A method and system for highway bridge and tunnel structure health monitoring multi-source sensing data fusion analysis

Abstract

This application provides a method and system for multi-source sensor data fusion analysis in highway bridge and tunnel structural health monitoring, belonging to the field of civil engineering structural health monitoring technology. The method involves: acquiring monitoring data from various sensors and aligning them to generate a standardized monitoring data sequence; using historical standardized monitoring data sequences and corresponding environmental and load data to establish and dynamically update an adaptive benchmark model of the bridge and tunnel structural response; inputting the standardized monitoring data of the current moment into the adaptive benchmark model to obtain damage-sensitive residual signals, extracting multi-dimensional features, and then filtering and fusing them before inputting them into a health assessment model to obtain fused damage indicators and their confidence levels; fusing the damage indicators, confidence levels, and damage evidence from visual monitoring, and using evidence theory for comprehensive reasoning to generate a comprehensive structural health index and anomaly tracing information, and then issuing graded early warnings based on preset thresholds. This application accurately monitors bridge and tunnel health, improving operation and maintenance safety and efficiency.

Description

Technical Field

[0001] This application belongs to the field of civil engineering structural health monitoring technology, specifically relating to a method and system for multi-source sensor data fusion analysis of highway bridge and tunnel structural health monitoring. Background Technology

[0002] The bridge and tunnel structures of highways are crucial to transportation networks, and their safe operation is directly related to the safety of life and property and socio-economic development. Structural health monitoring of highway bridges and tunnels, by deploying various sensors on the structure to acquire information such as dynamic response, static deformation, and environmental loads in real time or periodically, is an important means of assessing structural condition and providing early warning of potential risks.

[0003] Currently, bridge and tunnel structural health monitoring typically integrates multiple sensing devices such as vibration sensors (e.g., accelerometers), strain sensors, crack monitoring sensors, and temperature and humidity sensors, accumulating massive amounts of multi-source heterogeneous monitoring data. However, existing data analysis methods are mostly limited to threshold alarms and simple trend analysis for single types of data or single indicators (e.g., frequency, deflection), which has significant shortcomings. First, the data is isolated; data from different types of sensors are not effectively aligned and correlated in time and space, resulting in fragmented information and making it difficult to form a unified understanding of the structural status. Second, the anti-interference capability is poor; the structural response is severely affected by operational factors such as ambient temperature and traffic load, and directly using raw data or simply filtered data can easily lead to false alarms or missed alarms. Third, the damage identification accuracy is low; there is a lack of an effective mechanism for deep fusion of multi-source information, making it insensitive to early and minor damage and difficult to distinguish damage from environmental / operational changes. Finally, the assessment results are simplistic, typically outputting a simple "normal / abnormal" binary judgment or a single health score, lacking in-depth analysis and quantitative assessment of the source, type, and degree of development of abnormalities, making it difficult to support accurate maintenance decisions.

[0004] Therefore, there is an urgent need for an analytical method that can deeply integrate multi-source sensor data, effectively isolate operational and environmental interference, and achieve high-precision damage identification and traceable health status assessment, so as to improve the intelligence level of bridge and tunnel structure health monitoring systems. Summary of the Invention

[0005] In a first aspect, embodiments of this application provide a method for multi-source sensor data fusion and analysis for health monitoring of highway bridge and tunnel structures, comprising the following steps: S1. Acquire raw monitoring data from various sensors deployed on bridge and tunnel structures, and perform time synchronization and spatial interpolation alignment processing on the raw monitoring data to generate a standardized monitoring data sequence with spatiotemporal alignment. S2. Using historical standardized monitoring data sequences and corresponding environmental and load data, establish and dynamically update an adaptive benchmark model of bridge and tunnel structural response; S3. Input the standardized monitoring data at the current moment into the adaptive baseline model to obtain the damage-sensitive residual signal stripped of environmental and load effects; S4. Extract multi-dimensional features from the damage-sensitive residual signal, and filter and fuse the features corresponding to different sensors in the multi-dimensional features to generate a fused feature vector; S5. Input the fused feature vector into the pre-trained health assessment model to calculate the fused damage index and its confidence level, which reflect the local and global structural states. S6. Integrate damage indicators, confidence levels, and damage evidence from visual monitoring, and use evidence theory to conduct comprehensive reasoning to generate a comprehensive structural health index and anomaly tracing information. Then, combine these with preset thresholds to issue graded early warnings.

[0006] Furthermore, the specific steps of step S1 are as follows: S11. Synchronously acquire the raw monitoring data streams from vibration sensors, strain sensors, temperature sensors, and crack sensors; S12. Use the master clock of the monitoring system as a reference to align the original monitoring data stream with timestamps; S13. For target analysis locations where sensors are not directly deployed, based on the known spatial coordinates of the sensors and the aligned monitoring data, the Kriging spatial interpolation algorithm is used to calculate the estimated physical quantities of the target analysis location at each time point, generating spatially continuous monitoring field data. S14. Output a standardized monitoring data sequence with spatiotemporal alignment of all sensor locations and interpolation locations.

