Bridge safety monitoring system based on sensor data

Abstract

The present invention relates to the field of safety monitoring technology, and more specifically, to a bridge safety monitoring system based on sensor data. It includes a sensor measurement unit that monitors the key physical parameters of the bridge in real time; a data acquisition and transmission unit that reads the sensor data and transmits the read sensor data; a data processing and analysis unit that extracts key features reflecting the health status of the bridge from the sensor data based on a multimodal physical perception feature fusion method, identifies abnormal patterns in the sensor data through a support vector machine model, and evaluates the remaining life of the bridge based on a carbonization-rust-crack closed-loop feedback control rule; an information management unit that stores the analysis results of the data processing and analysis unit and displays the monitored physical parameters of the bridge. The design of the present invention achieves high-precision feature extraction of the response of the bridge structure by introducing a multimodal physical perception feature fusion method, combined with adversarial learning, graph wavelet transform, and three-dimensional convolutional neural network.

Description

Technical Field

[0001] The present invention relates to the technical field of safety monitoring, and in particular to a bridge safety monitoring system based on sensor data. Background Art

[0002] With increasing traffic volume and increasing service life, infrastructure such as bridges is increasingly facing the risk of aging and damage. However, traditional manual inspection methods make it difficult to achieve real-time, comprehensive monitoring of bridge health. Accurately separating structural response signals from environmental noise is particularly challenging in complex natural environments, particularly those affected by factors such as severe weather and vibration. Traditional monitoring methods often struggle to effectively distinguish between the two, compromising the accuracy of monitoring data. At the same time, with the continuous advancement of sensor technology, the amount of data acquired is exploding. Efficiently processing this massive amount of data and extracting useful information to assess bridge health has become an urgent issue. Furthermore, traditional life prediction models are mostly based on static or linear assumptions and cannot fully reflect the true deterioration process of bridge structures over time, particularly failing to account for nonlinear degradation effects caused by factors such as carbonization and corrosion. Therefore, a bridge safety monitoring system based on sensor data is being designed. Summary of the Invention

[0003] The purpose of the present invention is to provide a bridge safety monitoring system based on sensor data to solve the problems raised in the above-mentioned background technology, namely, the inability of traditional manual inspection to achieve real-time and comprehensive monitoring, the interference of environmental noise affecting data accuracy, the difficulty in efficiently processing and feature extraction of massive sensor data, and the inability of traditional life prediction models to accurately reflect the nonlinear degradation process of the structure.

[0004] To achieve the above objectives, the present invention provides a bridge safety monitoring system based on sensor data, comprising:

[0005] A sensor measurement unit, which monitors key physical parameters of the bridge in real time;

[0006] A data acquisition and transmission unit, wherein the data acquisition and transmission unit reads sensor data and transmits the read sensor data;

[0007] a data processing and analysis unit, which extracts key features reflecting the health status of the bridge from the sensor data based on a multimodal physical perception feature fusion method, identifies abnormal patterns in the sensor data using a support vector machine (SVM) model, and assesses the remaining life of the bridge based on a carbonization-corrosion-crack closed-loop feedback control rule;

[0008] An information management unit stores the analysis results of the data processing and analysis unit and displays the monitored physical parameters of the bridge.

[0009] As a further improvement of this technical solution, the data acquisition and transmission unit includes a data reading module and a data transmission module;

[0010] The data reading module reads sensor data by connecting to the sensor and performs preliminary sorting and packaging on the read sensor data;

[0011] The data transmission module transmits the packaged sensor data to the server where the data processing and analysis unit is located through wireless transmission.

[0012] As a further improvement of the present technical solution, the data processing and analysis unit includes a data feature extraction module and a data anomaly analysis module;

[0013] The data feature extraction module receives the pre-processed sensor data and extracts key features reflecting the health status of the bridge based on a multimodal physical perception feature fusion method;

[0014] The data anomaly analysis module identifies abnormal data patterns based on the extracted key features through a support vector machine model, and evaluates the remaining life of the bridge using a carbonization-corrosion-crack closed-loop feedback control rule based on the extracted key features and abnormal data patterns.

[0015] As a further improvement of this technical solution, the multimodal physical perception feature fusion method is used to extract key features reflecting the health status of the bridge, including the following steps:

[0016] S1.1. Receive preprocessed multidimensional time series sensor data transmitted from a data acquisition and transmission unit, use the preprocessed multidimensional time series sensor data as input to a bridge structural dynamics model, generate corresponding structural response predictions under given loads and boundary conditions, use the predicted structural response values ​​as a physical benchmark, calculate the physical residual between the actual structural response measurements and the predicted structural response values, apply an adversarial learning method to separate environmental noise from structural response, and generate a structural response feature vector and environmental characteristics;

[0017] S1.2. Perform a fast Fourier transform on the physical residual signal to obtain the frequency domain characteristics of the physical residual signal. Align the frequency components of the physical residual signal with the structural main frequency characteristics of the bridge through weighted adjustment. Optimize the frequency band division by constructing a sensor graph structure. Extract the main frequency amplitude and frequency band energy anomaly index of the aligned frequency.

[0018] S1.3. Based on the sensor deployment topology, construct a sensor physical relationship graph. Perform spatial decomposition of the structural response features separated in step S1.1 and the frequency domain features in step S1.2 using graph wavelet transform, and calculate the subgraph energy distribution entropy and cross-sensor frequency response consistency.

[0019] S1.4. Integrate the time domain, frequency domain, and spatial features extracted in steps S1.2 and S1.3 to construct a three-dimensional feature tensor, and use a three-dimensional convolution operation to fuse the time domain, frequency domain, and spatial features, and output the fused high-dimensional feature tensor;

[0020] S1.5. Based on the fused high-dimensional feature tensor, use the mutual information method to filter the initial feature vectors and perform normalization on each filtered feature vector.

