Full-time-domain structural health monitoring system and method

Abstract

The invention relates to the technical field of structural health monitoring, and discloses a full-time-domain structural health monitoring system and method. The system comprises a multi-sensor data acquisition module, a dynamic modeling module, a feature extraction module, a health state evaluation module, a self-adaptive adjustment module and an anomaly analysis module. The multi-sensor data acquisition module acquires multi-modal data of the structural body in real time; the dynamic modeling module generates a dynamic response model adaptive to the actual operation state of the structural body; the feature extraction module extracts features of the model and compresses the features to generate compressed feature data; the health state evaluation module analyzes structural health parameters based on the data, and generates an intuitive health state rating; the self-adaptive adjustment module constructs a safety threshold according to the rating, and generates a real-time monitoring strategy to reasonably allocate resources; the exception analysis module analyzes the deviation between the monitoring data and the strategy and generates exception mark data. The system realizes full-time-domain and omnibearing monitoring, and ensures the safe operation of the structure body.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring technology, specifically to a full-time-domain structural health monitoring system and method. Background Technology

[0002] In the field of engineering construction, the safe and stable operation of structures has always been a key focus of the industry. Whether it is bridges, buildings, large machinery and equipment, or water conservancy facilities, they are all subject to the dual effects of external environmental factors and internal aging during long-term use, gradually leading to performance degradation and accumulated damage. If these problems are not detected and addressed in a timely manner, they may, over time, cause safety hazards to the structure, or even lead to serious safety accidents, resulting in huge casualties and property losses.

[0003] Currently, structural health monitoring methods are gradually shifting from traditional manual inspection to automated monitoring. Traditional manual inspection methods often require personnel to go to the site of the structure and obtain its status information through visual observation and handheld devices. This method is not only inefficient, but also limited by the experience of the inspectors and the inspection environment, making it difficult to achieve full-time, all-round monitoring of the structure. This is especially true for some large and complex structures or structures in harsh environments, where manual inspection is even more difficult, and the accuracy and reliability of the inspection results are hard to guarantee.

[0004] With the development of sensor and data processing technologies, automated monitoring systems have begun to be applied in the field of structural health monitoring. Existing automated monitoring systems typically deploy a number of sensors on the structure to collect data on parameters such as vibration, strain, and temperature. The collected data is then processed and analyzed to determine the structure's health status. However, existing automated monitoring systems still have many shortcomings in practical applications. On the one hand, most monitoring systems can only collect data from a single type of sensor, failing to achieve simultaneous acquisition of multimodal data. This results in incomplete information about the structure's condition, making it difficult to accurately reflect the overall health status of the structure. For example, some systems only collect vibration data, ignoring the influence of parameters such as strain and temperature on the structure's health status, making it prone to misjudgments or omissions when analyzing structural damage.

[0005] Existing monitoring systems lack effective dynamic modeling capabilities in data processing and analysis. Most rely on fixed analytical models to process collected data, failing to dynamically adjust model parameters based on the actual operating status of the structure and data changes. This results in poor model adaptability and accuracy. Furthermore, in feature extraction, existing systems often cannot effectively compress the extracted feature data, leading to massive data volumes. This not only increases data storage and transmission costs but also affects the efficiency of subsequent health status assessments. In addition, the health status assessment standards of existing monitoring systems are relatively fixed, unable to dynamically adjust safety thresholds and monitoring strategies based on real-time monitoring data and the structure's health status. When abnormalities occur, the monitoring frequency and scope cannot be adjusted promptly, resulting in slow response times and hindering early warning and timely handling of structural safety hazards.

[0006] Existing monitoring systems lack effective anomaly analysis capabilities when they detect deviations from normal data. They can only provide a simple indication of an anomaly, failing to accurately analyze its cause, location, and severity. This makes it difficult for staff to take targeted repair and maintenance measures, increasing maintenance costs and potentially leading to further structural damage due to delayed handling of anomalies, thus affecting the safe operation and service life of the structure. Therefore, the field of structural health monitoring currently needs a system capable of full-time, multi-modal data acquisition, efficient dynamic modeling, accurate feature extraction and compression, flexible health status assessment, adaptive monitoring strategy adjustment, and accurate anomaly analysis. This system would address the shortcomings of existing monitoring systems and meet the needs of the structural health monitoring field for efficient, accurate, and comprehensive monitoring. Summary of the Invention

[0007] The purpose of this invention is to provide a full-time structural health monitoring system and method to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides a full-time-domain structural health monitoring system, the system comprising:

[0009] A multi-sensor data acquisition module is used to acquire multi-modal sensor data of the structure in real time.

[0010] A dynamic modeling module is used to generate a dynamic response model of the structure based on the multimodal sensor data;

[0011] The feature extraction module is used to extract and compress features from the dynamic response model to generate compressed feature data.

[0012] The health status assessment module is used to perform structural health status parameter analysis based on the compressed feature data and generate a health status rating.

[0013] An adaptive adjustment module is used to construct a safety threshold based on the health status rating and to generate a real-time monitoring strategy using the safety threshold.

[0014] The anomaly analysis module is used to analyze the deviation between real-time monitoring data and the preset standard health status data of the real-time monitoring strategy, and generate structural anomaly marker data, including anomaly type, location of occurrence, severity and occurrence time.

[0015] Preferably, the dynamic modeling module includes:

[0016] Key feature points are extracted from the synchronous sensor data stream to generate structural response vector nodes;

[0017] The structural response vector nodes are fused with timestamp information to form a dynamic response model.

[0018] Preferably, the feature extraction module includes:

[0019] The dynamic response model is hierarchically encoded to generate a multi-dimensional feature matrix;

[0020] Principal components in the multi-dimensional feature matrix are filtered by a preset feature importance threshold to generate compressed feature data.

[0021] Preferably, the health status assessment module includes:

[0022] Pattern recognition is performed on the compressed feature data to generate a structural damage feature vector;

[0023] The safety coefficient of the structural damage feature vector is evaluated by combining the preset structural health database, and a health status rating is generated.

[0024] Preferably, the adaptive adjustment module includes:

[0025] A safety threshold corridor is established based on the health status rating. The safety threshold corridor consists of two boundary values, an upper limit and a lower limit, which is a reasonable range for safe operation formed by the area between the two thresholds.