[0007] Furthermore, the specific steps of step S2 are as follows: S21. Collect key response data from historical standardized monitoring data sequences. Ambient temperature data and traffic load data ; S22. Based on the collected historical standardized monitoring data sequences, the following adaptive benchmark model is constructed:

[0008] in, For nonlinear functions fitted by support vector regression or neural networks, This is the parameter vector of the time-varying model; Using a forgetting factor Online update of time-varying model parameter vector using recursive least squares method The updated formula is:

[0009] in, Here is the gain matrix and the forgetting factor. This is used to control the rate at which the model's memory of historical data decays, thereby enabling the adaptive baseline model to adapt to the gradual changes in the long-term performance of the structure.

[0010] Furthermore, key response data For at least one of the following physical quantities extracted from the standardized monitoring data sequence: The fundamental frequency of the structure, the strain value at the preset position, and the deflection value at the preset position.

[0011] Furthermore, the specific steps of step S3 are as follows: S31. Extract key response data from the standardized monitoring data at the current moment. Ambient temperature data and traffic load data Input into the updated adaptive baseline model To obtain the predicted response value at the current moment. ; S32. Calculate the damage-sensitive residual signal using the following formula. : .

[0012] Furthermore, the specific steps of step S4 are as follows: S41. Extract time-domain statistical features, frequency-domain spectral features, and time-frequency wavelet packet energy features in parallel from the damage-sensitive residual signal to form an initial multi-dimensional feature set; S42. The maximum correlation minimum redundancy algorithm is used to construct the following objective function to filter and fuse the initial multi-dimensional feature set, obtaining the optimal feature subset combination:

[0013] in, For the subset of features to be selected, For damage status labels, Indicates mutual information, Given the subset size, the objective function aims to maximize the correlation between the feature subset and the damage state while minimizing the redundancy between features. S43. Use the selected optimal feature subset to generate a fused feature vector.

[0014] Furthermore, the specific steps of step S5 are as follows: S51. Construct a lightweight deep belief network as the health assessment model, the network structure of which includes the following in sequence: An input layer has the same number of nodes as the dimension of the fused feature vector; At least one hidden layer composed of a restricted Boltzmann structure; An output layer using the Softmax function, the number of nodes in the output layer is the same as the number of preset damage state categories; S52. Collect historical fusion feature vector samples and corresponding damage state labels labeled manually or by a refined model to form a training dataset; S53. The contrastive divergence algorithm is used to perform unsupervised layer-by-layer pre-training of the lightweight deep belief network to initialize the network weight parameters; S54. Based on the pre-training, using the training dataset, the entire lightweight deep belief network is fine-tuned in a supervised manner through the error backpropagation algorithm to optimize the network parameters and obtain the trained health assessment model. S55. Input the fused feature vector into the trained health assessment model, and output the probability distribution vector. , of which each Indicates that the input feature belongs to the first... Predicted probability of damage state; S56. From the probability distribution vector In the middle, it will have the highest probability value. The damage status category is determined as fusion damage index And based on the probability distribution vector The probability concentration of each category is used to calculate the confidence level. :

[0015] in, For probability distribution The second highest probability value.

[0016] Furthermore, step S6 is detailed as follows: S61. Obtain the change in crack geometric features obtained through visual image analysis. As independent evidence of visual impairment; S62. Fusion damage indicators and corresponding confidence level Transformed into the first basic probability assignment function in DS evidence theory Evidence of visual impairment Transform into the second basic probability assignment function ; S63. The following improved Dempster combination rule is adopted to introduce a weighted allocation based on the reliability of the evidence source to address conflicts of evidence:

[0017]

[0018] Where A, B, C, and D represent propositions or subsets in the identification framework Θ, and the identification framework Θ contains at least two basic propositions: health and abnormality. The conflict factor is calculated according to the traditional DS rule: Used to quantize the first basic probability assignment function With the second basic probability assignment The degree of inconsistency between them; The weighted average probability assignment for proposition A; , The first basic probability assignment function is respectively Second basic probability allocation The normalized reliability weights satisfy ; S64. Assign based on the fused basic probability. The reliability score of the health proposition is calculated as the structural comprehensive health index. ,Right now ; S65. Analysis of the First Basic Probability Assignment With the second basic probability assignment The contribution to the final anomalous proposition is used to generate anomalous source information.

[0019] Furthermore, the specific method for issuing tiered early warnings in step S6 is as follows: Preset safety threshold With warning threshold ,and ; like The situation is deemed safe, and no warning will be issued. For structural comprehensive health index; like If the situation is determined to be in a state of alert, an alert-level warning will be issued, along with information on the source of the anomaly. like The situation is determined to be in a warning state, and a warning-level alert is issued, along with information on the source of the anomaly and suggested key detection areas.

[0020] Secondly, embodiments of this application also provide a multi-source sensor data fusion and analysis system for monitoring the health of highway bridge and tunnel structures, comprising: The data acquisition and preprocessing module is used to acquire raw monitoring data from various sensors deployed on bridge and tunnel structures, and to perform time synchronization and spatial interpolation alignment processing on the raw monitoring data to generate a standardized monitoring data sequence with spatiotemporal alignment. The adaptive benchmark modeling module is used to establish and dynamically update the adaptive benchmark model of bridge and tunnel structure response using historical standardized monitoring data sequences and corresponding environmental and load data. The damage signal extraction module is used to input the standardized monitoring data at the current moment into the adaptive baseline model to obtain the damage-sensitive residual signal that has been stripped of environmental and load effects. The feature fusion module is used to extract multi-dimensional features from damage-sensitive residual signals, and to filter and fuse the features corresponding to different sensors in the multi-dimensional features to generate a fused feature vector. The health assessment module is used to input the fused feature vector into the pre-trained health assessment model and calculate the fused damage index and its confidence level, which reflect the local and global structural states. The comprehensive assessment and early warning module is used to integrate damage indicators, confidence levels, and damage evidence from visual monitoring. Through comprehensive reasoning based on evidence theory, it generates a comprehensive structural health index and anomaly tracing information, and then issues graded early warnings based on preset thresholds.