[0021] S1.6. Combine the normalized feature vectors into the final feature vector.

[0022] As a further improvement of the present technical solution, in S1.1, the physical residual between the actual structural response measurement value and the structural response prediction value is used to separate the environmental noise from the structural response by an adversarial learning method to generate a structural response feature vector and environmental features, including the following steps:

[0023] S1.11. Build a physically driven structural response model based on bridge structural parameters, input the current load and boundary conditions, and output the theoretically predicted response value at the same sensor location;

[0024] S1.12. Collect the actual structural response measurement value of the sensor and calculate the physical residual signal;

[0025] S1.13. Design a generator and a discriminator to build an adversarial network.

[0026] S1.14. Randomly initialize the parameters of the generator and discriminator, and train them by alternately optimizing the generator and discriminator.

[0027] S1.15. Use the cross-entropy loss function to measure the performance of the generator and discriminator.

[0028] S1.16. After training, the generator is used to extract the structural response feature vector from the physical residual signal, and the generated structural response feature vector and environmental feature vector are standardized.

[0029] As a further improvement of the present technical solution, in S1.2, constructing a sensor graph structure to optimize frequency band division and extracting the main frequency amplitude and frequency band energy anomaly index of the aligned frequency includes the following steps:

[0030] S1.21. Collect bridge sensor deployment information and the physical connection relationship between each sensor;

[0031] S1.22, build graph structure;

[0032] Among them, each node in the graph structure corresponds to a sensor, and the weight of the edge represents the correlation coefficient of the structural response between the two sensors;

[0033] S1.23. Use graph structure to analyze the response consistency of each sensor at different frequencies;

[0034] S1.24, cluster analysis is performed on the graph structure to identify subsets of sensors with similar frequency response patterns;

[0035] S1.25. Based on the cluster analysis results, retain the frequency bands within ±a% of the structural main frequency characteristics;

[0036] S1.26. Find the maximum amplitude and its corresponding frequency in the current frequency band, record the amplitude as the main frequency amplitude feature, calculate the energy of the current frequency band, and output the frequency band energy anomaly index.

[0037] As a further improvement of the present technical solution, in S1.4, a three-dimensional feature tensor is constructed, and a three-dimensional convolution operation is used to fuse the time domain, frequency domain, and spatial features to output the fused high-dimensional feature tensor, including the following steps:

[0038] S1.41. Determine tensor dimensions, where the tensor dimensions include the number of time windows, the number of physically aligned frequency bands, and the number of sensor nodes.

[0039] S1.42, for each time window, each frequency band, and each sensor node, fill the corresponding eigenvalue into the three-dimensional tensor;

[0040] S1.43. Define a 3D convolutional layer using a deep learning framework and set the convolution kernel size.

[0041] S1.44. Pass the constructed 3D feature tensor as input to the 3D convolutional layer;

[0042] S1.45. After the three-dimensional convolution layer, a high-dimensional feature tensor is output that combines the time domain, frequency domain and spatial characteristics.

[0043] As a further improvement of the present technical solution, the method of identifying abnormal data patterns through a support vector machine model and evaluating the remaining life of a bridge using a carbonization-corrosion-crack closed-loop feedback control rule based on the extracted key features and abnormal data patterns includes the following steps:

[0044] S2.1. Data acquired during the trouble-free operation of the bridge were selected as the training set. The crack propagation rate was predicted using a spatiotemporal graph convolutional network, and the carbonization depth was inverted using a hyperspectral image fusion CNN model.

[0045] S2.2. Using the extracted key features, crack growth rate, and carbonization depth as input, a support vector machine model is trained. During the training process, the parameters of the support vector machine model are adjusted through cross-validation.

[0046] S2.3. Use the trained support vector machine model to score the training data. If the score exceeds the preset threshold b, the bridge data for this time period is marked as abnormal; otherwise, it is considered normal.

[0047] S2.4. Based on long-term bridge monitoring data, construct a bridge performance degradation sequence and use the ratio of the carbonization zone ion diffusion rate to the rust expansion stress intensity to determine the bridge performance degradation sequence. As an indicator, it drives the carbonization depth evolution model, and then triggers the correction item in the degradation sequence through the dynamic results output by the carbonization depth evolution model;

[0048] S2.5. Model the bridge performance degradation sequence as a non-negative, non-decreasing random process and The values ​​construct three types of degradation patterns;

[0049] S2.6. Set the degradation threshold d for bridge structure failure and The value dynamically adjusts the degradation threshold d;

[0050] S2.7. Based on the current accumulated degradation level, shape parameters, and scale parameters, calculate the time from now until the degradation threshold d is expected to be reached. This time is the remaining life of the bridge.

[0051] As a further improvement of this technical solution, in S2.5, according to The three types of degradation models are constructed, including the following steps:

[0052] S2.51, if Value less than Base point of value , using Gaussian distribution to generate multiple degradation increment samples, simulate the degradation situation in the future period, and add the generated degradation increments to the current degradation level to update the bridge performance degradation sequence;

[0053] S2.52, if Value greater than or equal to The base point of the value and is less than or equal to Upper limit of value , use Gamma distribution to generate multiple degradation increment samples to simulate the degradation situation in the future;

[0054] S2.53, if Value greater than Upper limit of value , multiple degradation increment samples are generated using Weibull distribution to simulate the degradation situation in the future. The generated degradation increments are added to the current degradation level to update the bridge performance degradation sequence.

