A bridge structure health monitoring method and system based on multi-source sensor fusion

Abstract

This application provides a bridge structural health monitoring method and system based on multi-source sensor fusion, belonging to the field of health monitoring technology. The method involves acquiring multi-source sensor raw data of key bridge components and generating standardized data sequences; performing fusion analysis on the standardized data sequences to extract cross-modal correlation features; constructing a multi-parameter coupled state vector based on the cross-modal correlation features and comparing it with a health state benchmark domain to generate a structural deviation characterization; performing damage sensitivity analysis on the structural deviation characterization and outputting a structural anomaly confidence distribution; performing temporal accumulation and spatial extrapolation on the structural anomaly confidence distribution to generate a structural health state evolution map; and conducting health monitoring based on the structural health state evolution map. This application, through multi-source sensor fusion and cross-modal correlation feature extraction, can achieve accurate localization and type identification of bridge structural damage, improving the anti-interference capability of health monitoring against environmental disturbances.

Description

Technical Field

[0001] This application relates to the field of health monitoring technology, and more specifically, to a method and system for monitoring the health of bridge structures based on multi-source sensor fusion. Background Technology

[0002] In existing bridge structural health monitoring technologies, commonly used methods mainly include static weighing systems based on strain sensors, modal analysis techniques based on accelerometers, and deflection monitoring methods based on displacement gauges. These technologies typically use a single type of sensor to independently collect and analyze specific physical quantities of the bridge, such as measuring local stress in the structure using strain gauges, identifying the natural frequencies and mode shapes of the structure using accelerometers, and recording vertical deformation using total stations or displacement gauges. In recent years, some monitoring systems have attempted to combine multiple sensors and use linear fusion methods such as weighted averaging or principal component analysis to comprehensively process multi-source data. In terms of data transmission, existing technologies generally employ timed trigger sampling and unified resampling alignment strategies. At the damage identification level, mainstream methods rely on finite element model correction or anomaly detection algorithms based on statistical process control, judging the structural health status by comparing the residuals between the measured response and the theoretical calculation values.

[0003] In existing bridge structural health monitoring technologies, multi-source sensor data are typically analyzed independently or only superficially linearly fused. This fails to effectively uncover the nonlinear causal relationships and dynamic response sequence changes among different physical quantities caused by structural damage. Consequently, the sensitivity for identifying early local damage is insufficient, making it difficult to accurately distinguish between environmental disturbances and actual structural anomalies. Therefore, how to achieve precise localization and type identification of bridge structural damage through multi-source sensor fusion and cross-modal correlation feature extraction, and improve the resilience of health monitoring to environmental disturbances, is a challenge facing the industry. Summary of the Invention

[0004] This application provides a bridge structural health monitoring method and system based on multi-source sensor fusion. Through multi-source sensor fusion and cross-modal correlation feature extraction, it can achieve accurate location and type identification of bridge structural damage, and improve the anti-interference capability of health monitoring against environmental disturbances.

[0005] In a first aspect, this application provides a bridge structural health monitoring method based on multi-source sensor fusion, the health monitoring method comprising the following steps:

[0006] Acquire multi-source sensor raw data of key parts of the bridge, perform time alignment and anomaly removal on the multi-source sensor raw data, and generate a standardized data sequence;

[0007] The standardized data sequence is fused and analyzed to extract cross-modal correlation features between different physical quantities;

[0008] Based on the cross-modal correlation features, a multi-parameter coupled state vector is constructed and compared with the health state benchmark domain to generate a structural deviation characterization. Damage sensitivity analysis is performed on the structural deviation characterization, and the structural anomaly confidence distribution at each monitoring location is output.

[0009] The structural anomaly confidence distribution is accumulated over time and extrapolated spatially to generate a structural health status evolution map, and then health monitoring is carried out based on the structural health status evolution map.

[0010] In this embodiment, the multi-source sensing raw data includes strain data, acceleration data, temperature data, and deflection data.