[0026] The safety threshold corridor is used to search for feasible monitoring strategies and screen strategies that meet the preset continuity rules to obtain a set of candidate strategies. The monitoring strategy dynamically adjusts the safety threshold according to the actual health status rating of the structure. When the health status of the structure is good, the safety threshold can be appropriately relaxed and the monitoring frequency can be reduced. When the health status of the structure shows a downward trend or the rating is low, the safety threshold range is narrowed in time, the monitoring frequency is increased, and the monitoring range is expanded.

[0027] Set weight coefficients for each candidate strategy in the candidate strategy set to generate an adaptive strategy library;

[0028] Target strategies are selected from the adaptive strategy library according to the weight coefficients, and the target strategies are optimized to generate real-time monitoring strategies.

[0029] Preferably, the system further includes:

[0030] Obtain the modal parameters of the structure under its current vibration state, calculate the probability of future risks using the modal parameters, and generate a risk distribution map;

[0031] The risk distribution map is overlaid with the health status rating to generate weighting coefficients.

[0032] Preferably, the anomaly parsing module includes:

[0033] Obtain the difference parameters between preset standard health status data and real-time monitoring data;

[0034] Using the difference parameters, multi-layer semantic anomaly labeling data is generated, including high-level semantic anomaly labels and low-level semantic anomaly labels. The high-level semantic anomaly labels include anomaly level, anomaly type, and anomaly confidence, while the low-level semantic anomaly labels only record anomaly prompt information and timestamps.

[0035] The multi-layer semantic anomaly labeling data is used to identify anomaly types and generate structural anomaly labeling data.

[0036] The process of generating multi-layer semantic anomaly marker data using the difference parameters includes:

[0037] A joint analysis of the aforementioned difference parameters is performed to generate a confidence score;

[0038] The confidence score is compared with a preset anomaly confidence threshold; based on the judgment result, multi-layer semantic anomaly labeling data is generated.

[0039] Preferably, the system further includes:

[0040] A communication control module is used to dynamically adjust network bandwidth allocation and transmit the structural anomaly marker data to the monitoring center; the communication control module includes:

[0041] Real-time monitoring of signal quality parameters and transmission load parameters of primary and backup communication channels;

[0042] The channel priority score is calculated using the signal quality parameters and the transmission load parameters.

[0043] Bandwidth resources are allocated based on the channel priority score, and channel switching decisions are generated.

[0044] Based on the channel switching decision, the network bandwidth allocation is adjusted, and the structural anomaly marker data is transmitted to the monitoring center.

[0045] Preferably, the multi-sensor data acquisition module includes:

[0046] Acquire sensor data sequences and timestamp information under a preset sampling period;

[0047] The sensor data sequence is validated, and the multi-sensor data is aligned through a data synchronization mechanism to generate a synchronized sensor data stream.

[0048] Preferably, the present invention also includes a full-time-domain structural health monitoring method, the method comprising all modules and method flow of the full-time-domain structural health monitoring system described above.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] The multi-sensor data acquisition module enables real-time acquisition of multimodal sensor data from the structure, breaking the limitation of existing monitoring systems that can only collect single-type data. It can simultaneously acquire various parameter data such as vibration, strain, temperature, and displacement of the structure, making the acquired structural status information more comprehensive and abundant. This multimodal data acquisition method can reflect the operational status and health condition of the structure from different dimensions, avoiding the one-sidedness of information caused by a single data type. It provides more sufficient and comprehensive data support for subsequent dynamic modeling and health status assessment, enabling the judgment of the structure's health status to be based on more complete information, thus more closely reflecting the actual situation of the structure.

[0051] The dynamic modeling module generates a dynamic response model of the structure based on multimodal sensor data. Compared to existing monitoring systems that use fixed models, this module can fully utilize the rich information contained in multimodal data, combining the actual operating state of the structure with data change trends, and dynamically adjust the model's parameters and structure. This allows the generated dynamic response model to more accurately simulate the structure's operating characteristics under different working conditions. This dynamic modeling capability makes the model more adaptable. Whether the structure is in normal operation or is subject to external environmental interference or minor damage, the model can adjust in a timely manner according to data changes, thus more accurately reflecting the dynamic changes of the structure. This lays a solid foundation for subsequent feature extraction and health status assessment, and helps improve the accuracy of structural health status analysis.

[0052] The feature extraction module extracts and compresses features from the dynamic response model, generating compressed feature data. This process not only extracts key feature information reflecting the structure's health status from the dynamic response model but also effectively compresses the extracted feature data. Feature extraction removes redundant information from the data, focusing on core features closely related to the structure's health status and avoiding interference from irrelevant information in subsequent evaluation stages. Feature compression significantly reduces the data volume, lowering the pressure on data storage and transmission while ensuring no loss of key feature information. This reduces costs associated with storage devices and transmission bandwidth, while also improving the data processing efficiency of the subsequent health status evaluation module. It prevents delays in the evaluation process due to excessive data volume, ensuring timely acquisition of the structure's health status rating results.

[0053] The health status assessment module analyzes structural health status parameters and generates a health status rating based on compressed feature data. This module fully utilizes key information from the compressed feature data to conduct a detailed and comprehensive analysis of the structure's health status parameters. Through in-depth mining of the feature data, it can accurately identify whether the structure has damage, the approximate type of damage, and generate a clear and explicit health status rating according to a pre-set assessment system, allowing staff to intuitively understand the structure's current health status. This assessment method based on compressed feature data ensures the accuracy of the assessment results while improving assessment efficiency due to data compression, enabling the assessment process to be completed quickly and efficiently, and providing timely feedback on the structure's health status to staff.

[0054] The adaptive adjustment module constructs safety thresholds based on health status ratings and uses these thresholds to generate real-time monitoring strategies, giving the monitoring system flexible adaptive capabilities. Unlike existing monitoring systems with fixed safety thresholds and monitoring strategies, this system can dynamically adjust safety thresholds based on the actual health status rating of the structure. When the structure's health is good, the safety thresholds can be appropriately relaxed, reducing monitoring frequency and unnecessary resource consumption. When the structure's health shows a downward trend or a low rating, the safety threshold range can be promptly narrowed, the monitoring frequency increased, and the monitoring range expanded to ensure closer monitoring of changes in the structure's status. This dynamically adjusted monitoring strategy not only ensures timely detection of potential safety hazards but also rationally conserves monitoring resources when the structure is in good condition, achieving a balance between monitoring efficiency and resource costs.