[0055] As a further improvement of this technical solution, the information management unit includes a storage management module, an alarm management module, a report generation module and a video monitoring module;

[0056] The storage management module stores bridge physical parameters and abnormality analysis results by establishing a database;

[0057] The alarm management module receives abnormal data from the data processing and analysis unit and pushes alarm information according to the preset level;

[0058] The report generation module generates reports, historical data reports, and diagnostic reports for managing bridge data;

[0059] The video monitoring module is connected to and controls the bridge on-site camera for remote video monitoring.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. This sensor data-based bridge safety monitoring system introduces a multimodal physical perception feature fusion method, combined with adversarial learning, graph wavelet transform, and three-dimensional convolutional neural network, to achieve high-precision feature extraction of bridge structure response and effective separation of environmental noise, thereby significantly improving the accuracy and robustness of bridge health status assessment. It can not only effectively identify early structural damage, but also has good anti-interference ability, making it suitable for long-term bridge monitoring needs in complex environments.

[0062] 2. This sensor-based bridge safety monitoring system utilizes a closed-loop feedback control rule for carbonization, corrosion, and cracks. Combining real-time sensor data with a support vector machine (SVM) anomaly recognition model, it constructs a dynamically evolving bridge performance degradation model and, based on this, enables intelligent prediction of bridge remaining life. By introducing the carbonization and corrosion coupling coefficient, k, to dynamically adjust the degradation process and failure threshold, this system overcomes the limitations of traditional linear degradation models, enabling bridge life assessments to more closely align with actual degradation mechanisms. This provides a scientific basis for bridge maintenance decisions and improves bridge operation and maintenance safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is the overall flow chart of the present invention;

[0064] The meaning of each number in the figure is:

[0065] 1. Sensor measurement unit; 2. Data acquisition and transmission unit; 21. Data reading module; 22. Data transmission module; 3. Data processing and analysis unit; 31. Data feature extraction module; 32. Data anomaly analysis module; 4. Information management unit; 41. Storage management module; 42. Alarm management module; 43. Report generation module; 44. Video monitoring module. DETAILED DESCRIPTION

[0066] The following will clearly and completely describe 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. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0067] Example: See Figure 1 As shown, a bridge safety monitoring system based on sensor data is provided, including:

[0068] The sensor measurement unit 1 monitors the key physical parameters of the bridge (including stress, strain, displacement, etc.) in real time. In this embodiment, these sensors are usually installed in key locations of the bridge to ensure comprehensive and accurate acquisition of bridge health data. The sensors include stress sensors, strain sensors, displacement sensors, and hyperspectral imaging sensors deployed in key areas of the bridge deck. The hyperspectral imaging sensors are installed on the bottom surface of the bridge main beam and the surface of the pier, fixed by brackets and equipped with anti-shake modules. Each sensor covers an area of ​​10m×10m. The light source adopts natural light compensation mode during data collection, and the output data is a three-dimensional hyperspectral cube.

[0069] The data acquisition and transmission unit 2 reads the sensor data and transmits the read sensor data;

[0070] In this embodiment, the data acquisition and transmission unit 2 includes a data reading module 21 and a data transmission module 22;

[0071] The data acquisition and transmission unit 2 is responsible for safely and efficiently transmitting the data collected by the sensors to the data processing and analysis unit 3. This unit includes a data collector, transmission equipment, and the corresponding communication network.

[0072] The data reading module 21 reads sensor data by connecting to the sensor, and performs preliminary sorting and packaging on the read sensor data;

[0073] The data transmission module 22 transmits the packaged sensor data to the server or data center where the data processing and analysis unit 3 is located through wired or wireless transmission.

[0074] Data processing and analysis unit 3 extracts key features reflecting the health status of the bridge from the sensor data based on a multimodal physical perception feature fusion method, identifies abnormal patterns in the sensor data using a support vector machine model, and estimates the remaining life of the bridge based on a carbonization-corrosion-crack closed-loop feedback control rule;

[0075] In this embodiment, the data processing and analysis unit 3 includes a data feature extraction module 31 and a data anomaly analysis module 32;

[0076] The data feature extraction module 31 receives pre-processed sensor data (such as stress, strain, displacement, etc.) and extracts key features reflecting the health status of the bridge based on a multimodal physical perception feature fusion method;

[0077] The data anomaly analysis module 32 identifies abnormal data patterns based on the extracted key features through a support vector machine model, and evaluates the remaining life of the bridge using a carbonization-corrosion-crack closed-loop feedback control rule based on the extracted key features and the abnormal data patterns.

[0078] Furthermore, a multimodal physical perception feature fusion method accurately extracts key features reflecting the health of bridge structures from complex sensor data. Adversarial learning is used to separate environmental noise from structural response (S1.1), combined with a physical model benchmark (S1.11), eliminating interference from temperature and humidity to ensure feature reliability. Time-domain residuals, frequency-domain dominant frequency amplitudes (S1.2), and spatial sensor topology (S1.3) are integrated, and three-dimensional convolution is used to generate a high-dimensional feature tensor (S1.4) to comprehensively capture bridge dynamic behavior. The frequency band energy anomaly index (S1.26) and subgraph energy distribution entropy (S1.3) can identify micro-damage (such as early cracks) and support a support vector machine model to accurately locate abnormal patterns (S2.3). Through the fusion of physical model guidance and deep learning, a triple breakthrough has been achieved in noise robustness, early damage sensitivity, and life prediction accuracy, providing a key technical support for bridge monitoring to move from "post-event alarm" to "pre-event prediction."