[0011] In this embodiment, the fusion analysis of the standardized data sequence to extract cross-modal correlation features between different physical quantities specifically includes:

[0012] The standardized data sequence is time-scale aligned to establish a synchronous mapping relationship between multiple data sources;

[0013] Based on the propagation hysteresis characteristics of different physical quantities in the structural response, a time delay correlation matrix across physical quantities is constructed;

[0014] Principal component analysis was performed on the time delay correlation matrix to determine the initial cross-modal correlation characteristics;

[0015] By introducing a transfer entropy operator, the initial cross-modal correlation features are nonlinearly enhanced to determine the information flow intensity between each pair of physical quantities, thereby obtaining the cross-modal correlation features between different physical quantities.

[0016] In this embodiment, constructing the time delay correlation matrix across physical quantities specifically includes:

[0017] The response propagation time lag between each pair of sensing sequences is estimated using a cross-correlation function, and then a time delay matrix is ​​constructed.

[0018] Construct a confidence matrix based on the magnitude of the cross-correlation peaks;

[0019] The time delay correlation matrix is ​​determined based on the time delay matrix and the confidence matrix.

[0020] In this embodiment, the construction of a multi-parameter coupled state vector based on the cross-modal correlation features and its comparison with the health state baseline domain to generate a structural deviation characterization specifically includes:

[0021] The cross-modal correlation features are flattened and normalized within a time window, and then spliced ​​together to form a multi-parameter coupled state vector.

[0022] During the calibration window period in the early stage of bridge operation, a set of multi-parameter coupled state vectors under no abnormal conditions is collected, the stable attractor manifold of the set of multi-parameter coupled state vectors in the state space is learned, and the stable attractor manifold is used as the reference domain for the healthy state.

[0023] The multi-parameter coupled state vector at the current moment is projected onto the stable attractor manifold to determine the projection residual, which is then used as a characterization of structural deviation.

[0024] In this embodiment, the learning method for the stable attractor manifold specifically includes:

[0025] A local linear embedding algorithm is used to perform dimensionality reduction manifold learning on the multi-parameter coupled state vector within the calibration window to obtain low-dimensional embedding coordinates.

[0026] The convex hull surface of all calibrated sample points is determined in the low-dimensional embedding space, and the convex hull surface is extended outward by an adaptive tolerance distance to determine the boundary of the stable attractor manifold.

[0027] In this embodiment, the damage sensitivity analysis of the structural deviation characterization and the output of the structural anomaly confidence distribution at each monitoring location specifically include:

[0028] The structural deviation characterization is organized according to the spatial location of the sensor to form a spatial deviation distribution field;

[0029] Establish a set of parallel response difference enhancement operators, each corresponding to a preset damage type;

[0030] The spatial deviation distribution field is input into each response difference enhancement operator, and each response difference enhancement operator outputs the probability of occurrence of the corresponding damage type at each monitoring location;

[0031] The probability of occurrence of the output of each response difference enhancement operator is competitively normalized, thereby outputting the structural anomaly confidence distribution at each monitoring location.

[0032] In this embodiment, the temporal accumulation and spatial extrapolation of the structural anomaly confidence distribution to generate a structural health state evolution diagram specifically includes:

[0033] Based on the structural anomaly confidence distribution at each monitoring location, the time-series evolution curve of structural anomaly confidence at each monitoring location is determined;

[0034] The time-series evolution curve of the structural anomaly confidence is smoothed, and the smoothed time-series evolution curve of the structural anomaly confidence is extrapolated to the continuous space of the full bridge to generate a spatial anomaly confidence distribution cloud map under each time window.

[0035] The spatial anomaly confidence distribution cloud maps of multiple consecutive time windows are stacked in chronological order to form a structural health state evolution map.

[0036] In this embodiment, the structural health status evolution diagram refers to a three-dimensional data structure that stores the change of anomaly confidence at each monitoring location over time, with spatial coordinates as the planar dimension and time as the vertical axis.

[0037] Secondly, this application provides a bridge structural health monitoring system based on multi-source sensor fusion, used to execute a bridge structural health monitoring method based on multi-source sensor fusion, the health monitoring system comprising:

[0038] The data acquisition module is used to acquire multi-source sensor raw data of key parts of the bridge, perform time alignment and anomaly removal on the multi-source sensor raw data, and generate a standardized data sequence.