[0021] As can be seen from the above technical solutions, this application has the following advantages: The multi-source sensor data fusion analysis method and system for highway bridge and tunnel structural health monitoring provided in this application improves the accuracy, timeliness, and reliability of bridge and tunnel structural damage identification through spatiotemporal alignment preprocessing, adaptive benchmark modeling, multi-dimensional feature fusion, and improved evidence theory reasoning. It effectively eliminates the interference of environment and load on monitoring results, reduces the probability of false alarms and missed alarms, and solves the problems of spatiotemporal misalignment of multi-source data, large environmental load interference, feature redundancy, and evidence conflict in traditional bridge and tunnel health monitoring. It realizes intelligent and precise monitoring throughout the entire process from data acquisition to early warning issuance. It can be adapted to different types of bridge and tunnel structures and sensor configurations. The generated anomaly tracing information and hierarchical early warning can provide accurate support for operation and maintenance decisions, extend the service life of bridges and tunnels, reduce operation and maintenance costs, and ensure the safety of highway traffic. Attached Figure Description

[0022] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the multi-source sensor data fusion and analysis method for health monitoring of highway bridge and tunnel structures according to the present invention.

[0024] Figure 2 This is a schematic diagram of the multi-source sensor data fusion analysis system for monitoring the structural health of highway bridges and tunnels according to the present invention. Detailed Implementation

[0025] The various embodiments of this disclosure will be described more fully in the following detailed description of the specific steps of the multi-source sensor data fusion analysis method for monitoring the structural health of highway bridges and tunnels. This disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of this disclosure to the specific embodiments disclosed herein, but rather this disclosure should be understood to cover all adjustments, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of this disclosure.

[0026] This embodiment provides a method for multi-source sensor data fusion analysis for highway bridge and tunnel structural health monitoring. It integrates multi-source sensor and visual data, removes environmental interference, accurately identifies bridge and tunnel damage, generates source tracing information and hierarchical early warning, and ensures traffic safety.

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Please see Figure 1 The diagram shows a flowchart of a multi-source sensor data fusion and analysis method for health monitoring of highway bridge and tunnel structures in a specific embodiment. The method includes the following steps: S1. Acquire raw monitoring data from various sensors deployed on bridge and tunnel structures, and perform time synchronization and spatial interpolation alignment processing on the raw monitoring data to generate a standardized monitoring data sequence with spatiotemporal alignment. It should be noted that the raw monitoring data, after standardization, forms a unified format and spatiotemporally consistent sequence, eliminating systematic errors in the data acquisition process. The spatiotemporally aligned standardized data provides a high-quality, highly consistent data source for subsequent steps such as adaptive benchmark modeling and damage signal extraction, avoiding model distortion and damage identification bias caused by data misalignment, and providing a foundation for the reliability of the entire analysis process. At the same time, standardization improves the versatility of the data, facilitating the fusion and application of data from different types of sensors. S2. Using historical standardized monitoring data sequences and corresponding environmental and load data, establish and dynamically update an adaptive benchmark model of bridge and tunnel structural response; It should be noted that this step constructs and dynamically updates an adaptive benchmark model, combining historical standardized data with environmental and load data, enabling the model to accurately characterize the response patterns of bridge and tunnel structures under normal working conditions. This overcomes the shortcomings of traditional fixed benchmark models, which cannot adapt to the gradual changes in long-term structural performance and fluctuations in environmental loads. The dynamic update mechanism optimizes model parameters in real time to ensure that the benchmark always matches the actual state of the structure, providing a basis for separating environmental and load effects, effectively improving the identification of damage signals, reducing the impact of external interference on monitoring results, and ensuring the accuracy of damage identification. S3. Input the standardized monitoring data at the current moment into the adaptive baseline model to obtain the damage-sensitive residual signal stripped of environmental and load effects; It should be noted that this step inputs the current standardized data into the adaptive benchmark model and obtains the damage-sensitive residual signal through difference calculation. This effectively separates the environmental and load effects, solving the problem that damage signals are masked by external interference signals in traditional monitoring. The damage-sensitive residual signal can focus on the damage characteristics of the structure itself, improve the signal-to-noise ratio, provide accurate core data for subsequent multi-dimensional feature extraction, avoid damage misjudgment caused by external factors, and provide support for accurate damage identification. S4. Extract multi-dimensional features from the damage-sensitive residual signal, and filter and fuse the features corresponding to different sensors in the multi-dimensional features to generate a fused feature vector; It should be noted that this step extracts multi-dimensional features from the residual signal and then filters and fuses them. The extraction of multi-domain features achieves comprehensive coverage of damage information and avoids missing key damage details due to single-domain features. The filtering and fusion of features from different sensors eliminates redundant information and retains core features, achieving complementary advantages of multi-source data. The generation of fused feature vectors not only reduces the computational cost of subsequent model training but also improves the ability of features to represent the damage state, solving the problem of fragmented information caused by traditional multi-sensor data working independently. This provides high-quality input for the health assessment model and enhances the reliability and accuracy of damage identification. S5. Input the fused feature vector into the pre-trained health assessment model to calculate the fused damage index and its confidence level, which reflect the local and global structural states. It should be noted that this step calculates the fusion damage index and confidence level through a pre-trained health assessment model, realizing a quantitative assessment of the local and overall health status of the structure. Compared with traditional qualitative assessment, this approach is more reasonable and objective. The fusion damage index can accurately reflect the degree and distribution of structural damage, while the confidence level quantifies the reliability of the assessment results, providing a precise and credible basis for comprehensive reasoning. This step takes into account both local and overall state assessment, avoiding the one-sidedness of single-dimensional assessment. At the same time, model-based processing improves assessment efficiency and meets the needs of real-time monitoring. S6. Integrate damage indicators, confidence levels, and damage evidence from visual monitoring, and use evidence theory to conduct comprehensive reasoning to generate a comprehensive structural health index and anomaly tracing information, and then combine these with preset thresholds to issue graded early warnings. It should be noted that this step integrates multi-source evidence and generates a comprehensive health index and anomaly tracing information through evidence theory reasoning. This achieves synergistic linkage between physical sensor data, visual monitoring data, and confidence levels, solving the problems of insufficient support from single evidence and difficulty in handling evidence conflicts in traditional assessments. Improved evidence theory enhances the robustness of multi-evidence fusion, the comprehensive health index makes structural health status intuitive and quantifiable, anomaly tracing information can accurately locate the source of anomalies, and tiered early warning enables refined management of health status. This step transforms the assessment results into actionable operation and maintenance guidance, providing decision support for operation and maintenance personnel, reducing the risk of structural failure, and ensuring traffic safety.

