly detection based on an autoencoder, to obtain: ; Where: fθ(⋅) is the encoding-decoding process of AutoEncoder; L is the reconstruction error. If L exceeds the threshold, it is judged as abnormal; This paper constructs a shield construction environment anomaly detection model through a multimodal data fusion method based on deep learning. It uses an LSTM-Transformer architecture for time series modeling and combines it with an AutoEncoder reconstruction error detection mechanism to effectively identify abnormal geological conditions, seepage mutations, and surrounding rock deformation risks, thereby improving the reliability of intelligent monitoring of shield construction. 2) Mathematical modeling for cave volume calculation The distribution of rock caves follows the Gaussian process model, which is: ; Among them, r=(x,y,z) is the spatial coordinate; μ = (μx, μy, μz) is the center position of the cave; Σ is the covariance matrix, describing the uncertainty of cave morphology; The cave volume can be estimated by probability integration: ;; Where, Ω is the spatial area in front of the shield; P(x,y,z) reflects the possibility of cave formation, and the integral result is the expected volume of the cave; This paper proposes a cave volume prediction method based on probability distribution modeling. This method uses Gaussian process regression (GPR) to model the probability distribution of caves and calculates the expected cave volume through triple integrals. Compared with traditional geometric calculation methods, this method effectively reduces the impact of geological uncertainty on calculation accuracy and improves the ability to intelligently assess geological conditions ahead of shield construction. 3) Machine Learning Method for Predicting Water Permeability Coefficient The factors affecting the permeability coefficient K include: Formation lithology R Fissure development degree F Groundwater level H Porosity ϕ Water flow rate v The permeability coefficient can be predicted by the machine learning model: ; Among them: GBDT model is used to train the nonlinear regression function f The training data set can come from drilling tests, formation monitoring data, and seepage numerical simulation results; This paper uses a gradient boosted decision tree (GBDT) model, combined with multi-source data (stratum lithology, fracture development, groundwater level, etc.), to develop an intelligent prediction model for water permeability. By training on historical data, this method can effectively predict water permeability in shield tunneling areas, providing scientific guidance for tunnel excavation and improving water hazard prevention and control capabilities.

[0031] As can be seen from the above, the adaptive sampling method can dynamically adjust the sampling frequency according to the dynamic changes of the bridge, the load status and the importance of the sensor, thereby reducing the sampling frequency when the bridge is in a stable state, reducing the collection of redundant data, and saving storage and computing resources: Adaptive sampling reduces the redundant data generated when the data collection frequency is too high, thereby reducing the storage space requirements and the bandwidth requirements for data transmission, and improving computing efficiency; when the bridge load is heavy, increasing the sampling frequency of key sensors helps to obtain more accurate real-time data, improving the response speed and real-time performance of the monitoring system; Multi-source data fusion enhances accuracy and robustness: The complementary characteristics of different sensors enhance the robustness of the system, enabling more accurate assessment of abnormal or highly uncertain data and reducing potential misjudgments caused by single sensor data. Data fusion also helps reduce noise interference and avoids distortion of a single data source in high-noise environments, thereby improving the accuracy of the health assessment model. The health monitoring and early warning system for special-shaped steel arch bridges, based on multi-source data fusion, significantly improves the accuracy, sensitivity, and real-time response capabilities of bridge health monitoring by applying advanced theories such as adaptive sampling, data fusion, anomaly detection, and fuzzy reasoning. Furthermore, the system's scalability, flexibility, and diverse decision support capabilities enable it to not only address current bridge health monitoring tasks but also adapt to evolving needs and technological advancements in the future.

[0032] Example 2: Bridge health monitoring based on adaptive sampling and data fusion: Application scenario: A city's special-shaped steel arch bridges require long-term health monitoring. The system integrates multiple sensors (such as vibration sensors, strain gauges, and temperature sensors) and uses an adaptive sampling method to optimize data collection frequency, reduce redundancy, and improve data response speed.

[0033] System parameters: Sensor configuration: Vibration sensor: sampling frequency 100 Hz, arranged at the support points of the bridge.

[0034] Strain gauges: sampling frequency 50 Hz, arranged in the middle and arch structures of the bridge.

[0035] Temperature sensor: sampling frequency 10Hz, arranged on the bridge deck and support.

[0036] Accelerometer: sampling frequency 200 Hz, arranged at the vibration analysis point of the bridge.