[0039] The feature extraction module is used to perform fusion analysis on the standardized data sequence, and then extract cross-modal correlation features between different physical quantities;

[0040] The confidence output module is used to construct a multi-parameter coupled state vector based on the cross-modal correlation features, compare it with the health state reference domain, generate a structural deviation characterization, perform damage sensitivity analysis on the structural deviation characterization, and output the structural anomaly confidence distribution at each monitoring location.

[0041] The health monitoring module is used to perform temporal accumulation and spatial extrapolation of the structural anomaly confidence distribution, generate a structural health state evolution map, and then perform health monitoring based on the structural health state evolution map.

[0042] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0043] The process involves acquiring raw multi-source sensor data from key bridge components, performing time alignment and anomaly removal on the raw data to generate a standardized data sequence, and then performing fusion analysis on the standardized data sequence to extract cross-modal correlation features between different physical quantities. Based on these cross-modal correlation features, a multi-parameter coupled state vector is constructed and compared with a health state baseline domain to generate a structural deviation characterization. Damage sensitivity analysis is performed on the structural deviation characterization to output the structural anomaly confidence distribution at each monitoring location. The structural anomaly confidence distribution is then subjected to temporal accumulation and spatial extrapolation to generate a structural health state evolution map, which is subsequently used for health monitoring.

[0044] Therefore, this application firstly generates a standardized data sequence with synchronized time and unified dimensions by performing time alignment and anomaly removal on the raw data from multi-source sensing, providing a data foundation for the accurate extraction of subsequent cross-modal correlation features; secondly, by fusing and analyzing the standardized data sequence and extracting cross-modal correlation features, the extracted correlation features contain the physical essence of the structural response and are highly sensitive to cross-modal coupling anomalies caused by damage, providing robust feature inputs for subsequent structural state characterization and damage identification; and thirdly, by constructing a multi-parameter coupled state vector and aligning it with a stable attractor manifold reference... By comparing domains, high-dimensional cross-modal correlation features are compressed into local deviation fields, enabling a quantitative description of the overall structural health status. This allows structural deviations to be mapped to the probability of different damage types occurring at each monitoring location, achieving damage location and type identification, and providing an intuitive decision-making basis for subsequent health monitoring and early warning. Finally, by accumulating the confidence distribution of structural anomalies over time and extrapolating it spatially, the anomaly information of discrete monitoring points is expanded into a dynamic evolution map of the entire bridge in continuous space, enabling accurate location and type identification of bridge structural damage and improving the anti-interference capability of health monitoring against environmental disturbances.

[0045] In summary, the technical solution adopted in this application can achieve accurate location and type identification of bridge structural damage through multi-source sensor fusion and cross-modal correlation feature extraction, thereby improving the anti-interference capability of health monitoring against environmental disturbances. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is an exemplary flowchart of a bridge structural health monitoring method based on multi-source sensor fusion provided in this application;

[0048] Figure 2 This is a schematic diagram of a device scenario for a bridge structural health monitoring system based on multi-source sensor fusion, as provided in this application.

[0049] Figure 3 This is a module structure diagram of a bridge structural health monitoring system based on multi-source sensor fusion provided in this application. Detailed Implementation

[0050] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0051] This application provides a method and system for bridge structural health monitoring based on multi-source sensor fusion. The core of this method involves acquiring raw multi-source sensor data from key bridge components, performing time alignment and anomaly removal on the raw data to generate a standardized data sequence, performing fusion analysis on the standardized data sequence to extract cross-modal correlation features between different physical quantities, constructing a multi-parameter coupled state vector based on the cross-modal correlation features, comparing it with a health state benchmark domain to generate a structural deviation characterization, performing damage sensitivity analysis on the structural deviation characterization, and outputting the structural anomaly confidence distribution at each monitoring location, performing temporal accumulation and spatial extrapolation on the structural anomaly confidence distribution to generate a structural health state evolution map, and then performing health monitoring based on the structural health state evolution map.