[0029] This embodiment uses spatiotemporal data alignment, adaptive benchmark modeling, and feature fusion, combined with improved evidence theory, to accurately remove environmental load interference, achieve quantitative assessment of bridge and tunnel damage, anomaly tracing and graded early warning, reduce operation and maintenance costs, and improve structural safety assurance capabilities.

[0030] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, in order to fully illustrate the specific implementation process in this embodiment, another method for multi-source sensor data fusion analysis for highway bridge and tunnel structural health monitoring is provided. Taking a two-way four-lane highway mountain tunnel as the monitoring object, the tunnel is 1200m long, with a main tunnel net width of 10.5m, a net height of 5.0m, and an operating life of 8 years as an example, the method includes the following steps: S1. Acquire raw monitoring data from various sensors deployed on the bridge and tunnel structure, and perform time synchronization and spatial interpolation alignment processing on the raw monitoring data to generate a standardized monitoring data sequence with spatiotemporal alignment; the specific steps of step S1 are as follows: S11. Synchronously acquire the raw monitoring data streams from vibration sensors, strain sensors, temperature sensors, and crack sensors; S12. Use the master clock of the monitoring system as a reference to align the original monitoring data stream with timestamps; S13. For target analysis locations where sensors are not directly deployed, based on the known spatial coordinates of the sensors and the aligned monitoring data, the Kriging spatial interpolation algorithm is used to calculate the estimated physical quantities of the target analysis location at each time point, generating spatially continuous monitoring field data. S14. Output the spatiotemporally aligned standardized monitoring data sequence of all sensor points and interpolation points; For example, the section from K0+200 to K0+800 of the tunnel was selected as the core monitoring area, and multiple types of sensors were deployed: 12 vibration sensors (sampling frequency 100Hz), 18 strain sensors (sampling frequency 50Hz), 20 temperature sensors (sampling frequency 1Hz), and 8 crack sensors (sampling frequency 0.1Hz) were deployed at the key sections of the arch, sidewalls, and invert. At the same time, a high-definition industrial camera (frame rate 15fps) was configured for visual monitoring. Simultaneously collect raw data streams from each sensor: vibration sensor collects vibration acceleration signals of tunnel structure, strain sensor collects concrete strain values, temperature sensor collects ambient and structural surface temperature, and crack sensor collects crack width data. Using the monitoring system's master clock (accuracy 1μs) as a reference, a unified timestamp is added to all raw data streams to correct the acquisition delay of vibration and temperature sensors (maximum delay 0.02s) and achieve timestamp alignment. For key stress sections such as K0+350 and K0+550 of the tunnel where no sensors are deployed, based on the spatial coordinates (error ±5cm) of three adjacent sensors and the aligned data, the Kriging spatial interpolation algorithm (interpolation error ≤3%) is used to calculate the strain and temperature estimates of the target section at each time point, generating spatially continuous monitoring field data. Output spatiotemporally aligned and standardized monitoring data sequences of 32 sensor points and 15 interpolation points. The data format is uniformly JSON for easy access by subsequent modules. S2. Using historical standardized monitoring data sequences and corresponding environmental and load data, establish and dynamically update an adaptive benchmark model for the bridge and tunnel structure response; the specific steps of step S2 are as follows: S21. Collect key response data from historical standardized monitoring data sequences. Ambient temperature data and traffic load data Key response data For at least one of the following physical quantities extracted from the standardized monitoring data sequence: The fundamental frequency of the structure, the strain value at the preset position, and the deflection value at the preset position; S22. Based on the collected historical standardized monitoring data sequences, the following adaptive benchmark model is constructed:

[0031] in, For nonlinear functions fitted by support vector regression or neural networks, This is the parameter vector of the time-varying model; Using a forgetting factor Online update of time-varying model parameter vector using recursive least squares method The updated formula is:

[0032] in, Here is the gain matrix and the forgetting factor. The range of values ​​is This is used to control the rate at which the model's memory decays over historical data, thereby enabling the adaptive baseline model to adapt to the gradual changes in the long-term performance of the structure. For example, a standardized monitoring data sequence of the tunnel over the past two years was collected, and key response data X(t) (including crown deflection value, sidewall strain value, and structural fundamental frequency, where the fundamental frequency was obtained through vibration signal spectrum analysis with an accuracy of ±0.01Hz), ambient temperature data T(t) (range -5℃ to 38℃), and traffic load data L(t) (average daily traffic flow of 8,000-12,000 vehicles, converted to equivalent load by vehicle type) were extracted. Valid historical data (excluding data from extreme weather and traffic control periods, with valid data accounting for 92%) were selected to construct a dataset, in which the arch deflection value (unit mm, range 0-50) and the maximum strain value of the sidewall (unit με, range -200 to 200) were selected as the core response parameters for X(t); Constructing an adaptive baseline model: ,in A BP neural network was used for fitting (3 nodes in the input layer, 2 hidden layers, and 2 nodes in the output layer), and θ(t) is the time-varying model parameter vector (the dimension is the weight matrix of the hidden layer and the output layer, with a total of 64 parameters). Online update using recursive least squares with forgetting factor λ Let λ = 0.98, and set the initial value of the gain matrix K(t) as the identity matrix, using the formula... Update the parameters hourly to adapt the model to the long-term performance changes caused by tunnel concrete creep (monthly average deflection change ≤ 0.1 mm). S3. Input the standardized monitoring data at the current moment into the adaptive baseline model to obtain the damage-sensitive residual signal stripped of environmental and load effects; The specific steps of step S3 are as follows: S31. Extract key response data from the standardized monitoring data at the current moment. Ambient temperature data and traffic load data Input into the updated adaptive baseline model To obtain the predicted response value at the current moment. ; S32. Calculate the damage-sensitive residual signal using the following formula. : ; For example, the morning rush hour (7:00-8:00) on a certain weekday is selected as the current monitoring period, and the ambient temperature during this period is... =18℃, equivalent traffic load =25kN / m², extract the crown deflection from the current standardized data. =8.2mm, sidewall strain =-125με.

[0033] Will , , Input the updated adaptive baseline model to obtain the response prediction value. (Ceiling deflection 8.0 mm, sidewall strain -122 με); Calculate damage-sensitive residual signal The residual value of the arch deflection is 8.2-8.0=0.2mm, and the residual value of the sidewall strain is -125-(-122)=-3με. The residual signals are all within a reasonable fluctuation range (preset residual thresholds ±0.5mm and ±5με). S4. Extract multi-dimensional features from the damage-sensitive residual signal, and filter and fuse the features corresponding to different sensors in the multi-dimensional features to generate a fused feature vector; The specific steps of step S4 are as follows: S41. Extract time-domain statistical features, frequency-domain spectral features, and time-frequency wavelet packet energy features in parallel from the damage-sensitive residual signal to form an initial multi-dimensional feature set; S42. The maximum correlation minimum redundancy algorithm is used to construct the following objective function to filter and fuse the initial multi-dimensional feature set, obtaining the optimal feature subset combination:

[0034] in, For the subset of features to be selected, For damage status labels, Indicates mutual information, Given the subset size, the objective function aims to maximize the correlation between the feature subset and the damage state while minimizing the redundancy between features. S43. Use the selected optimal feature subset to generate a fused feature vector; For example, based on the residual signal obtained in step S3, multi-dimensional feature extraction and fusion are performed: The initial 20-dimensional feature set is composed of time-domain statistical features (mean, variance, peak factor, a total of 6 features), frequency-domain spectral features (fundamental frequency, harmonic amplitude, a total of 4 features), and time-frequency domain wavelet packet energy features (decomposed into 4 layers, extracting 8 frequency bands of energy, a total of 8 features). The maximum correlation minimum redundancy algorithm is used to screen features. Let the damage state label be Y (healthy=0, minor damage=1, severe damage=2). Through the objective function calculation, 8 redundant features (mutual information with Y <0.1) are eliminated, and 12 core features (including strain residual variance, wavelet packet third layer energy, etc.) are retained. The selected features are normalized (mapped to the 0-1 interval) to generate a 12-dimensional fused feature vector: [0.32,0.45,0.28,0.51,0.37,0.62,0.41,0.29,0.53,0.34,0.48,0.27]; S5. Input the fused feature vector into the pre-trained health assessment model to calculate the fused damage index and its confidence level, which reflect the local and global structural states. The specific steps of step S5 are as follows: S51. Construct a lightweight deep belief network as the health assessment model, the network structure of which includes the following in sequence: An input layer has the same number of nodes as the dimension of the fused feature vector; At least one hidden layer composed of a restricted Boltzmann structure; An output layer using the Softmax function, the number of nodes in the output layer is the same as the number of preset damage state categories; S52. Collect historical fusion feature vector samples and corresponding damage state labels labeled manually or by a refined model to form a training dataset; S53. The contrastive divergence algorithm is used to perform unsupervised layer-by-layer pre-training of the lightweight deep belief network to initialize the network weight parameters; S54. Based on the pre-training, using the training dataset, the entire lightweight deep belief network is fine-tuned in a supervised manner through the error backpropagation algorithm to optimize the network parameters and obtain the trained health assessment model. S55. Input the fused feature vector into the trained health assessment model, and output the probability distribution vector. , of which each Indicates that the input feature belongs to the first... Predicted probability of damage state; S56. From the probability distribution vector In the middle, it will have the highest probability value. The damage status category is determined as fusion damage index And based on the probability distribution vector The probability concentration of each category is used to calculate the confidence level. :