[0079] The key features reflecting the health status of the bridge are extracted based on the multimodal physical perception feature fusion method, which includes the following steps:

[0080] S1.1. Receive pre-processed multi-dimensional time series sensor data (including stress, strain, displacement, etc.) transmitted from the data acquisition and transmission unit 2, and use the pre-processed multi-dimensional time series sensor data as the input of the bridge structure dynamics model (the bridge structure dynamics model is constructed based on structural mechanics and dynamics control equations, and is used to simulate the dynamic response behavior of the bridge under various loads and boundary conditions. The model first establishes the geometric topology of the bridge, defines the material properties of each component (including elastic modulus, density, damping, etc.), and inputs time-varying external excitations such as vehicle loads, wind loads, and temperature changes. At the same time, boundary conditions such as support types and connection forms are set. On this basis, the model solves the problem. The system uses the structural dynamics equation to obtain the time-varying responses at key nodes, including displacement, acceleration, stress, or strain values, as the structural response prediction results. The output is aligned with the actual sensor data and used to construct physical residuals to support subsequent feature extraction, anomaly identification, and remaining life prediction. Under given loads and boundary conditions, the corresponding structural response prediction value is generated. The structural response prediction value is used as a physical benchmark to calculate the physical residual between the actual structural response measurement value and the structural response prediction value. The adversarial learning method is applied to separate environmental noise from structural response, and the structural response feature vector and environmental characteristics are generated (to improve the ability to identify the true state of the structure and enhance the accuracy and robustness of feature extraction).

[0081] The physical residual between the actual structural response measurement value and the structural response prediction value is calculated, and the adversarial learning method is applied to separate the environmental noise and the structural response to generate the structural response feature vector and the environmental characteristics. The following steps are included:

[0082] S1.11. Construct a physics-driven structural response model based on bridge structural parameters (including geometric parameters, material parameters, etc.). (A physics-driven structural response model is an analytical or numerical dynamic model constructed based on bridge structural parameters, boundary conditions, and load inputs. Its core function is to predict the theoretical response value (stress, strain, displacement) at each sensor location. This model inputs the current external load information and boundary conditions, combines the geometric information, material properties, and structural topology of the bridge itself, and outputs the response variable corresponding to the sensor measurement point as the physical reference value of the structural response.) Input the current load and boundary conditions and output the theoretical predicted response value (stress, strain, displacement) at the same sensor location.

[0083] S1.12. Collect actual structural response measurements from sensors and calculate physical residual signals (the difference between actual structural response measurements and predicted values. Physical residuals are used to reflect structural anomalies. By comparing actual structural response measurements from sensors with predicted values ​​based on the bridge dynamics model, the resulting residual signals can effectively reflect structural deviations and potential anomalies, helping to locate the source of anomalies).

[0084] S1.13. Design a generator and a discriminator to construct an adversarial network. The generator attempts to extract structural response features from the input (physical residual signal) while mimicking the distribution of environmental noise as much as possible. The discriminator attempts to distinguish the generated structural response features from the real structural response features.

[0085] S1.14. Randomly initialize the parameters of the generator and discriminator. Train by alternately optimizing the generator and discriminator. In each iteration, first fix the discriminator parameters and update the generator to minimize the discriminator's ability to identify its output (i.e., make the generated structural response features more difficult to identify as forged). Then, fix the generator parameters and update the discriminator to better distinguish between real and generated structural response features.

[0086] S1.15. Use the cross-entropy loss function to measure the performance of the generator and discriminator. (The metric is to minimize the discriminator's accuracy in distinguishing between real and generated samples (the discriminator wants to minimize the loss, i.e., correct classification), while maximizing the generator's ability to deceive generated samples (the generator wants to maximize the discriminator's loss, i.e., incorrect classification). The two compete with each other and are optimized together until a balance is reached.)

[0087] S1.16. After training, the generator is used to extract the structural response feature vector from the physical residual signal, and the generated structural response feature vector and environmental feature vector are standardized.

[0088] S1.2. Perform a fast Fourier transform on the physical residual signal to obtain its frequency domain characteristics. Then, through weighted adjustment, align the frequency components of the physical residual signal with the structural dominant frequency characteristics of the bridge (a weighting function is defined for each sensor's amplitude spectrum, assigning higher weights to locations close to the structural dominant frequency characteristics. A weighted strategy is used to adjust the spectrum to eliminate frequency offsets caused by boundary condition changes or external disturbances). A sensor graph structure is constructed (node ​​= sensor, edge = physical association strength) to optimize frequency band division, focusing on structurally sensitive frequency bands. The dominant frequency amplitude and frequency band energy anomaly index of the aligned frequencies are extracted (building a graph structure based on the physical connection relationships between sensors and using it for cluster analysis can help identify subsets of sensors with similar frequency response patterns. This not only helps focus on key features that truly reflect the structural state, but also effectively filters out noise information unrelated to structural health, improving the quality of feature extraction in subsequent analysis steps).

[0089] The following steps are involved: constructing a sensor graph structure (node ​​= sensor, edge = physical association strength), optimizing frequency band division, focusing on structurally sensitive frequency bands, and extracting the dominant frequency amplitude and frequency band energy anomaly index of the aligned frequencies.

[0090] S1.21. Collect bridge sensor deployment information (location, type) and the physical connection relationship between each sensor;

[0091] S1.22, build graph structure;

[0092] Each node in the graph corresponds to a sensor, and the edge weight represents the correlation coefficient of the structural response between the two sensors (the correlation coefficient between the two sensors is calculated based on the modal vibration shape vector);

[0093] S1.23. Use graph structure to analyze the response consistency of each sensor at different frequencies;

[0094] S1.24. Perform cluster analysis (spectral clustering) on ​​the graph structure to identify subsets of sensors with similar frequency response patterns.