[0055] The anomaly analysis module analyzes the deviation between real-time monitoring data and the preset standard health status data of the real-time monitoring strategy, generating structural anomaly marker data. This module solves the problem that existing monitoring systems can only indicate anomalies but cannot analyze them. When a deviation occurs between real-time monitoring data and the monitoring strategy based on safety thresholds, the anomaly analysis module can deeply analyze the cause of the deviation. Through comprehensive analysis of multimodal monitoring data, dynamic response models, and health status ratings, it accurately determines the location, type, and severity of the anomaly, and generates anomaly marker data containing this key information. Using the anomaly marker data, staff can quickly understand the specific details of the anomaly, thereby developing targeted repair and maintenance plans, avoiding blind maintenance work, reducing maintenance costs, and enabling timely handling of anomalies to prevent further structural damage, ensuring the safe operation of the structure and extending its service life. Attached Figure Description

[0056] Figure 1 This is a timing diagram of the full-time-domain structural health monitoring system described in this invention;

[0057] Figure 2 This is a comparison chart of data from multiple sensors;

[0058] Figure 3 A flowchart illustrating how the adaptive adjustment module works. Detailed Implementation

[0059] 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.

[0060] Please see Figure 1 This invention provides a full-time-domain structural health monitoring system, the system comprising:

[0061] The multi-sensor data acquisition module acquires multi-modal sensor data from the structure in real time. This module reads the raw data sequence output by the sensors according to a preset sampling period and adds timestamp information. The data is then verified and synchronized to generate a synchronized sensor data stream. The dynamic modeling module receives this data stream and extracts key feature points to form structural response vector nodes. These nodes are then fused with timestamps to construct a dynamic response model. The feature extraction module performs hierarchical encoding on the dynamic response model to obtain a multi-dimensional feature matrix. Principal components are selected based on preset feature importance thresholds to generate compressed feature data. The health status assessment module performs pattern recognition on the compressed feature data to extract structural damage feature vectors. Combined with a structural health database, a safety factor assessment is performed, and a health status is output. The system includes a health status rating module; an adaptive adjustment module establishes a safety threshold corridor based on the health status rating, searches for feasible monitoring strategies and filters a set of candidate strategies that meet the continuity rule, assigns weight coefficients to each candidate strategy to form an adaptive strategy library, and then selects the target strategy for optimization to generate a real-time monitoring strategy; an anomaly analysis module obtains the deviation between the real-time monitoring data and the preset standard health status data of the real-time monitoring strategy, calculates the difference parameters and generates multi-layer semantic anomaly labeling data, and outputs structural anomaly labeling data through anomaly type identification; in addition, the system also includes a communication control module, which monitors the signal quality and transmission load parameters of the primary and backup communication channels, calculates the channel priority score and allocates bandwidth resources accordingly, and realizes channel switching decisions to ensure that structural anomaly labeling data is reliably transmitted to the monitoring center.

[0062] Example 1: The multi-sensor data acquisition module initiates the data collection process within a fixed sampling period. This period is preset according to the characteristics of the structure and monitoring requirements, and is usually triggered periodically in milliseconds or seconds. The module controls various sensors distributed in key parts of the structure, including accelerometers, strain gauges, temperature sensors, and displacement sensors. These sensors synchronously collect multimodal physical quantities such as vibration, deformation, and temperature changes. Each sensor generates corresponding timestamp information when collecting data. The timestamps are uniformly managed by a high-precision clock to ensure that all data points have a unified time reference. The collected raw data sequence first enters the data verification stage. The verification process includes logical rationality checks and physical range verification, such as whether the strain data is within the elastic range of the material and whether the temperature data conforms to the local environmental change pattern. For outliers that exceed the reasonable range, median filtering or threshold rejection methods are used for preprocessing, and then the data synchronization stage begins.

[0063] The data synchronization phase employs a clock synchronization mechanism based on a network time protocol. This mechanism sends synchronization signals to each sensor node through the main controller, correcting time deviations caused by transmission delays or node clock drift. The aligned multi-sensor data is integrated according to the time sequence to form a synchronized sensor data stream with a unified time axis. This data stream not only contains the measurement values ​​of each sensor but also retains complete timestamp information and sensor identification, providing a time-series correlation basis for subsequent processing. The synchronized sensor data stream is transmitted to the dynamic modeling module for processing. This module first extracts key feature points. The feature point extraction algorithm adopts different strategies according to different sensor types. For vibration acceleration data, a peak detection algorithm is used to identify local extreme points of the vibration response. For strain data, a sliding window variance calculation is used to capture abrupt changes in strain. For temperature data, gradient calculation is used to identify inflection points of temperature changes. Each extracted feature point is transformed into a structural response vector node. The node data includes the feature point's numerical value, occurrence time, sensor location information, and feature type identifier. These nodes are arranged in chronological order to form a preliminary feature sequence.

[0064] The structural response vector nodes are then deeply fused with timestamp information. This fusion process goes beyond simply appending timestamps to node data; it incorporates time information as a crucial dimension of the nodes in model construction. Each node's timestamp is converted into both relative time offset and absolute time value. The relative time offset describes the temporal relationship between feature points, while the absolute time value records the specific moment the feature occurred. This fusion method allows the dynamic response model to simultaneously reflect both the temporal and spatial distribution characteristics of the structural response. The fusion of structural response vector nodes and timestamp information employs a spatiotemporal anchoring mechanism: each structural response vector node contains sensor location codes (e.g., "main beam-mid-span-number 3" in bridge monitoring), physical quantity values, and feature type identifiers (e.g., "vibration peak value" or "strain mutation point"). During fusion, the timestamps are first converted into a dual-dimensional format of absolute time and relative time offset (absolute time accurate to milliseconds; relative time offset starting from the first data acquisition of the day, in seconds). Then, a hash mapping is used to create a unique association index by combining the time dimension data with the "location code-feature type" of the structural response vector node.