[0052] Example 1: To better understand the above technical solution, the following will provide a detailed description of the technical solution in conjunction with the accompanying drawings and specific implementation methods. (Refer to...) Figure 1 As shown in the figure, this is an exemplary flowchart of a bridge structure health monitoring method based on multi-source sensor fusion according to this embodiment of the present application. The health monitoring method includes the following steps:

[0053] In step S1, multi-source sensor raw data of key parts of the bridge are acquired, and time alignment and anomaly removal are performed on the multi-source sensor raw data to generate a standardized data sequence.

[0054] In practical implementation, multi-source sensor raw data is acquired from key parts of the bridge. This involves deploying a multi-source sensor system at these critical locations, including: the mid-span of the main span, quarter points, supports, the top and bottom of the towers, and anchorage zones of stay cables or suspenders—areas sensitive to structural stress. The sensor system consists of four types of sensors: fiber optic strain sensors, MEMS capacitive accelerometers, platinum resistance temperature sensors, and wire-type displacement sensors or BeiDou displacement monitoring terminals. The strain sensors measure the microscopic deformation of the structure under external loads; the accelerometers collect the vibration response of the structure; the temperature sensors collect ambient temperature field data to help eliminate temperature interference; and the displacement sensors monitor the vertical deflection and support displacement of the bridge. All sensors share the same hardware clock source. Each data point is timestamped with high precision upon acquisition. In subsequent processing, instead of resampling at a fixed frequency, the time axis of the accelerometer is used as the reference. For other low-frequency sensors, a zero-order hold that retains the previous effective value is used to fill the gaps between adjacent sampling points, and the filling marker is recorded. The collected strain data, acceleration data, temperature data, and deflection data are used as the raw data for multi-source sensing.

[0055] In addition, in the specific implementation, time alignment and anomaly removal are performed on the raw data from multiple sensors to generate a standardized data sequence. That is, all sensors share the same hardware clock source during acquisition, and each data point carries a high-precision timestamp. The time alignment operation uses the time axis of the accelerometer as the reference time axis. For each timestamp on the acceleration time axis, the data point with the smallest absolute deviation from that timestamp is searched in the strain, temperature, and deflection data sequences as the matching point. When the time deviation between the acceleration timestamp and the nearest neighbor temperature data point exceeds half of the temperature sampling interval, a zero-order hold is used. This means that the previous valid temperature value is retained as an approximation for that moment, and this point is recorded as a zero-order hold fill point to distinguish it from the true sampling point. Anomaly detection based on local outlier factors is performed on the single-channel data of each type of sensor. For each data point, the ratio of its local reachability density to several nearest neighbor data points in its surrounding neighborhood is calculated. When this ratio is lower than a preset threshold, the data point is identified as an outlier. For cases where a single isolated outlier is identified, the median of the two normal data points before and after that point is used for replacement. For cases where multiple consecutive data points are identified as anomalous, such as data gaps caused by a momentary communication interruption, linear interpolation based on adjacent normal data points is used for filling. However, if the length of a consecutive anomalous segment exceeds one second, the data for that period is considered unrecoverable, marked as invalid data, and skipped in subsequent fusion analysis without being filled. For temperature and deflection data, the first-order difference method is used to calculate the rate of change between adjacent sampling points. When the rate of change of a sampling point exceeds five times the standard deviation of the sensor's historical rate of change mean, that point is identified as a jump peak, and the measurement value at the moment before the jump occurs is used for replacement.

[0056] In addition, in the specific implementation, after time alignment and anomaly removal are completed, standardization adopts the commonly used mean-variance normalization method. For each type of sensor, its mean and standard deviation are calculated within the calibration window during the initial operation of the bridge. Then, for each original measurement value, the mean of the physical quantity is subtracted and then divided by the standard deviation to obtain the dimensionless standardized value. The standardized data sequence is then organized into a two-dimensional matrix in chronological order as the standardized data sequence.

[0057] In step S2, the standardized data sequence is fused and analyzed to extract cross-modal correlation features between different physical quantities.