[0037] Adaptive sampling parameters: The sampling frequency of the sensors is dynamically adjusted based on the health of the bridge and environmental changes. The sampling frequency of each sensor ranges from 20Hz to 300Hz.

[0038] Adaptive Sampling: Assume that the vibration sensor collects data at the stress point of the bridge. When the bridge is in a stable state, the sampling frequency of the vibration sensor is 100 Hz.

[0039] When the bridge load is heavy (such as high traffic volume), the sampling frequency of the vibration sensor is automatically increased to 200Hz to capture higher-frequency vibration signals.

[0040] Data Fusion: In the system, data from vibration sensors, strain gauges, and temperature sensors are fused using weighted projection pursuit (WPP). Dynamic weighting increases the weight of the vibration sensor when the load is greater, while the weights of the strain gauge and temperature sensors are adjusted accordingly.

[0041] Health assessment and early warning: A fuzzy inference system is used to evaluate the health of the fused data and calculate a health index. If the health index falls below a set threshold (e.g., 0.5), an alert is triggered.

[0042] For example, when the vibration sensor's data sampling frequency is 200 Hz, the data volatility increases, and the system will detect that the bridge may be unstable and immediately issue an early warning.

[0043] Specific parameter data and calculation process: Vibration sensor (Sensor 1): Sampling frequency: 100 Hz, data volatility: standard deviation 0.2.

[0044] When the load is large, the sampling frequency is increased to 200 Hz, and the data volatility is: standard deviation 0.4.

[0045] Strain gauge (sensor 2): Sampling frequency: 50 Hz, data volatility: standard deviation 0.05.

[0046] When the load changes, the data volatility increases and the sampling frequency increases.

[0047] Accelerometer (Sensor 3): Sampling frequency: 200 Hz, data volatility: standard deviation 0.8.

[0048] Calculate the fused health assessment value: S=0.8·proj1(x1)+0.5·proj2(x2)+0.7·proj3(x3) in: proj1(x1)=0.4 (based on the vibration sensor projection result).

[0049] proj2(x2)=0.3 (based on strain gauge projection results).

[0050] proj3(x3)=0.5 (based on accelerometer projection results).

[0051] The health assessment value is: S=0.8×0.4+0.5×0.3+0.7×0.5=0.32+0.15+0.35=0.82 If the evaluation value S is less than 0.5, the system will trigger an early warning, indicating that there is an abnormality in the bridge.

[0052] As can be seen from the above, the system can reduce redundant data collection by dynamically adjusting the sampling frequency, while improving monitoring accuracy when the bridge is under heavy load. Data fusion provides a more accurate health assessment.

[0053] This method improves the sensitivity and efficiency of system response, can monitor the health of the bridge in real time, and provide early warning when needed, helping managers take timely maintenance measures.

[0054] Example 3: Anomaly Detection and Health Assessment Based on VAE and Density Estimation Application Scenario: A special-shaped steel arch bridge has been exposed to various environmental changes and traffic loads over its long-term operation. To monitor the bridge's health and promptly alert users of potential failures, the system employed an anomaly detection and health assessment method based on a variational autoencoder (VAE) and density estimation.

[0055] System parameters: Sensor configuration: Vibration sensors: sampling frequency 100 Hz, arranged at different locations on the bridge.

[0056] Strain gauges: Sampling frequency 50 Hz, arranged at key support positions of the bridge.

[0057] Temperature sensor: sampling frequency 10Hz, arranged on the bridge deck and support.

[0058] VAE model parameters: Encoder: 3-layer fully connected network with 64 hidden layer nodes.

[0059] Decoder: A 3-layer fully connected network with the number of output nodes being the dimension of the sensor data.

[0060] Activation function: ReLU, optimizer: Adam, learning rate 0.001.

[0061] Density estimation parameters: Kernel function: Gaussian kernel, bandwidth 0.2.

[0062] Data collection and calculation process: Data collection and preprocessing: Data is collected from vibration sensors, strain gauges, and temperature sensors, and the data is pre-processed and denoised before entering the VAE model.

[0063] VAE training: The VAE model is trained using historical normal data to learn the latent representation and reconstruction process of the normal state of the bridge.

[0064] Anomaly Detection: Input the real-time data into the trained VAE model and calculate the reconstruction error; ; Where x is the input data, is the result of model reconstruction.

[0065] If the reconstruction error exceeds the set threshold (for example, 0.3), the anomaly detection stage is entered, and the density estimation method is used to calculate its probability density p(r|x).