[0035] in, For probability distribution The second most probable value in the range; For example, a lightweight deep belief network is constructed as a health assessment model. The network structure is as follows: 12 nodes in the input layer (corresponding to the dimension of the fused feature vector), 2 hidden layers (containing 16 and 8 restricted Boltzmann machine nodes respectively), and 3 nodes in the output layer (corresponding to 3 types of damage states). 5000 historical fusion feature vector samples were collected (including 100 samples of minor damage and 50 samples of severe damage). The model was pre-trained using the contrastive divergence algorithm (100 iterations, learning rate 0.01) and then fine-tuned using the backpropagation algorithm (batch size 32, 50 iterations). The model achieved an accuracy of 96.8% in testing. Input the fused feature vector generated in step S4 into the model, and output the probability distribution vector P=[0.92,0.06,0.02], which represents a 92% probability of being in a healthy state, a 6% probability of minor injury, and a 2% probability of severe injury. Determine fusion damage index =Health (0), second highest probability value =0.06, calculate the confidence level. =1-0.06=0.94, indicating that the evaluation result is highly reliable; S6. Integrate damage indicators, confidence levels, and damage evidence from visual monitoring, and use evidence theory to conduct comprehensive reasoning to generate a comprehensive structural health index and anomaly tracing information, and then combine these with preset thresholds to issue graded early warnings. The specific steps of step S6 are as follows: S61. Obtain the change in crack geometric features obtained through visual image analysis. As independent evidence of visual impairment; S62. Fusion damage indicators and corresponding confidence level Transformed into the first basic probability assignment function in DS evidence theory Evidence of visual impairment Transform into the second basic probability assignment function ; S63. The following improved Dempster combination rule is adopted to introduce a weighted allocation based on the reliability of the evidence source to address conflicts of evidence:

[0036]

[0037] Where A, B, C, and D represent propositions or subsets in the identification framework Θ, and the identification framework Θ contains at least two basic propositions: health and abnormality. The conflict factor is calculated according to the traditional DS rule: Used to quantize the first basic probability assignment function With the second basic probability assignment The degree of inconsistency between them; The weighted average probability assignment for proposition A; , The first basic probability assignment function is respectively Second basic probability allocation The normalized reliability weights satisfy ; S64. Assign based on the fused basic probability. The reliability score of the health proposition is calculated as the structural comprehensive health index. ,Right now ; S65. Analysis of the First Basic Probability Assignment With the second basic probability assignment The contribution to the final anomalous proposition is used to generate anomalous source information; The specific method for issuing tiered early warnings in step S6 is as follows: Preset safety threshold With warning threshold ,and ; like The situation is deemed safe, and no warning will be issued. For structural comprehensive health index; like If the situation is determined to be in a state of alert, an alert-level warning will be issued, along with information on the source of the anomaly. like If the situation is determined to be in a warning state, a warning-level alert will be issued, along with information on the source of the anomaly and suggested key areas for detection. For example, combining visual monitoring with evidence fusion reasoning, the specific implementation is as follows: Images of the tunnel sidewalls and arch were captured using high-definition cameras, and the changes in the geometric features of the cracks were obtained through image recognition. =0.02mm (the increase in crack width over the past 30 days, with a preset visual safety threshold of 0.1mm), as evidence of visual damage; Transform D = healthy, C = 0.94 into the first basic probability assignment m1 (healthy = 0.94, minor injury = 0.05, severe injury = 0.01); =0.02mm is converted into the second basic probability allocation m2 (healthy = 0.93, minor injury = 0.06, severe injury = 0.01). An improved Dempster combination rule was used for fusion, with reliability weights w1=0.9 (determined based on confidence C) and w2=0.1 (determined based on image quality assessment). The evidence conflict factor k=0.007 (minimum conflict) was calculated and substituted into the formula to obtain the basic probability allocation m after fusion (healthy=0.939, minor injury=0.058, severe injury=0.003). Calculate the overall health index of the structure =0.939; Analyzing the contribution of evidence, both m1 and m2 play a dominant role in the health proposition, with no abnormal source information. Preset safety threshold =0.85, warning threshold =0.7, because =0.939≥ The status is determined to be safe, and no warning will be issued.