[0058] In this embodiment, the standardized data sequence is fused and analyzed to extract cross-modal correlation features between different physical quantities. This can be achieved through the following steps:

[0059] The standardized data sequence is time-scale aligned to establish a synchronous mapping relationship between multiple data sources;

[0060] Based on the propagation hysteresis characteristics of different physical quantities in the structural response, a time delay correlation matrix across physical quantities is constructed;

[0061] Principal component analysis was performed on the time delay correlation matrix to determine the initial cross-modal correlation characteristics;

[0062] By introducing a transfer entropy operator, the initial cross-modal correlation features are nonlinearly enhanced to determine the information flow intensity between each pair of physical quantities, thereby obtaining the cross-modal correlation features between different physical quantities.

[0063] In practice, firstly, the standardized data sequences are time-scale aligned to establish a synchronous mapping relationship between multi-source data. That is, within each time window, the data vectors of the four physical quantities—strain, acceleration, temperature, and deflection—correspond one-to-one in time, collectively constituting the multi-physical quantity observation sample at that moment. Secondly, based on the transmission hysteresis characteristics of different physical quantities in the structural response, a cross-physical quantity time-delay correlation matrix is ​​constructed. This process will be elaborated in subsequent steps. Then, principal component analysis is performed on the time-delay correlation matrix to determine the initial cross-modal correlation characteristics. Specifically, the time-delay correlation matrix is ​​treated as a high-dimensional vector, and principal component analysis is performed on the time-delay correlation matrix samples collected under multiple time windows, extracting those with a cumulative contribution rate exceeding 85%. The first few principal components are used as initial cross-modal correlation features, which mainly describe linear or weakly nonlinear correlations. Finally, a transfer entropy operator is introduced to nonlinearly enhance the initial cross-modal correlation features, determining the information flow strength between each pair of physical quantities. This yields cross-modal correlation features between different physical quantities. Specifically, for each pair of physical quantities, the response transfer time lag estimated by the time delay correlation matrix is ​​used as a priori offset. Transfer entropy is calculated within the offset time window. When the calculated transfer entropy value exceeds a preset threshold, a genuine correlation driven by structural response is determined between the pair of physical quantities; otherwise, it is determined to be a pseudo-correlation driven by environmental noise, and the transfer entropy value of the pair of physical quantities is set to zero. For physical quantity pairs determined to be genuinely correlated, their transfer entropy value is used as the nonlinearly enhanced correlation strength and fused with the principal component scores in the initial cross-modal correlation features. That is, the transfer entropy value is used as a weight multiplied by the corresponding principal component's feature vector to form a weighted correlation feature vector. The weighted correlation feature vector is continuously output according to the time window, and the resulting time series is used as the cross-modal correlation feature. Cross-modal correlation features imply causal coupling relationships between different physical quantities in space, as well as evolutionary information over time. Among them, the transfer entropy operator is an asymmetric metric function based on information theory, used to quantify the direction and intensity of information flow between two sensing sequences of different physical quantities.

[0064] In this embodiment, the construction of the time delay correlation matrix across physical quantities can be achieved through the following steps:

[0065] The response propagation time lag between each pair of sensing sequences is estimated using a cross-correlation function, and then a time delay matrix is ​​constructed.

[0066] Construct a confidence matrix based on the magnitude of the cross-correlation peaks;

[0067] The time delay correlation matrix is ​​determined based on the time delay matrix and the confidence matrix.