[0065] The final dynamic response model is expressed in the form of a multi-dimensional vector, with each time point corresponding to a feature vector containing feature point information from all sensors at that moment. For time locations without feature points, linear interpolation or state-preserving methods are used to fill in the data, ensuring the continuity of the model. The entire model is stored and managed in the form of a time-series database, supporting fast retrieval and access by time range or by sensor type. Data redundancy and noise handling were also considered during model construction. For adjacent time points with similar features, feature clustering was used for appropriate merging to reduce data storage. Simultaneously, the Kalman filter algorithm was used to smooth the feature point data, eliminating interference from random noise and ensuring the reliability and accuracy of the model data. The final dynamic response model comprehensively reflects the dynamic behavior characteristics of the structure under load and environmental influences, providing detailed data support for the assessment of structural health.

[0066] See Figure 2 The core technologies of the full-time-domain structural health monitoring system are presented. Figures (a) and (b) show the spatiotemporal distribution characteristics of the multi-sensor system during the period from 08-18 to 08-24. Acceleration monitoring data (Figure a) shows that the vibration response amplitudes at locations 1 and 2 remain in the range of 0.00-0.25 m / s². The periodic compression phenomenon at location 1 around 08-21, which shortened from 4 hours to 2.5 hours, may indicate a decrease in the bolt preload at the structural connection nodes. This micro-loosening will change the structural stiffness distribution and thus affect the vibration energy transfer path. Strain monitoring data (Figure b) reveals a systematic shift of 60-80 με between the two measuring points. This continuous difference reflects the characteristics of stress redistribution within the structure. The 40 με baseline drift at location 1 during the period from 08-20 to 08-21 may be related to material creep or the development of local microcracks. The feature extraction module effectively captures this abrupt change feature through sliding window variance calculation.

[0067] Figures (c) and (d) detail the environmental parameters and displacement response. Temperature monitoring data (Figure c) shows that the curves at the two locations highly overlap within the normal range of 20-30°C, and their diurnal fluctuations are consistent with the solar radiation pattern, proving the effectiveness of the temperature compensation algorithm. Displacement monitoring data (Figure d) records that position 1 is consistently higher than position 2 by approximately 0.05 mm, and later diverges to a difference of 0.12 mm. This separation of time-varying displacement trajectories may indicate wear accumulation at the structural hinge points. The health status assessment module generates an anomaly marker with a deviation confidence level of 0.87 through Mahalanobis distance calculation, triggering an adaptive increase in the sampling frequency from 10Hz to 50Hz. The feature extraction module uses a three-layer coding architecture to process temporal feature vectors, constructing a multi-dimensional feature matrix through time-domain statistics, frequency-domain transformation, and spatial fusion. After principal component screening, compressed feature data is formed. Large-scale bridge monitoring cases show that this process, through the combination of kurtosis index and spectral entropy value, can effectively identify typical damage such as cable stress loss and concrete carbonation. Finally, safety factor assessment and status classification are achieved through health database matching, providing decision support for optimizing structural monitoring strategies.

[0068] Example 2: The feature extraction module receives output data from the dynamic modeling module. This data is organized in the form of a time-series feature vector, containing multi-dimensional information such as the vibration response, strain distribution, and temperature field of the structure at different times. The module first starts a hierarchical encoding process. The first layer of encoding divides the continuous time-series data into fixed time windows. The data in each window is processed independently, and time-domain statistical features such as mean, variance, peak factor, and waveform indicators are extracted. These features can reflect the amplitude characteristics and fluctuations of the structural response. The second layer of encoding performs a fast Fourier transform on the time-domain feature sequence in each time window to convert the signal to the frequency domain space and extract frequency domain features such as dominant frequency components, frequency band energy distribution, and spectral centroid. The third layer of encoding integrates the results of the first two layers and introduces sensor spatial location information to construct a multi-dimensional feature matrix containing time, frequency, and spatial dimensions.

[0069] The row dimension of the multi-dimensional feature matrix corresponds to the time series, recording the pattern of feature changes over time. The column dimension contains various feature indicators, covering different feature parameters in the time and frequency domains. The depth dimension reflects the spatial distribution relationship of different sensor nodes. This three-dimensional matrix structure completely preserves the key information in the original data. However, the data volume is large and requires further compression processing. The module calls the preset feature importance thresholds to perform principal component screening. These thresholds are obtained by analyzing historical health status data and stored in the system configuration file. The screening process calculates the variance contribution rate of each feature dimension to evaluate the explanatory power of the feature on the overall data changes. At the same time, it calculates the mutual information score between each feature and the structural damage state to measure the correlation between the feature and the damage.

[0070] The feature selection algorithm traverses all feature dimensions in the feature matrix and calculates the comprehensive importance score for each feature. This score is a weighted composite of variance contribution rate and mutual information score. The weight coefficients are adjusted according to the feature type and monitoring target. For feature dimensions with scores exceeding a preset threshold, the system marks them as principal component features and retains them in a new feature set. Features with scores below the threshold are considered redundant information and are removed. After selection, compressed feature data is generated. This data retains the key information that best reflects changes in structural state from the original data. The amount of data is significantly reduced, but the information density is increased.

[0071] After compressed feature data is input into the health status assessment module, it first undergoes pattern recognition processing. An unsupervised learning algorithm is used to cluster the feature data, identifying different response pattern categories. Simultaneously, a supervised classification algorithm is used to match and identify known damage patterns. This process extracts feature combinations directly related to structural damage, forming a structural damage feature vector. This vector contains the coordinates of the possible location of the damage, a quantitative indicator of the damage severity, and a predicted value for the damage development trend. The system accesses a pre-set structural health database, which stores a large amount of historical cases and experimental data, including safety factor comparison tables for various damage types. The assessment algorithm matches and compares the current structural damage feature vector with the records in the database, calculates a similarity score, and finds the closest reference case. Based on the matching result, the system determines the current structural safety factor value. This factor comprehensively reflects the structure's load-bearing capacity margin and risk level. Finally, based on the safety factor's value range and according to predefined health status level standards, a corresponding health status rating output is generated. This rating result serves as the input parameter for the subsequent adaptive adjustment module, guiding the system to adjust its monitoring strategy.