[0038] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0039] like Figure 2 As shown, the following are embodiments of the highway bridge and tunnel structure health monitoring multi-source sensor data fusion analysis system provided in this disclosure. This system and the highway bridge and tunnel structure health monitoring multi-source sensor data fusion analysis method of the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the highway bridge and tunnel structure health monitoring multi-source sensor data fusion analysis system, please refer to the embodiments of the highway bridge and tunnel structure health monitoring multi-source sensor data fusion analysis method of the above embodiments.

[0040] The system includes: The data acquisition and preprocessing module is used to acquire raw monitoring data from various sensors deployed on bridge and tunnel structures, and to perform time synchronization and spatial interpolation alignment processing on the raw monitoring data to generate a standardized monitoring data sequence with spatiotemporal alignment. The adaptive benchmark modeling module is used to establish and dynamically update the adaptive benchmark model of bridge and tunnel structure response using historical standardized monitoring data sequences and corresponding environmental and load data. The damage signal extraction module is used to input the standardized monitoring data at the current moment into the adaptive baseline model to obtain the damage-sensitive residual signal that has been stripped of environmental and load effects. The feature fusion module is used to extract multi-dimensional features from damage-sensitive residual signals, and to filter and fuse the features corresponding to different sensors in the multi-dimensional features to generate a fused feature vector. The health assessment module is used to input the fused feature vector into the pre-trained health assessment model and calculate the fused damage index and its confidence level, which reflect the local and global structural states. The comprehensive assessment and early warning module is used to integrate damage indicators, confidence levels, and damage evidence from visual monitoring. Through comprehensive reasoning based on evidence theory, it generates a comprehensive structural health index and anomaly tracing information, and then issues graded early warnings based on preset thresholds.

[0041] This embodiment achieves the integration of multi-source sensor and visual data, the removal of environmental interference, accurate identification of bridge and tunnel damage, generation of source tracing information and graded early warning, and ensures traffic safety through the interactive collaboration of data acquisition and preprocessing module, adaptive benchmark modeling module, damage signal extraction module, feature fusion module, health assessment module, and comprehensive assessment and early warning module.

[0042] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for multi-source sensing data fusion analysis of highway bridge and tunnel structure health monitoring, characterized in that, Includes the following steps: S1. Acquire raw monitoring data from various sensors deployed on bridge and tunnel structures, and perform time synchronization and spatial interpolation alignment processing on the raw monitoring data to generate a standardized monitoring data sequence with spatiotemporal alignment. S2. Using historical standardized monitoring data sequences and corresponding environmental and load data, establish and dynamically update an adaptive benchmark model of bridge and tunnel structural response; S3. Input the standardized monitoring data at the current moment into the adaptive baseline model to obtain the damage-sensitive residual signal stripped of environmental and load effects; S4. Extract multi-dimensional features from the damage-sensitive residual signal, and filter and fuse the features corresponding to different sensors in the multi-dimensional features to generate a fused feature vector; S5. Input the fused feature vector into the pre-trained health assessment model to calculate the fused damage index and its confidence level, which reflect the local and global structural states. S6. Integrate damage indicators, confidence levels, and damage evidence from visual monitoring, and use evidence theory to conduct comprehensive reasoning to generate a comprehensive structural health index and anomaly tracing information. Then, combine these with preset thresholds to issue graded early warnings.

2. The method for multi-source sensor data fusion and analysis for highway bridge and tunnel structural health monitoring according to claim 1, characterized in that, The specific steps of step S1 are as follows: S11. Synchronously acquire the raw monitoring data streams from vibration sensors, strain sensors, temperature sensors, and crack sensors; S12. Use the master clock of the monitoring system as a reference to align the original monitoring data stream with timestamps; S13. For target analysis locations where sensors are not directly deployed, based on the known spatial coordinates of the sensors and the aligned monitoring data, the Kriging spatial interpolation algorithm is used to calculate the estimated physical quantities of the target analysis location at each time point, generating spatially continuous monitoring field data. S14. Output a standardized monitoring data sequence with spatiotemporal alignment of all sensor locations and interpolation locations.

3. The method for multi-source sensor data fusion and analysis for highway bridge and tunnel structural health monitoring according to claim 1, characterized in that, The specific steps of step S2 are as follows: S21. Collect key response data from historical standardized monitoring data sequences. Ambient temperature data and traffic load data ; S22. Based on the collected historical standardized monitoring data sequences, the following adaptive benchmark model is constructed: in, For nonlinear functions fitted by support vector regression or neural networks, This is the parameter vector of the time-varying model; Using a forgetting factor Online update of time-varying model parameter vector using recursive least squares method The updated formula is: in, Here is the gain matrix and the forgetting factor. This is used to control the rate at which the model's memory of historical data decays, thereby enabling the adaptive baseline model to adapt to the gradual changes in the long-term performance of the structure.

4. The method for multi-source sensor data fusion and analysis for highway bridge and tunnel structural health monitoring according to claim 3, characterized in that, Key Response Data For at least one of the following physical quantities extracted from the standardized monitoring data sequence: The fundamental frequency of the structure, the strain value at the preset position, and the deflection value at the preset position.