[0068] In practical implementation, firstly, a cross-correlation function can be used to estimate the response propagation time lag between each pair of sensing sequences, and then a time delay matrix can be constructed. That is, for any two standardized sensing sequences of different physical quantities, denoted as A and B, sequence B is translated relative to sequence A in time steps, and the inner product of the two sequences is calculated at each translation amount. When the two sequences reach the highest similarity at a certain translation amount, that translation amount is the response propagation time lag. The response propagation time lag refers to the time lag after which physical quantity A changes after the structure is excited, and physical quantity B shows a corresponding change. The above cross-correlation analysis is performed on all sensor channels in pairs to obtain the response propagation time lag between each pair of sensors. All response transmission time lags are filled into a square matrix according to the sensor arrangement. Rows represent source physical quantities, columns represent target physical quantities, and diagonal elements are zero, meaning the lag of the same sensor relative to itself is zero. This forms the time delay matrix. Next, a confidence matrix is ​​constructed using the amplitude of the cross-correlation peaks. The amplitude of the cross-correlation function at its peak represents the similarity between two sequences in their optimal alignment. All sensor channels are paired, and the amplitude values ​​of the cross-correlation peaks are recorded. These values ​​are then filled into another square matrix in the same row and column order as the time delay matrix, forming the confidence matrix. Each element in the confidence matrix ranges from -1 to 1, and the absolute value is used as the confidence level. A larger value indicates a more reliable estimate of the lag between the sensor pairs. Finally, the time delay correlation matrix is ​​determined based on the time delay matrix and the confidence matrix. This is achieved by multiplying the corresponding positions of the time delay matrix and the confidence matrix. For each sensor pair, its lag is multiplied by the corresponding confidence level, thus obtaining the time delay correlation matrix. The time delay correlation matrix contains information about the timing of response transmission and the coupling strength. The signs of the values ​​in the time delay correlation matrix indicate the order of responses: a positive sign indicates that the source physical quantity is ahead of the target physical quantity, and a negative sign indicates that the source physical quantity is behind the target physical quantity. The magnitude of the absolute value reflects the length of the lag time and the strength of the correlation.

[0069] In step S3, a multi-parameter coupled state vector is constructed based on the cross-modal correlation features and compared with the health status benchmark domain to generate a structural deviation characterization. Damage sensitivity analysis is performed on the structural deviation characterization, and the structural anomaly confidence distribution at each monitoring location is output.

[0070] In this embodiment, a multi-parameter coupled state vector is constructed based on the cross-modal correlation features and compared with the health state baseline domain to generate a structural deviation characterization. This can be achieved through the following steps:

[0071] The cross-modal correlation features are flattened and normalized within a time window, and then spliced ​​together to form a multi-parameter coupled state vector.

[0072] During the calibration window period in the early stage of bridge operation, a set of multi-parameter coupled state vectors under no abnormal conditions is collected, the stable attractor manifold of the set of multi-parameter coupled state vectors in the state space is learned, and the stable attractor manifold is used as the reference domain for the healthy state.

[0073] The multi-parameter coupled state vector at the current moment is projected onto the stable attractor manifold to determine the projection residual, which is then used as a characterization of structural deviation.

[0074] In practical implementation, firstly, cross-modal associated features can be flattened and normalized within a time window, and then concatenated to form a multi-parameter coupled state vector. That is, all associated feature values ​​within each time window are flattened, essentially flattening the two-dimensional feature tensor into a one-dimensional vector. This one-dimensional vector is then normalized using a maximum-minimum normalization method, mapping the numerical range of each feature dimension to between zero and one. After flattening and normalization, this one-dimensional vector serves as the multi-parameter coupled state vector for that time window. Then, during the calibration window in the initial stage of bridge operation, a set of multi-parameter coupled state vectors under anomaly-free conditions is collected to learn the multi-parameter... The set of multi-parameter coupled state vectors is placed on a stable attractor manifold in the state space, and this stable attractor manifold is used as the reference domain for the healthy state. This process will be elaborated in subsequent steps. Finally, the multi-parameter coupled state vectors at the current time are projected onto the stable attractor manifold to determine the projection residuals. These projection residuals are used as a representation of structural bias. Specifically, for each multi-parameter coupled state vector, a local linear embedding transformation identical to that used for the calibration window is applied to map it to a low-dimensional embedding space, obtaining its coordinates in the low-dimensional space. Several healthy sample points closest to these coordinates are found in this low-dimensional space, and the current coordinates are locally reconstructed using a linear combination of these nearest neighbors. The projection residual is the Euclidean distance between the current coordinates and their locally reconstructed values, and the obtained projection residual is used as a representation of structural bias.

[0075] In this embodiment, the learning method for the stable attractor manifold can be implemented using the following steps:

[0076] A local linear embedding algorithm is used to perform dimensionality reduction manifold learning on the multi-parameter coupled state vector within the calibration window to obtain low-dimensional embedding coordinates.