[0072] Taking the structural health monitoring of a large bridge as an example, the feature extraction module receives bridge vibration response data from the dynamic modeling module. These data come from a group of acceleration sensors arranged on the main beam, piers and cables, with a sampling frequency of 100Hz. The vibration response of traffic flow was continuously collected for 30 minutes. The dynamic response model contains a sequence of vibration feature points of the bridge under normal traffic load. Each feature point has a timestamp and sensor location information.

[0073] The module initiates hierarchical encoding processing. The first layer of encoding divides the 30-minute data into 180 10-second time windows. For the vibration acceleration data within each window, time-domain statistical characteristics are calculated, including mean, standard deviation, peak value, and kurtosis. The mean reflects the average level of vibration, the standard deviation shows the fluctuation amplitude, the peak value captures extreme vibration values, and the kurtosis characterizes the sharpness of the signal distribution. The second layer of encoding performs Fourier transform on the data in each window, extracts the top 5 main frequency components and their amplitudes, and calculates the frequency band energy ratio and spectral entropy value. The frequency band energy ratio shows the energy distribution of each frequency component, and the spectral entropy value quantifies the complexity of the frequency distribution. The third layer of encoding fuses the time-domain and frequency-domain features and introduces sensor spatial information to form a feature matrix containing three dimensions: time, frequency, and space. The rows of the matrix correspond to the time window sequence, the columns contain 20 feature indicators, and the depth dimension represents the spatial distribution of 12 main sensor nodes. The generated multi-dimensional feature matrix enters the principal component screening stage. The system calls the preset feature importance thresholds, which are trained based on the bridge's health monitoring data over the past year. The screening process calculates the variance contribution rate of each feature dimension to evaluate the feature's explanatory power for overall data changes. At the same time, it calculates the mutual information score between each feature and the bridge's damage state to measure the correlation between the feature and the structural health status. After calculation, 8 of the 20 features have a variance contribution rate exceeding 0.85 and a mutual information score greater than 0.7. These features are retained as principal components, including 3 time-domain features (standard deviation, peak value, and kurtosis) and 5 frequency-domain features (first-order frequency amplitude, third-order frequency amplitude, frequency band energy ratio, and spectral entropy). The remaining 12 features are judged as redundant information and are removed. The final compressed feature data volume is reduced by 60%, but retains the key information that best reflects the changes in the bridge's condition.

[0074] After the compressed feature data is input into the health status assessment module, it first undergoes pattern recognition processing. A clustering algorithm is used to divide the feature data of 180 time windows into three main response mode categories, corresponding to light load, medium load, and heavy load traffic states, respectively. At the same time, a classification algorithm is used to identify abnormal vibration modes. This process extracts feature combinations related to bridge damage to form a structural damage feature vector. The vector contains the coordinates of the possible damage location, the quantitative indicators of the damage degree, and the predicted value of the damage development trend.

[0075] The system accesses a pre-set bridge health database, which stores monitoring data and periodic inspection records for the bridge over the past three years. This database includes safety factor comparison tables for various damage types. The evaluation algorithm matches the current structural damage feature vector with historical records in the database, calculates a similarity score, and identifies the closest reference case. Based on the matching results, it determines the current safety factor of the bridge structure. Finally, according to the safety factor range and predefined health status level standards, it generates a corresponding health status rating output. This rating result serves as input parameters for the subsequent adaptive adjustment module, guiding the system to adjust its monitoring strategy. Throughout the process, the feature extraction module successfully compressed the original vibration data from 5.4 million data points to 3240 feature values, reducing the data volume by 99.94%, while preserving key information completely. Based on the compressed feature data, the health status assessment module accurately identifies the bridge's current operating status and potential risks, providing reliable data support for bridge maintenance decisions.

[0076] Example 3: See Figure 3 The adaptive adjustment module receives a health status rating as input parameter, which reflects the overall safety status of the structure. The module first establishes a safety threshold corridor based on the rating value. This corridor consists of two boundary values: an upper limit and a lower limit. The boundary values ​​are set considering factors such as structure type, service life, and environmental conditions. The upper threshold represents the characteristic value range when the structure is in an ideal safety state, while the lower threshold corresponds to the critical state that requires warning. The area between the two thresholds forms a reasonable range for safe operation. The width of the threshold corridor is not fixed but is flexibly adjusted according to the dynamic changes in the health status rating. When the rating shows that the structure is in good condition, the corridor width is appropriately widened to improve system robustness. When the rating shows that the structure is deteriorating, the corridor width is narrowed to enhance monitoring sensitivity. Based on the established safety threshold corridor, the module begins to search for feasible monitoring strategies. The strategy search space includes multiple adjustable parameters, such as sensor sampling frequency, data acquisition duration, transmission interval, and node sleep cycle. The algorithm enumerates all possible parameter combinations to generate a set of candidate strategies. Each strategy needs to pass the continuity rule test, which requires that the change of strategy parameters in adjacent time periods cannot exceed the set maximum jump limit to avoid drastic fluctuations in the monitoring strategy. Strategies that pass the test are retained to form a preliminary candidate strategy library.

[0077] Candidate strategies need to be further evaluated for their applicability. The module assigns weight coefficients to each strategy. The weight calculation comprehensively considers the strategy's execution effect and risk factors. Risk assessment is based on modal parameter analysis of the current vibration state of the structure. By monitoring the natural frequency, damping ratio, and mode shape data, combined with environmental load information, the risk probability distribution over a future period can be calculated. The risk distribution map displays the probability of anomalies occurring in various parts of the structure in the form of a probability cloud.

[0078] The weighting coefficients are calculated using the following formula:

[0079] in: Indicates the first The weight coefficients of each strategy Representing the The risk probability value of each risk area Representation Strategy Monitoring intensity and risk areas Spatial distance, This is the distance attenuation coefficient. For strategy Resource consumption indicators This is the total number of risk areas. This represents the total number of candidate strategies. This formula reflects the principle that high-risk areas require stronger monitoring, while also taking into account the rationality of resource allocation.

[0080] Before calculating the weighting coefficients, it is necessary to complete the overlay analysis of the risk distribution map and the health status rating. The specific steps are as follows:

[0081] Risk distribution map preprocessing: The risk distribution map generated based on modal parameters (such as the three-dimensional risk probability cloud map of offshore wind turbine tower) is divided into 10×10 grid cells. Each grid cell corresponds to a risk probability value (range [0,1]) and the spatial coordinates of the grid are marked (such as 0-10m in the tower height direction and 0-36° in the circumferential direction).