[0066] Health Assessment: If the density estimation result p(r|x) is less than the threshold value of 0.2, it means that the data deviates from the normal state, and the system generates a health assessment report and issues an early warning.

[0067] Vibration sensor (Sensor 1): Normal data: mean 0, standard deviation 0.2.

[0068] Reconstruction error during anomaly: 0.5.

[0069] Density estimation result: p(r|x)=0.1 (below the threshold of 0.2, indicating anomaly).

[0070] Strain gauge (sensor 2): Normal data: mean 0, standard deviation 0.05.

[0071] Reconstruction error during anomaly: 0.08.

[0072] Density estimation result: p(r|x)=0.15 (indicating slight anomaly).

[0073] Health assessment and early warning: The calculated reconstruction error and density estimation value show that the reconstruction error of the vibration sensor exceeds the threshold and the density estimation value is less than 0.2. The system detects a serious anomaly and immediately triggers an early warning.

[0074] As can be seen from the above, VAE can effectively learn the normal patterns of bridge health, and density estimation can effectively identify abnormal data. This method can detect potential structural problems by detecting subtle changes in the bridge's operating status.

[0075] This method improves the reliability and early warning capabilities of bridge monitoring systems through accurate anomaly detection and health assessment, helping bridge managers take timely actions to prevent potentially catastrophic events.

[0076] Example 4: This embodiment uses a special-shaped steel arch bridge in a certain city as an example to illustrate how to use this system for comprehensive health monitoring and early warning. It specifically describes the sensor configuration, collection parameters, data calculation process, and actual results.

[0077] 1. System composition and sensor layout Sensor configuration: Displacement sensor: installed on piers, main beams and supports to measure micro-displacement in real time.

[0078] Sampling frequency: 100Hz (when the bridge is stable), up to 200Hz (adaptive adjustment when the load increases sharply); Accuracy: ±0.1mm.

[0079] Settlement monitor: installed at the bottom of the bridge to measure structural settlement.

[0080] Sampling frequency: 10Hz, accuracy: ±0.05mm.

[0081] Strain gauges: Installed on key load-bearing components (such as arch ribs and connecting plates) to detect strain caused by tension and compression.

[0082] Sampling frequency: 50Hz, accuracy: ±0.002με.

[0083] Accelerometers and vibration sensors: deployed at key locations of supports and arch ribs to monitor vibration modes.

[0084] Sampling frequency: 200Hz, accuracy: ±0.01g.

[0085] High-definition camera system: regularly collects images of the bridge deck, arch ribs and steel plate surfaces for crack and erosion detection.

[0086] Layout description: Because special-shaped steel arch bridges differ significantly from conventional bridges in arch rib design and plate thickness distribution, the system sets up a dense layout scheme in the arch and support areas, and the corresponding sensor sampling frequency and weight are relatively high to capture local stress and deformation details.

[0087] 2. Data Collection and Adaptive Sampling Collection method: An adaptive sensor sampling method is used to set critical values ​​based on the rate of change of real-time monitoring signals, sensor signal stability, and engineering experience.

[0088] If the displacement sensor detects a rate of change exceeding 0.05 mm / s, the sampling frequency is automatically increased from 100 Hz to 200 Hz; Otherwise, maintain at 100Hz to reduce redundant data.

[0089] Sampling formula: The final sampled signal: ; Among them: wi(t) is dynamically adjusted according to the signal change rate and local importance (determined by the finite element model); xi(t) is the sampling data of the i-th sensor at time t.

[0090] 3. Data preprocessing and fusion calculation process Preprocessing: All sensor raw data are first decomposed by wavelet transform, and then an adaptive filter (parameter β is 0.3) is used to filter out high-frequency noise, and missing data are interpolated and supplemented.

[0091] Data Fusion: The fusion module uses the weighted projection pursuit method, combined with the theoretical stress distribution obtained by finite element calculation, to reduce the dimension of each sensor data and fuse them into a comprehensive health index S.

[0092] Calculation formula: S=0.4·proj displacement+0.3·proj settlement+0.2·proj strain+0.1·proj crack in: proj displacement = the value obtained by processing the displacement data through the projection function (i.e. the ratio of the actual measured value to the theoretical value); The same goes for proj settlement, proj strain, and proj crack, which are assigned different weights after normalization.

[0093] The data measured at a certain moment is: Displacement: 2.0mm (theoretical allowable value is 2.5mm); Settlement: 1.0mm (theoretical allowable value is 1.2mm); Strain: 300με (theoretical allowable value is 350με); Crack width: 0.2mm (theoretical allowable value is 0.3mm).