[0095] S1.25. Based on the cluster analysis results, retain frequency bands within ±a% of the structural dominant frequency characteristics (ensuring that the frequency bands of interest are closely surrounding the frequency components that best reflect the structural dynamic characteristics, further enhancing the pertinence and effectiveness of the structural health assessment. In this process, the dominant frequency amplitude and frequency band energy anomaly index serve as important characteristic indicators that can intuitively reflect the response characteristics and changing trends of the structure at specific frequencies).

[0096] S1.26. Find the maximum amplitude and its corresponding frequency within the current frequency band, record the amplitude as the main frequency amplitude feature, calculate the energy of the current frequency band, and output the frequency band energy anomaly index;

[0097] S1.3. Based on the sensor deployment topology, construct a sensor physical relationship graph (node ​​attributes include sensor type (including strain gauges) and installation location, and edge weights are determined by the strength of the structural physical connection between sensors). Perform spatial decomposition of the structural response features separated in step S1.1 and the frequency domain features in step S1.2 using graph wavelet transform. Extract the energy distribution of subgraphs at different scales and calculate the subgraph energy distribution entropy (calculating the subgraph energy distribution entropy to quantify the spatial heterogeneity of the structural response, that is, to measure the uniformity of the energy distribution within each subgraph) and the cross-sensor frequency response consistency (analyzing and comparing the similarities or differences in the responses of each sensor within the selected frequency band to reflect changes in the response coordination between sensors).

[0098] S1.4. Integrate the time-domain, frequency-domain, and spatial features extracted in steps S1.2 and S1.3 to construct a three-dimensional feature tensor (including the number of time windows, the number of physically aligned frequency bands, and the number of sensor nodes). Use a three-dimensional convolution operation to fuse the time-domain, frequency-domain, and spatial features, and output a fused high-dimensional feature tensor. (By integrating the time series segments of sensor data (time domain), the physically aligned frequency components (frequency domain), and the spatial distribution of sensors (spatial domain) into a three-dimensional tensor, this method can comprehensively capture the multi-level characteristics of the health status of the bridge structure. This multi-dimensional information integration helps to more accurately reflect the actual condition of the bridge. The three-dimensional convolution operation can automatically learn and extract the complex relationships between different dimensions, excavating the deep features hidden in the original data. Compared with traditional single-dimensional analysis methods, this method can provide richer and more representative feature representations, which helps to improve the accuracy of subsequent tasks (such as anomaly detection and remaining life prediction).

[0099] Among them, constructing a three-dimensional feature tensor and using a three-dimensional convolution operation to fuse time domain, frequency domain, and spatial features to output the fused high-dimensional feature tensor includes the following steps:

[0100] S1.41. Determine the tensor dimensions. The tensor dimensions include the number of time windows (the number of time series segments determined according to actual monitoring needs), the number of physically aligned frequency bands (the number of key frequency bands determined from S1.2), and the number of sensor nodes (the total number of sensors deployed on the bridge).

[0101] S1.42, for each time window, each frequency band, and each sensor node, fill the corresponding eigenvalue into the three-dimensional tensor;

[0102] S1.43. Define a 3D convolutional layer using a deep learning framework (PyTorch) and set the convolution kernel size.

[0103] S1.44. Pass the constructed 3D feature tensor as input to the 3D convolution layer. The convolution operation automatically slides the window over the entire tensor, performs a dot product and sums the results at each position, and thus generates a new feature map.

[0104] S1.45, after a series of three-dimensional convolutional layers, outputs a high-dimensional feature tensor that combines time domain, frequency domain, and spatial characteristics;

[0105] S1.5. Based on the fused high-dimensional feature tensor, use the mutual information method to filter the initial feature vectors and perform normalization on each filtered feature vector.

[0106] S1.6. Combine the normalized feature vectors into the final feature vector.

[0107] Furthermore, a support vector machine model is used to identify abnormal data patterns. Based on the extracted key features and abnormal data patterns, a closed-loop feedback control rule for carbonization, corrosion, and cracks is used to estimate the remaining life of the bridge. (The closed-loop feedback control rule for carbonization, corrosion, and cracks integrates the three fields of chemistry (carbonization), electrochemistry (corrosion), and mechanics (cracks) to capture the full chain of degradation evolution and improve the dynamic accuracy of the prediction. The feedback term is introduced, where rust expansion stress induces cracks, which in turn accelerate carbonization, avoiding the traditional "static" assumption and improving the dynamic accuracy of the prediction. The degradation rate is regulated by the k value to support the simulation of degradation jumps and self-amplification effects.) The following steps are included:

[0108] S2.1. Data acquired during the trouble-free operation of the bridge are selected as the training set. The crack propagation rate is predicted using a spatiotemporal graph convolutional network (the input of the spatiotemporal graph convolutional network is: a spatiotemporal graph structure (nodes represent sensor locations or specific structural key points, and edges represent the physical association strength (distance) between these points) and time series data (for each node, there is a series of data records that change over time, which reflect the historical changes in the structural parameters at that location); intermediate processing includes graph convolution operations, temporal convolution operations, and fusion mechanisms; the final output is an estimate of the crack propagation rate in the future, which is a numerical value or a set of numerical values ​​representing the prediction results in different time periods). The carbonization depth is inverted using a hyperspectral image fusion CNN model (hyperspectral image data is combined with a convolutional neural network (CNN) to identify and quantify the degree of carbonization in bridge concrete).