[0082] Health status rating mapping: The health status rating (e.g., Grade B corresponds to a risk coefficient range of [0.3, 0.6]) is converted into a risk correction coefficient for the grid cell. If the original risk probability of the grid cell is... The health status rating correction factor is (Grade B) The corrected risk probability is... ;

[0083] Overlay analysis: The corrected grid risk probability is correlated with the level range of health status rating (e.g., level B corresponds to a mild risk label) to generate an overlay data table of "spatial coordinates - corrected risk probability - risk level label";

[0084] Determining the base values ​​of weighting coefficients: Based on the overlay data table, assign base weight values ​​to each candidate monitoring strategy (e.g., the 45Hz sampling strategy for basic locations). If the proportion of cells with a correction risk probability ≥ 0.2 in the grid cells covered by the strategy is... Then the basic weight value (The higher the percentage, the greater the basic weight);

[0085] Strategy Adaptability Adjustment: Consider the adaptability of the monitoring parameters of candidate strategies to structural risk characteristics. For example, for high-risk foundational areas, the adaptability coefficient of the "45Hz sampling + full-node monitoring" strategy is adjusted. (The fit coefficient ranges from [0,1], determined based on expert experience and historical data), so the final weighting coefficients are... .

[0086] After calculating the weight coefficients of each strategy, the module establishes an adaptive strategy library, sorting all candidate strategies in descending order of weight. The highest-weighted strategies are selected as target strategies, and their parameters are optimized using an iterative trial-and-error method. The optimization fine-tunes the strategy parameters to better adapt to actual monitoring needs while ensuring resource constraints are met. The resulting real-time monitoring strategies guide the next stage of data collection and analysis, completing the system's adaptive adjustment cycle. This entire adjustment process forms a closed-loop feedback mechanism; the execution effect of the real-time monitoring strategies is fed back into the health status rating, influencing the establishment of the next round of safety threshold corridors and strategy optimization. This dynamic adjustment mechanism enables the system to adapt to changes in structural state and environmental conditions, maintaining optimal monitoring performance.

[0087] In the structural health monitoring system of a large offshore wind turbine, the adaptive adjustment module receives the rating result output by the health status assessment module, showing that the tower structure is currently in a Class B state (mild fatigue accumulation). Based on this rating, the module establishes a safety threshold corridor, with the upper threshold set at 1.2 times the normal vibration amplitude and the lower threshold set at 0.8 times the normal value, forming a safe range for allowable fluctuations. Considering the complexity of the marine environment, the width of the threshold corridor is dynamically adjusted according to real-time wind and wave conditions, appropriately widening to 1.5 times the normal value when the wind and waves are calm, and narrowing to 1.1 times the normal value during stormy weather. The module begins to search for feasible monitoring strategies. The strategy parameter space includes dimensions such as sampling frequency (selectable 10Hz, 25Hz, 50Hz), data transmission interval (1 minute, 5 minutes, 15 minutes), and sensor activation combination (full node, critical node, minimum node set). The algorithm generates 27 possible parameter combinations. After a continuity rule check, 8 strategies with excessively large jumps are eliminated, leaving 19 strategies in the candidate set. The continuity rule requires that the sampling frequency change between adjacent time periods does not exceed 20Hz and the data transmission interval change does not exceed 10 minutes.

[0088] The weighting coefficient calculation requires assessing the risk adaptability of each strategy. The system acquires the modal parameters of the current tower vibration state, including the first natural frequency of 0.32Hz, damping ratio of 0.8%, and the first three vibration modes. Combined with real-time environmental load data of wind speed of 12m / s and wave height of 2.5m, the risk probability of the tower foundation connection is calculated to be 0.25, the risk probability of the tower top nacelle is 0.18, and the risk probability of the middle section is 0.12, forming a three-dimensional risk distribution map. The weighting process comprehensively considers the risk distribution and strategy characteristics. For high-risk areas (probability > 0.2), the strategy of high sampling frequency (50Hz) and full node monitoring is preferred; for medium-risk areas (0.1 < probability ≤ 0.2), the strategy of medium sampling frequency (25Hz) and key node combination is matched; for low-risk areas (probability ≤ 0.1), the basic monitoring configuration (10Hz, minimum node set) is used. Calculations showed that the fine-grained monitoring strategy for the high-risk area of ​​the tower foundation received the highest weight of 0.28, the tower top nacelle monitoring strategy received 0.22, the intermediate section monitoring strategy received 0.15, and the remaining strategies ranged from 0.05 to 0.12. After establishing an adaptive strategy library, the module selected the three strategies with the highest weights as target strategies for optimization. For the tower foundation monitoring strategy, the sampling frequency was optimized from 50Hz to 45Hz to reduce resource consumption, while adding vibration direction-specific detection. For the tower top nacelle strategy, the data transmission interval was adjusted from 5 minutes to 7 minutes to balance real-time performance and energy consumption. The intermediate section strategy maintained its basic configuration but added a temperature compensation algorithm. The final real-time monitoring strategy formed a combined scheme: the foundation area used a 45Hz sampling frequency + full-node monitoring + 1-minute transmission interval; the nacelle area used 25Hz sampling + key nodes + 7-minute transmission; and the intermediate section used 10Hz sampling + minimum node set + 15-minute transmission.

[0089] This monitoring strategy was implemented during the subsequent 24 hours of wind turbine operation. The system dynamically adjusted based on actual monitoring data and the strategy's effectiveness. When the wind speed increased to 18 m / s, the risk probability distribution changed, and the system automatically increased the sampling frequency of the foundation components to 50 Hz and shortened the transmission interval to 30 seconds. When wind and wave conditions improved, the system gradually reverted to the original strategy configuration. This adaptive adjustment mechanism ensured optimal allocation of monitoring resources and effectively reduced system energy consumption while guaranteeing structural safety.