[0094] After normalization, we get: proj displacement = 2.0 / 2.5 = 0.8; proj settlement = 1.0 / 1.2 ≈ 0.83; proj strain = 300 / 350 ≈ 0.86; proj crack = 0.2 / 0.3≈0.67.

[0095] Then the comprehensive health index S=0.4×0.8+0.3×0.83+0.2×0.86+0.1×0.67=0.32+0.25+0.17+0.067≈0.807.

[0096] If the warning threshold is set to 0.75, the current state is close to the warning range and the system will automatically issue maintenance recommendations.

[0097] 4. Anomaly detection, health assessment, and risk prediction Anomaly Detection: Variational autoencoder (VAE) is used to train normal state data. If the reconstruction error of real-time data exceeds 0.3, it is judged as abnormal. At the same time, the abnormal probability is calculated by density estimation method. When p(r|x)<0.2, it is confirmed that there is a health risk.

[0098] Health Assessment: The outputs of multiple models (including displacement, settlement, strain and crack) are integrated and converted into health warning indices using fuzzy logic.

[0099] Early warning index y = μ (comprehensive model output); If y is lower than 0.5, an alarm is triggered.

[0100] Risk Profile: The long short-term memory network (LSTM) is used to capture temporal dependencies and combined with the Bayesian network to predict risk probability.

[0101] Over a period of five consecutive minutes, the comprehensive health indicator S gradually decreased (0.82→0.78→0.75→0.70→0.65). The LSTM model predicted that S might drop to 0.60 within the next 10 minutes. The Bayesian network calculated that the risk probability reached 80%, so the system recommended an emergency check.

[0102] All the above sensor data are uploaded to the central monitoring system in real time through wired / wireless transmission modules, and are initially processed in combination with local edge computing units to ensure data accuracy and latency of less than 1 second.

[0103] The data acquisition system has been calibrated in the laboratory and verified on site to ensure that the accuracy of each sensor meets the engineering requirements.

[0104] Through multi-source data fusion, professional engineering calculations (finite element, modal analysis), and advanced machine learning technology, the system can: accurately reflect the comprehensive health status of bridges under tension, compression, and bending conditions; When the load increases sharply or there is external disturbance, the sampling frequency can be increased in time to capture the changes in key parameters; Realize real-time calculation of comprehensive health index S and risk probability, providing quantitative basis for maintenance decision-making; In actual applications, local abnormalities are detected in advance through the early warning system, and the maintenance department intervenes at the early stage of the problem, effectively extending the service life of the bridge and reducing safety hazards.

[0105] As can be seen from the above, this embodiment proves the feasibility and superiority of the system in actual engineering through detailed parameter settings, data collection and processing processes. It not only meets the multidimensional needs of bridge health monitoring, but also realizes efficient and accurate health assessment and early warning through professional engineering methods and machine learning technology.

[0106] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion, characterized in that: include: The data acquisition module is used to obtain various damage parameter data such as displacement, settlement, strain, and cracks caused by tension, compression, and bending during the use of the bridge through multiple sensors; A data preprocessing module is used to perform denoising, data cleaning and missing value filling on the disease parameter data to obtain denoised data; A multi-source data fusion module is used to fuse the denoised data and perform comprehensive processing on the data through weighted projection pursuit method and fuzzy inference system to obtain fused data; An anomaly detection and pattern recognition module is used to perform anomaly detection on the fused data using a deep generative model and a density estimation method, identify abnormal patterns of the bridge structure in combination with finite element calculation results, and obtain anomaly detection results; A health assessment and early warning module, configured to assess the health of the bridge based on the abnormality detection results and modal analysis, and generate early warning information; A visual inspection and image analysis module, which collects images of bridge surfaces and uses convolutional neural networks to detect cracks in the images, providing visual information about the structural health. The risk prediction and decision support module is used to combine the warning information and the visualization information to perform bridge risk prediction and obtain prediction results.

2. The health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion according to claim 1 is characterized in that: The data acquisition module uses an adaptive sensor sampling method to automatically adjust the sensor data acquisition frequency according to the dynamic characteristics of the bridge, thereby optimizing the timeliness and accuracy of the data; The adaptive weighted sampling method dynamically adjusts the sampling frequency of each sensor according to the actual needs of bridge health monitoring, and performs adaptive optimization based on the signal importance of each sensor and environmental changes. The formula of the adaptive weighted sampling method is: ； Where f(t): the final integrated sampling signal; wi(t): The adaptive weight of the i-th sensor, which is dynamically adjusted according to the signal’s rate of change, importance, and stability; xi(t): sampling data of the i-th sensor at time t; n: total number of sensors; t: current time.