[0109] In the input layer of the hyperspectral image fusion CNN model, the hyperspectral data and the strain sensor time series data at the corresponding position (i.e., sensor measurement unit 1 monitors the strain parameters of the bridge in real time) are used as collaborative inputs and mapped into an H×W×32 feature map through a fully connected layer (H represents the height of the image feature map, and W represents the width of the image feature map). In the fusion layer, the spectrum and strain features are spliced ​​in the channel dimension to obtain an H×W×182 tensor. The spatiotemporal-spectral features are extracted through three layers of 3D convolution (kernel size 3×3×5, step size 1×1×2). The key bands are dynamically weighted through the channel attention module. Finally, the output layer predicts the carbonization depth through the regression head and the crack propagation rate using the spatiotemporal graph convolutional network. The data is simultaneously collected by the bridge deck hyperspectral imaging sensor and input into the hyperspectral CNN model to invert the carbonization depth.

[0110] S2.2. Using the extracted key features, crack growth rate, and carbonization depth as input, a support vector machine model is trained. During the training process, the parameters of the support vector machine model are adjusted through cross-validation.

[0111] S2.3. Use the trained support vector machine model to score the training data. If the score exceeds the preset threshold b, the bridge data for this time period is marked as abnormal; otherwise, it is considered normal.

[0112] S2.4. Based on long-term bridge monitoring data, construct a bridge performance degradation sequence to reflect the trend of structural performance over time, and use the ratio of the carbonization zone ion diffusion rate to the rust expansion stress intensity ( value) as an indicator to drive the carbonization depth evolution model , and then trigger the correction term in the degradation sequence through the dynamic results output by the carbonization depth evolution model (the correction term is ,when When the value changes, the evolution mode and degradation threshold of the entire degradation model will be dynamically adjusted to more accurately simulate the actual degradation process). By using the trigger correction term, the degradation process is no longer a simple linear accumulation, but a nonlinear evolution with a feedback amplification effect, reflecting the closed-loop feedback of carbonization depth → steel corrosion → microcrack expansion → carbonization acceleration.

[0113] Among them, the bridge performance degradation sequence is:

[0114] Where, Indicates that the bridge is at time The cumulative degradation level, i.e. the bridge performance degradation sequence, represents the base degradation increment (determined by environmental conditions), , represents the amplification function of the carbonization-corrosion coupling factor, , represents the ion diffusion rate in the carbonization zone, Indicates the rust expansion stress intensity, Indicates the carbonization depth at the current moment, The function that represents the influence of carbonization depth on degradation is a linear function. represents the weight coefficient, represents a small disturbance term or random noise, Indicates time;

[0115] The ratio of the ion diffusion rate in the carbonization zone to the rust expansion stress intensity ( The carbonization depth evolution model driven by the indicator (value) is as follows:

[0116] ;

[0117] Where, Indicates the carbonization depth at the current moment, represents the material damage coefficient;

[0118] S2.5. Model the bridge performance degradation sequence as a non-negative, non-decreasing random process and The values ​​construct three types of degradation patterns;

[0119] Among them, according to The three types of degradation models are constructed, including the following steps:

[0120] The three degradation modes are the real-time ratio of chemical diffusion (carbonization) to mechanical response (rust expansion stress). As a criterion for degradation state, it breaks through the traditional time / single physical field dependent model;

[0121] S2.51, if Value less than Base point of value (Carbonization and corrosion have not yet formed a significant coupling, the structure is in a relatively safe state, and the degradation rate is slow.) Use Gaussian distribution to generate multiple degradation increment samples to simulate the degradation situation over a period of time in the future. The generated degradation increments are added to the current degradation level to update the bridge performance degradation sequence.

[0122] During this stage, the incremental degradation of bridge performance ;

[0123] in, is the average value of the basic degradation rate, reflecting a slow and uniform degradation trend. is the standard deviation of fluctuation, reflecting the random disturbance of the environment;

[0124] Simulation method: Randomly generate multiple degradation increment samples from Gaussian distribution, and for each sample :

[0125] ;

[0126] Where, represents the time step index;

[0127] S2.52, if Value greater than or equal to The base point of the value and is less than or equal to Upper limit of value (The coupling effect begins to appear and some structural problems begin to appear, but have not yet reached a dangerous level). Use the Gamma distribution to generate multiple degradation increment samples to simulate the degradation situation in the future.

[0128] Among them, Gamma distribution is used for modeling:

[0129] ;

[0130] represents the shape parameter, which determines the degree of skewness of the distribution. Represents the scale parameter, controls the degradation rate, and the Gamma distribution is suitable for describing occasional acceleration in the degradation process (including microcrack connection and protective layer failure);

[0131] Simulation method: multiple Gamma distribution samples are sampled as future degradation increments, but the degradation sequence is not updated immediately. They are only used as input for deducing alternative paths or Monte Carlo predictions.

[0132] S2.53, if Value greater than Upper limit of value (The corrosion rate surges nonlinearly (exponential amplification effect), and the structure enters a stage of rapid deterioration.) Multiple degradation increment samples are generated using the Weibull distribution to simulate degradation over a period of time. The generated degradation increments are added to the current degradation level to update the bridge performance degradation sequence.

[0133] Its increment is modeled using a Weibull distribution:

[0134] ;

[0135] represents the scale parameter, which controls the degradation rate, Represents the shape parameter reflecting the “acceleration” characteristic;

[0136] For every moment , sampling multiple samples , update the degradation level:

[0137] ;

[0138] S2.6. Set the degradation threshold d for bridge structure failure. This value reflects that when the bridge structure reaches this level of degradation, it will no longer be able to operate safely. The value dynamically adjusts the degradation threshold d;

[0139] based on The value dynamically adjusts the degradation threshold as follows:

[0140] ;

[0141] Where, A reference point representing the dynamic coupling of carbonization and corrosion; express The width of the transition interval, indicating Changes within this range have a greater impact on structural degradation; Indicates the maximum ratio by which the degradation threshold can be reduced; represents the initial degradation threshold; Indicates based on The degradation threshold after the value is dynamically adjusted;

[0142] S2.7. Based on the current accumulated degradation level, shape parameters, and scale parameters, calculate the time from now until the degradation threshold d is expected to be reached. This time is the remaining life of the bridge.