[0090] Example 4: The anomaly analysis module receives deviation information between real-time monitoring data and the preset standard health status data of the real-time monitoring strategy. This deviation is calculated by comparing the difference between the actual collected data and the expected value of the strategy. The difference parameters include multiple dimensions such as root mean square error, correlation coefficient, and offset. The root mean square error reflects the overall degree of deviation in data fluctuation, the correlation coefficient measures the consistency of data trends, and the offset represents the systematic deviation at the numerical level. These parameters describe the deviation between the monitoring data and the expected behavior from different perspectives, providing multi-dimensional data support for anomaly identification. Refer to Table 1, which shows a set of typical difference parameter calculation results.

[0091] Table 1: Parameters showing differences in monitoring data

[0092] Monitoring point number Root mean square error Correlation coefficient Offset (μm) Timestamp S01 0.0245 0.87 12.3 2023-08-1514:30:22 S02 0.0312 0.92 8.7 2023-08-1514:30:23 S03 0.0568 0.75 23.1 2023-08-1514:30:24 S04 0.0189 0.94 5.4 2023-08-1514:30:25 S05 0.0723 0.68 31.6 2023-08-1514:30:26

[0093] The differential parameters are then incorporated into the joint analysis process. First, each parameter is standardized to eliminate the influence of dimensions. The root mean square error and offset are normalized using the maximum and minimum values, while the correlation coefficient is directly used with its original value. Subsequently, weight coefficients are assigned according to the importance of the parameters. The root mean square error is weighted at 0.4, the correlation coefficient at 0.3, and the offset at 0.3. The weighted sum is then used to obtain the overall confidence score, which ranges from 0 to 1. A higher score indicates a greater probability of an anomaly.

[0094] The confidence score is compared with a preset anomaly confidence threshold. The threshold is set based on historical anomaly data analysis results and is typically between 0.6 and 0.8. It can be adjusted according to structural importance and risk tolerance. When the score exceeds the threshold, a high-level semantic anomaly label is generated. The label contains detailed information such as anomaly level, anomaly type, and anomaly confidence. The anomaly level is divided into three levels: warning, attention, and severity. The anomaly type includes categories such as sensor failure, structural damage, and environmental interference. When the score is below the threshold, a low-level semantic anomaly label is generated, which only records the anomaly prompt information and timestamp for subsequent trend analysis. The multi-level semantic anomaly label data enters the anomaly type identification stage. This stage uses a combination of rule reasoning and pattern matching. Rule reasoning is based on logical judgment conditions formulated by expert experience, such as "if the offset continues to increase and the correlation coefficient decreases, it is determined to be structural damage." Pattern matching uses a classification model trained on historical anomaly data to identify anomaly feature patterns. Finally, structural anomaly label data is output, which clearly indicates key information such as anomaly type, location, severity, and time of occurrence.

[0095] Taking a bridge monitoring system as an example, the system detected a continuous increase in the offset of monitoring point 3 and a gradual decrease in the correlation coefficient. Although the root mean square error was still within a reasonable range, the overall confidence score reached 0.72, exceeding the set threshold of 0.65. The system generated a high-level semantic anomaly marker, marking the anomaly level as "attention," initially identifying it as "structural damage." Further analysis by the anomaly type identification module, combined with the historical data characteristics of this point and the correlation with adjacent monitoring points, ultimately confirmed it as a structural anomaly marker of "microcrack development." This marker data was transmitted to the monitoring center, prompting engineering technicians to conduct on-site verification. The entire anomaly analysis process employs a multi-level verification mechanism. The high-level semantic anomaly marker needs to undergo time-series verification and spatial correlation testing before final confirmation. Time-series verification requires the anomaly to persist for more than a set duration to avoid false alarms caused by transient interference. Spatial correlation testing requires coordinated changes in adjacent monitoring points to ensure the reliability of anomaly identification. This multi-level analysis method effectively distinguishes between genuine structural anomalies and temporary environmental interference, improving the accuracy and practicality of the monitoring system. The output of the anomaly analysis module includes not only structural anomaly marker data, but also detailed analysis process records and supporting data indexes. This information is stored in the system database, providing a complete data chain for subsequent anomaly tracing and analysis. Engineering technicians can query these records to understand the basis and reasoning process for anomaly identification, assisting in decision-making and taking countermeasures.

[0096] Example 5: The communication control module continuously monitors the operating status of the main and backup communication channels. The main channel uses a fiber optic network for transmission, while the backup channel uses a wireless microwave link. The module collects signal quality parameters for both channels in real time, including signal-to-noise ratio (SNR), bit error rate (BER), transmission delay, and signal strength. Simultaneously, it monitors transmission load parameters such as bandwidth utilization, packet queue length, transmission failure rate, and retransmission count. These parameters are updated at a second-level frequency, forming a dynamic performance profile. In signal quality assessment, SNR reflects the channel's anti-interference capability; a higher value indicates a cleaner signal. BER measures the accuracy of data transmission; a lower value indicates better transmission quality. Transmission delay records the time difference between sending and receiving data packets, affecting real-time performance. Signal strength indicates link stability, which is particularly important in the wireless channel. Regarding transmission load parameters, bandwidth utilization displays the current network capacity usage, packet queue length reflects the degree of transmission congestion, and transmission failure rate and retransmission count reflect the link's reliability characteristics. Based on these real-time parameters, the module calculates a priority score for each channel. The primary channel score calculation assigns higher weights to signal quality parameters, particularly bit error rate and transmission delay, because fiber optic channels typically exhibit more stable transmission characteristics. Backup channel score calculations focus more on indicators reflecting wireless link stability, such as signal strength and transmission failure rate. The score calculation formula employs a weighted summation model, with weighting coefficients dynamically adjusted according to channel type and transmission requirements to ensure that the evaluation results accurately reflect the actual transmission capabilities of the channels.