3. The health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion according to claim 2 is characterized in that: The data preprocessing module uses a hybrid denoising method of wavelet transform and adaptive filtering to extract the high-frequency components of the signal and remove the noise to obtain denoised data; The expression of the hybrid denoising method of wavelet transform and adaptive filtering is: ； in : denoised signal; ak: coefficient of wavelet transform, representing the weight of the signal in different frequency bands; ψk(t): wavelet basis function, which decomposes the signal at different scales; β: weight of the adaptive filter, used to adjust the strength of noise removal; filter(x(t)): The result of processing the original signal x(t) by the adaptive filter.

4. The health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion according to claim 1 is characterized in that: The weighted projection pursuit method and fuzzy reasoning reduce the dimensionality of high-dimensional data through the projection pursuit algorithm, and the fuzzy reasoning performs a comprehensive evaluation on the reduced dimensionality data. The expression of the weighted projection pursuit method and fuzzy reasoning is: ； Where S: comprehensive assessment of health status indicators; wi: the weight of the i-th sensor, which is dynamically adjusted based on the importance of the sensor; proji(xi): projection function for reducing the dimension of the input data xi of the i-th sensor; n: number of sensors.

5. The health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion according to claim 1 is characterized in that: The anomaly detection and pattern recognition module is based on the variational autoencoder model and density estimation method. It models normal data through a generative model and matches it with real-time data to identify potential abnormal patterns. The expression of the variational autoencoder model and density estimation method is: ； LVAE: loss function of variational autoencoder; X: reconstructed data; z: latent variable; E: represents the expectation of the latent variable z; Eq[log p(x|z)] represents the expected value of the log-likelihood of reconstructing the data x under the distribution of the latent variable z; q(z|x): latent variable distribution output by the encoder; p(x|z): the output distribution generated by the decoder; p(z): prior distribution, representing the prior distribution of the latent variable z; DKL: Kullback-Leibler divergence, which measures the difference between the distribution of the latent variable and the prior distribution.

6. The health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion according to claim 1 is characterized in that: The health assessment and early warning module uses a method that combines ensemble learning with fuzzy logic reasoning to integrate multiple health assessment models to generate a health early warning index for the bridge and issue early warning information based on the assessment results. The expression of the method combining ensemble learning with fuzzy logic reasoning is: ； ； Where y: comprehensive health assessment output; fe: ensemble learning model, combining the prediction results of multiple single models; x1,x2,...,xn: input data for each evaluation model; : Early warning index generated by fuzzy reasoning; μ: Fuzzy inference function used to convert health assessment results into warning levels.

7. The health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion according to claim 1 is characterized in that: The risk prediction and decision support module uses a long short-term memory network combined with a Bayesian network to model time series data, predict future health trends and perform risk reasoning to generate decision support information; The expression of the long short-term memory network combined with the Bayesian network is: ； ； Where ht: hidden state in LSTM, representing the memory information of time step t; h t-1 Represents the hidden state at the previous moment; σ: activation function, sigmoid function, used to map the input to the [0,1] interval; Wh, Wx: LSTM weight matrix, acting on hidden state and input data respectively; xt: input data of the current time step; n: number of sensors; p(r|x): risk probability given observation data x in the Bayesian network; p(x|r): likelihood of observation data under risk state; p(r): prior probability of risk; p(x): marginal probability of observing data x.

8. The health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion according to claim 1 is characterized in that: The system also includes a data visualization and reporting module for visually displaying the prediction results in the form of charts or reports; The system sets up differentiated sensor layout schemes and data processing algorithms based on the special geometric configuration and stress characteristics of special-shaped steel arch bridges, which is different from the monitoring methods of conventional bridges.

9. The health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion according to claim 1 is characterized in that: The data visualization and reporting module adopts a dynamic interactive data visualization algorithm combined with an incremental learning model to update charts and reports in real time, ensuring the real-time and accuracy of data display.

10. The health monitoring and early warning system for special-shaped steel arch bridges based on multi-source data fusion according to claim 8, characterized in that: The data processing algorithm includes a multimodal data fusion method, a cave volume calculation method and a permeability coefficient prediction method.

Images