[0143] The information management unit 4 stores the analysis results of the data processing and analysis unit 3 and displays the monitored physical parameters of the bridge;

[0144] In this embodiment, the information management unit 4 includes a storage management module 41, an alarm management module 42, a report generation module 43 and a video monitoring module 44;

[0145] The storage management module 41 establishes a database to store bridge physical parameters and abnormality analysis results. Users can access the system anytime and anywhere to view real-time data and historical records of the bridge. By tracing and analyzing the monitoring data, the system can promptly identify potential safety issues of the bridge.

[0146] The alarm management module 42 receives abnormal data from the data processing and analysis unit 3 and pushes alarm information according to preset levels. Through various types of sensors deployed on the bridge surface, the system can monitor the structural parameters of the bridge and set a multi-level alarm function. When the data exceeds the alarm upper limit, the system will automatically issue alarm information of different levels and convey it to relevant leaders and responsible persons in various ways;

[0147] The report generation module 43 generates reports, historical data reports, and diagnostic reports for managing bridge data;

[0148] The video monitoring module 44 is connected to and controls the bridge's on-site camera for remote video monitoring. Users can see the on-site video images in real time from the monitoring terminal. At the same time, the on-site camera lens and pan / tilt head can be controlled through software to meet the user's monitoring needs for multi-directional on-site video images and ensure the safe operation of the bridge.

[0149] The basic principles, main features, and advantages of the present invention are shown and described above. It should be understood by those skilled in the art that the present invention is not limited to the above-described embodiments. The above-described embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention, and such changes and modifications fall within the scope of the invention claimed.

Claims

1. A bridge safety monitoring system based on sensor data, characterized in that: include: A sensor measurement unit (1), wherein the sensor measurement unit (1) monitors key physical parameters of the bridge in real time; A data acquisition and transmission unit (2), wherein the data acquisition and transmission unit (2) reads sensor data and transmits the read sensor data; A data processing and analysis unit (3) is configured to extract key features reflecting the health status of the bridge from the sensor data based on a multimodal physical perception feature fusion method, identify abnormal patterns in the sensor data using a support vector machine (SVM) model, and evaluate the remaining life of the bridge based on a carbonization-corrosion-crack closed-loop feedback control rule; The support vector machine (SVM) model is used to identify abnormal patterns in sensor data, and the remaining life of the bridge is evaluated based on the carbonization-corrosion-crack closed-loop feedback control rule, which includes the following steps: S2.

1. Data acquired during the trouble-free operation of the bridge were selected as the training set. The crack propagation rate was predicted using a spatiotemporal graph convolutional network, and the carbonization depth was inverted using a hyperspectral image fusion CNN model. S2.

2. Using the extracted key features, crack growth rate, and carbonization depth as input, a support vector machine model is trained. During the training process, the parameters of the support vector machine model are adjusted through cross-validation. S2.

3. Use the trained support vector machine model to score the training data. If the score exceeds the preset threshold b, the bridge data for this time period is marked as abnormal; otherwise, it is considered normal. S2.

4. Based on long-term bridge monitoring data, a bridge performance degradation sequence is constructed. The ratio k of the ion diffusion rate in the carbonization zone to the rust expansion stress intensity is used as an indicator to drive the carbonization depth evolution model. The dynamic results output by the carbonization depth evolution model are then used to trigger the correction terms in the degradation sequence. S2.

5. Model the bridge performance degradation sequence as a non-negative, non-decreasing random process and construct three types of degradation patterns based on the k value; S2.

6. Set a degradation threshold d for bridge structure failure and dynamically adjust the degradation threshold d based on the k value; S2.

7. Based on the current accumulated degradation level, shape parameters, and scale parameters, calculate the time from now until the degradation threshold d is expected to be reached. This time is the remaining life of the bridge; An information management unit (4) stores the analysis results of the data processing and analysis unit (3) and displays the monitored physical parameters of the bridge.

2. The bridge safety monitoring system based on sensor data according to claim 1, characterized in that: The data acquisition and transmission unit (2) comprises a data reading module (21) and a data transmission module (22); The data reading module (21) reads sensor data by connecting to the sensor, and performs preliminary sorting and packaging on the read sensor data; The data transmission module (22) sends the packaged sensor data to the server where the data processing and analysis unit (3) is located via wireless transmission.

3. The bridge safety monitoring system based on sensor data according to claim 1, characterized in that: The data processing and analysis unit (3) includes a data feature extraction module (31) and a data anomaly analysis module (32); The data feature extraction module (31) receives the pre-processed sensor data and extracts key features reflecting the health status of the bridge based on a multimodal physical perception feature fusion method; The data anomaly analysis module (32) identifies abnormal data patterns based on the extracted key features through a support vector machine model, and evaluates the remaining life of the bridge using a carbonization-corrosion-crack closed-loop feedback control rule based on the extracted key features and the abnormal data patterns.

4. The sensor data-based bridge safety monitoring system according to claim 3, characterized in that: The method for extracting key features reflecting the health status of a bridge based on multimodal physical perception feature fusion includes the following steps: S1.