[0097] Based on the priority scores, the module formulates a bandwidth resource allocation scheme, granting higher-scoring channels a larger proportion of bandwidth resources. Simultaneously, it generates channel switching decisions, including specific switching trigger conditions. For example, switching is triggered when the primary channel's score is lower than the backup channel's score for three consecutive periods, and the difference exceeds a set threshold. Detailed parameters such as switching time window selection, data migration order, and connection maintenance strategies are also included. During actual execution, the system detected a sudden increase in the bit error rate of the primary fiber channel, with transmission latency increasing from the normal 5 milliseconds to 50 milliseconds, while bandwidth utilization reached 85%. Meanwhile, the backup wireless channel maintained a stable signal-to-noise ratio and a low transmission failure rate. Priority score calculations showed that the backup channel's score surpassed the primary channel's, initiating a switching preparation process. First, the data transmission volume of the primary channel was gradually reduced, while the actual transmission capacity of the backup channel was tested. Once usability was confirmed, data migration began. The switching process employed a smooth transition method. First, historical data transmission tasks with lower real-time requirements were switched to the backup channel, while real-time monitoring data transmission continued on the primary channel. Subsequently, the real-time data stream was gradually transferred, ensuring packet sequence continuity and transmission integrity to avoid data loss or duplication. The entire switching process is completed within 300 milliseconds, with transmission interruption time controlled within 50 milliseconds, meeting the continuity requirements of the real-time monitoring system. After the switch is completed, the system continues to monitor the status changes of the two channels. When the signal quality of the main channel stabilizes and the score surpasses that of the backup channel again, the system initiates a back-switching process, gradually migrating data transmission back to the main channel. This dynamic channel management and bandwidth allocation mechanism ensures that structural anomaly marker data can always be transmitted to the monitoring center through the optimal path, guaranteeing the reliability and real-time nature of the monitoring data.

[0098] The data received by the monitoring center includes complete timestamps and channel tagging information, facilitating subsequent analysis of transmission quality trends and optimization of network configuration. The system records all channel switching events and performance parameter change history, providing data support for network maintenance and upgrades. Through this intelligent communication control mechanism, the structural health monitoring system can maintain stable data transmission capabilities in complex field environments, providing reliable communication assurance for structural safety monitoring.

[0099] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A full-time-domain structural health monitoring system, characterized in that, The system includes: A multi-sensor data acquisition module is used to acquire multi-modal sensor data of the structure in real time. A dynamic modeling module is used to generate a dynamic response model of the structure based on the multimodal sensor data; The feature extraction module is used to extract and compress features from the dynamic response model to generate compressed feature data. The health status assessment module is used to perform structural health status parameter analysis based on the compressed feature data and generate a health status rating. An adaptive adjustment module is used to construct a safety threshold based on the health status rating and to generate a real-time monitoring strategy using the safety threshold. The anomaly analysis module is used to analyze the deviation between real-time monitoring data and the preset standard health status data of the real-time monitoring strategy, and generate structural anomaly marker data, including anomaly type, location of occurrence, severity and time of occurrence. The communication control module is used to dynamically adjust the network bandwidth allocation and transmit the structural anomaly marker data to the monitoring center; The communication control module includes: Real-time monitoring of signal quality parameters and transmission load parameters of primary and backup communication channels; The channel priority score is calculated using the signal quality parameters and the transmission load parameters. Bandwidth resources are allocated based on the channel priority score, and channel switching decisions are generated. Based on the channel switching decision, adjust the network bandwidth allocation and transmit the structural anomaly marker data to the monitoring center; The adaptive adjustment module includes: A safety threshold corridor is established based on the health status rating. The safety threshold corridor consists of two boundary values, an upper limit and a lower limit, which is a reasonable range for safe operation formed by the area between the two thresholds. The safety threshold corridor is used to search for feasible monitoring strategies and screen strategies that meet the preset continuity rules to obtain a set of candidate strategies. The monitoring strategy dynamically adjusts the safety threshold according to the actual health status rating of the structure. When the health status of the structure is good, the safety threshold can be appropriately relaxed and the monitoring frequency can be reduced. When the health status of the structure shows a downward trend or the rating is low, the safety threshold range is narrowed in time, the monitoring frequency is increased, and the monitoring range is expanded. Set weight coefficients for each candidate strategy in the candidate strategy set to generate an adaptive strategy library; Target strategies are selected from the adaptive strategy library according to the weight coefficients, and the target strategies are optimized to generate real-time monitoring strategies. The anomaly parsing module includes: Obtain the difference parameters between preset standard health status data and real-time monitoring data; Using the difference parameters, multi-layer semantic anomaly labeling data is generated, including high-level semantic anomaly labels and low-level semantic anomaly labels. The high-level semantic anomaly labels include anomaly level, anomaly type, and anomaly confidence, while the low-level semantic anomaly labels only record anomaly prompt information and timestamps. The multi-layer semantic anomaly labeling data is used to identify anomaly types and generate structural anomaly labeling data. The process of generating multi-layer semantic anomaly marker data using the difference parameters includes: A joint analysis of the aforementioned difference parameters is performed to generate a confidence score; The confidence score is compared with a preset anomaly confidence threshold; based on the judgment result, multi-layer semantic anomaly labeling data is generated.

2. The full-time-domain structural health monitoring system according to claim 1, characterized in that, The dynamic modeling module includes: Key feature points are extracted from the synchronous sensor data stream to generate structural response vector nodes; The structural response vector nodes are fused with timestamp information to form a dynamic response model.

3. The full-time-domain structural health monitoring system according to claim 1, characterized in that, The feature extraction module includes: The dynamic response model is hierarchically encoded to generate a multi-dimensional feature matrix; Principal components in the multi-dimensional feature matrix are filtered by a preset feature importance threshold to generate compressed feature data.

4. The full-time-domain structural health monitoring system according to claim 1, characterized in that, The health status assessment module includes: Pattern recognition is performed on the compressed feature data to generate a structural damage feature vector; The safety coefficient of the structural damage feature vector is evaluated by combining the preset structural health database, and a health status rating is generated.

5. The full-time-domain structural health monitoring system according to claim 1, characterized in that, The system also includes: Obtain the modal parameters of the structure under its current vibration state, calculate the probability of future risks using the modal parameters, and generate a risk distribution map; The risk distribution map is overlaid with the health status rating to generate weighting coefficients.

6. The full-time-domain structural health monitoring system according to claim 1, characterized in that, The multi-sensor data acquisition module includes: Acquire sensor data sequences and timestamp information under a preset sampling period; The sensor data sequence is validated, and the multi-sensor data is aligned through a data synchronization mechanism to generate a synchronized sensor data stream.

7. A method for monitoring structural health across the entire time domain, characterized in that, It includes all modules and method flows of a full-time-domain structural health monitoring system as described in any one of claims 1 to 6.

Images