5. The method for multi-source sensor data fusion and analysis for highway bridge and tunnel structural health monitoring according to claim 1, characterized in that, The specific steps of step S3 are as follows: S31. Extract key response data from the standardized monitoring data at the current moment. Ambient temperature data and traffic load data Input into the updated adaptive baseline model To obtain the predicted response value at the current moment. ; S32. Calculate the damage-sensitive residual signal using the following formula. : 。 6. The method for multi-source sensor data fusion and analysis for highway bridge and tunnel structural health monitoring according to claim 1, characterized in that, The specific steps of step S4 are as follows: S41. Extract time-domain statistical features, frequency-domain spectral features, and time-frequency wavelet packet energy features in parallel from the damage-sensitive residual signal to form an initial multi-dimensional feature set; S42. The maximum correlation minimum redundancy algorithm is used to construct the following objective function to filter and fuse the initial multi-dimensional feature set, obtaining the optimal feature subset combination: in, For the subset of features to be selected, For damage status labels, Indicates mutual information, Given the subset size, the objective function aims to maximize the correlation between the feature subset and the damage state while minimizing the redundancy between features. S43. Use the selected optimal feature subset to generate a fused feature vector.

7. The method for multi-source sensor data fusion and analysis for highway bridge and tunnel structural health monitoring according to claim 1, characterized in that, The specific steps of step S5 are as follows: S51. Construct a lightweight deep belief network as the health assessment model, the network structure of which includes the following in sequence: An input layer has the same number of nodes as the dimension of the fused feature vector; At least one hidden layer composed of a restricted Boltzmann structure; An output layer using the Softmax function, the number of nodes in the output layer is the same as the number of preset damage state categories; S52. Collect historical fusion feature vector samples and corresponding damage state labels labeled manually or by a refined model to form a training dataset; S53. The contrastive divergence algorithm is used to perform unsupervised layer-by-layer pre-training of the lightweight deep belief network to initialize the network weight parameters; S54. Based on the pre-training, using the training dataset, the entire lightweight deep belief network is fine-tuned in a supervised manner through the error backpropagation algorithm to optimize the network parameters and obtain the trained health assessment model. S55. Input the fused feature vector into the trained health assessment model, and output the probability distribution vector. , of which each Indicates that the input feature belongs to the first... Predicted probability of damage state; S56. From the probability distribution vector In the middle, it will have the highest probability value. The damage status category is determined as fusion damage index And based on the probability distribution vector The probability concentration of each category is used to calculate the confidence level. : in, For probability distribution The second highest probability value.

8. The method for multi-source sensor data fusion and analysis for health monitoring of highway bridge and tunnel structures according to claim 1, characterized in that, The specific steps of step S6 are as follows: S61. Obtain the change in crack geometric features obtained through visual image analysis. As independent evidence of visual impairment; S62. Fusion damage indicators and corresponding confidence level Transformed into the first basic probability assignment function in DS evidence theory Evidence of visual impairment Transform into the second basic probability assignment function ; S63. The following improved Dempster combination rule is adopted to introduce a weighted allocation based on the reliability of the evidence source to address conflicts of evidence: Where A, B, C, and D represent propositions or subsets in the identification framework Θ, and the identification framework Θ contains at least two basic propositions: health and abnormality. The conflict factor is calculated according to the traditional DS rule: Used to quantize the first basic probability assignment function With the second basic probability assignment The degree of inconsistency between them; The weighted average probability assignment for proposition A; , The first basic probability assignment function is respectively Second basic probability allocation The normalized reliability weights satisfy ; S64. Assign based on the fused basic probability. The reliability score of the health proposition is calculated as the structural comprehensive health index. ,Right now ; S65. Analysis of the First Basic Probability Assignment With the second basic probability assignment The contribution to the final anomalous proposition is used to generate anomalous source information.

9. The method for multi-source sensor data fusion and analysis for highway bridge and tunnel structural health monitoring according to claim 8, characterized in that, The specific method for issuing tiered early warnings in step S6 is as follows: Preset safety threshold With warning threshold ,and ; like The situation is deemed safe, and no warning will be issued. For structural comprehensive health index; like If the situation is determined to be in a state of alert, an alert-level warning will be issued, along with information on the source of the anomaly. like The situation is determined to be in a warning state, and a warning-level alert is issued, along with information on the source of the anomaly and suggested key detection areas.

10. A multi-source sensor data fusion and analysis system for monitoring the structural health of highway bridges and tunnels, characterized in that, include: The data acquisition and preprocessing module is used to acquire raw monitoring data from various sensors deployed on bridge and tunnel structures, and to perform time synchronization and spatial interpolation alignment processing on the raw monitoring data to generate a standardized monitoring data sequence with spatiotemporal alignment. The adaptive benchmark modeling module is used to establish and dynamically update the adaptive benchmark model of bridge and tunnel structure response using historical standardized monitoring data sequences and corresponding environmental and load data. The damage signal extraction module is used to input the standardized monitoring data at the current moment into the adaptive baseline model to obtain the damage-sensitive residual signal that has been stripped of environmental and load effects. The feature fusion module is used to extract multi-dimensional features from damage-sensitive residual signals, and to filter and fuse the features corresponding to different sensors in the multi-dimensional features to generate a fused feature vector. The health assessment module is used to input the fused feature vector into the pre-trained health assessment model and calculate the fused damage index and its confidence level, which reflect the local and global structural states. The comprehensive assessment and early warning module is used to integrate damage indicators, confidence levels, and damage evidence from visual monitoring. Through comprehensive reasoning based on evidence theory, it generates a comprehensive structural health index and anomaly tracing information, and then issues graded early warnings based on preset thresholds.

Images