1. receiving pre-processed multi-dimensional time series sensor data transmitted from the data acquisition and transmission unit (2), using the pre-processed multi-dimensional time series sensor data as input to the bridge structure dynamics model, generating corresponding structural response prediction values ​​under given loads and boundary conditions, and using the structural response prediction values ​​as physical benchmarks, calculating the physical residual between the actual structural response measurement values ​​and the structural response prediction values, applying an adversarial learning method to separate environmental noise from structural response, and generating a structural response feature vector and environmental characteristics; S1.

2. Perform a fast Fourier transform on the physical residual signal to obtain the frequency domain characteristics of the physical residual signal. Align the frequency components of the physical residual signal with the structural main frequency characteristics of the bridge through weighted adjustment. Optimize the frequency band division by constructing a sensor graph structure. Extract the main frequency amplitude and frequency band energy anomaly index of the aligned frequency. S1.

3. Based on the sensor deployment topology, construct a sensor physical relationship graph. Perform spatial decomposition of the structural response features separated in step S1.1 and the frequency domain features in step S1.2 using graph wavelet transform, and calculate the subgraph energy distribution entropy and cross-sensor frequency response consistency. S1.

4. Integrate the time domain, frequency domain, and spatial features extracted in steps S1.2 and S1.3 to construct a three-dimensional feature tensor, and use a three-dimensional convolution operation to fuse the time domain, frequency domain, and spatial features, and output the fused high-dimensional feature tensor; S1.

5. Based on the fused high-dimensional feature tensor, use the mutual information method to filter the initial feature vectors and perform normalization on each filtered feature vector. S1.

6. Combine the normalized feature vectors into the final feature vector.

5. The bridge safety monitoring system based on sensor data according to claim 4, characterized in that: In S1.1, the physical residual between the actual structural response measurement value and the structural response prediction value is calculated, and the adversarial learning method is applied to separate the environmental noise and the structural response to generate the structural response feature vector and the environmental feature, including the following steps: S1.

11. Build a physically driven structural response model based on bridge structural parameters, input the current load and boundary conditions, and output the theoretically predicted response value at the same sensor location; S1.

12. Collect the actual structural response measurement value of the sensor and calculate the physical residual signal; S1.

13. Design a generator and a discriminator to build an adversarial network. S1.

14. Randomly initialize the parameters of the generator and discriminator, and train them by alternately optimizing the generator and discriminator. S1.

15. Use the cross-entropy loss function to measure the performance of the generator and discriminator. S1.

16. After training, the generator is used to extract the structural response feature vector from the physical residual signal, and the generated structural response feature vector and environmental feature vector are standardized.

6. The bridge safety monitoring system based on sensor data according to claim 4, characterized in that: In S1.2, constructing a sensor graph structure to optimize frequency band division and extracting the main frequency amplitude and frequency band energy anomaly index of the aligned frequency includes the following steps: S1.

21. Collect bridge sensor deployment information and the physical connection relationship between each sensor; S1.22, build graph structure; Among them, each node in the graph structure corresponds to a sensor, and the weight of the edge represents the correlation coefficient of the structural response between the two sensors; S1.

23. Use graph structure to analyze the response consistency of each sensor at different frequencies; S1.24, cluster analysis is performed on the graph structure to identify subsets of sensors with similar frequency response patterns; S1.

25. Based on the cluster analysis results, retain the frequency bands within ±a% of the structural main frequency characteristics; S1.

26. Find the maximum amplitude and its corresponding frequency in the current frequency band, record the amplitude as the main frequency amplitude feature, calculate the energy of the current frequency band, and output the frequency band energy anomaly index.

7. The bridge safety monitoring system based on sensor data according to claim 4, characterized in that: In S1.4, a three-dimensional feature tensor is constructed, and a three-dimensional convolution operation is used to fuse the time domain, frequency domain, and spatial features to output the fused high-dimensional feature tensor, including the following steps: S1.

41. Determine tensor dimensions, where the tensor dimensions include the number of time windows, the number of physically aligned frequency bands, and the number of sensor nodes. S1.42, for each time window, each frequency band, and each sensor node, fill the corresponding eigenvalue into the three-dimensional tensor; S1.

43. Define a 3D convolutional layer using a deep learning framework and set the convolution kernel size. S1.

44. Pass the constructed 3D feature tensor as input to the 3D convolutional layer; S1.

45. After the three-dimensional convolution layer, a high-dimensional feature tensor is output that combines the time domain, frequency domain and spatial characteristics.

8. The sensor data-based bridge safety monitoring system according to claim 1, characterized in that: In S2.5, three types of degradation modes are constructed according to the k value, including the following steps: S2.

51. If the k value is less than the reference point k0 of the k value, use Gaussian distribution to generate multiple degradation increment samples to simulate the degradation situation in the future period, and add the generated degradation increments to the current degradation level to update the bridge performance degradation sequence; S2.52, if the k value is greater than or equal to the reference point of the k value and less than or equal to the upper limit k of the k value max , use Gamma distribution to generate multiple degradation increment samples to simulate the degradation situation in the future; S2.53, if the k value is greater than the upper limit k of the k value max , multiple degradation increment samples are generated using Weibull distribution to simulate the degradation situation in the future. The generated degradation increments are added to the current degradation level to update the bridge performance degradation sequence.

9. The bridge safety monitoring system based on sensor data according to claim 1, characterized in that: The information management unit (4) includes a storage management module (41), an alarm management module (42), a report generation module (43) and a video monitoring module (44); Wherein, the storage management module (41) stores bridge physical parameters and abnormality analysis results by establishing a database; The alarm management module (42) receives abnormal data from the data processing and analysis unit (3) and pushes alarm information according to a preset level; The report generation module (43) generates reports, historical data reports, and diagnostic reports for managing bridge data; The video monitoring module (44) is connected to and controls the bridge on-site camera for remote video monitoring.

Images