[0077] The convex hull surface of all calibrated sample points is determined in the low-dimensional embedding space, and the convex hull surface is extended outward by an adaptive tolerance distance to determine the boundary of the stable attractor manifold.

[0078] In practical implementation, firstly, a local linear embedding algorithm can be used to perform dimensionality reduction manifold learning on the multi-parameter coupled state vector within the calibration window to obtain low-dimensional embedding coordinates. That is, for each multi-parameter coupled state vector within the calibration window, the nearest neighboring points are searched in the high-dimensional Euclidean space. Under the constraint of maintaining the linear reconstruction relationship between each sample point and its neighbors, the optimal reconstruction weight coefficients when each sample point is linearly represented by these neighbors are solved. This solution process is a least squares problem with equality constraints, requiring that the sum of the reconstruction weights of each sample point equals one, and that the weights are non-zero only when neighbors are used for reconstruction. The reconstruction weight coefficients of all sample points are combined into a sparse reconstruction weight matrix according to the index relationship between the samples. The reconstructed weight matrix is ​​decomposed into eigenvalues, and the eigenvectors corresponding to the smallest non-zero eigenvalues ​​are taken as the low-dimensional embedding coordinates of all sample points. Then, the convex envelope surface of all calibrated sample points is determined in the low-dimensional embedding space, and the convex envelope surface is extended outward by an adaptive tolerance distance to determine the boundary of the stable attractor manifold. That is, after obtaining the low-dimensional embedding coordinates, all healthy sample points within the calibration window constitute a point cloud set with a certain shape and boundary, and the region occupied by these sample points is the main part of the stable attractor manifold. The convex envelope surface of all sample points is then calculated. The convex envelope surface is the surface of the smallest convex polyhedron containing all sample points, and the region inside the convex envelope surface is the densely distributed area of ​​healthy sample points. For each vertex on the convex envelope surface, the average distance to its nearest neighbor healthy sample points is calculated, and this average distance is multiplied by a preset expansion coefficient to obtain the local tolerance distance at that vertex. The expansion coefficient can be set by expert experience. By moving each vertex of the convex hull surface along its outward normal direction by its corresponding local tolerance distance, and then reconnecting adjacent vertices, the surface of the soft boundary is obtained. The region enclosed by this soft boundary is the boundary of the stable attractor manifold. In the low-dimensional embedding space, multi-parameter coupled state vectors located within this soft boundary are considered to be in a healthy state, while vectors located outside the soft boundary are considered to be in a potentially anomalous state. A stable attractor manifold refers to the low-dimensional geometric structure formed by the multi-parameter coupled state vectors of a bridge in a high-dimensional state space under healthy conditions.

[0079] In this embodiment, damage sensitivity analysis is performed on the structural deviation characterization to output the structural anomaly confidence distribution at each monitoring location. This can be achieved through the following steps:

[0080] The structural deviation characterization is organized according to the spatial location of the sensor to form a spatial deviation distribution field;

[0081] Establish a set of parallel response difference enhancement operators, each corresponding to a preset damage type;

[0082] The spatial deviation distribution field is input into each response difference enhancement operator, and each response difference enhancement operator outputs the probability of occurrence of the corresponding damage type at each monitoring location;

[0083] The probability of occurrence of the output of each response difference enhancement operator is competitively normalized, thereby outputting the structural anomaly confidence distribution at each monitoring location.

[0084] In practical implementation, firstly, the structural deviation representation is organized according to the spatial location of the sensors to form a spatial deviation distribution field. That is, in the low-dimensional embedding space, the magnitude of the components of the structural deviation representation in each coordinate axis direction is recorded. Using the reconstruction weight matrix of the local linear embedding algorithm, the components of the low-dimensional structural deviation representation are mapped back to the high-dimensional sensor space to obtain the local residual contribution value corresponding to each sensor channel. According to the actual physical installation location of the strain sensor, acceleration sensor, temperature sensor, and deflection sensor, these local residual contribution values ​​are assigned to the corresponding spatial coordinate points. For spatial locations where no sensors are installed, their residual contribution values ​​are temporarily empty, thus forming a spatial deviation distribution field. Secondly, a set of parallel response difference enhancement operators is established. Each response difference enhancement operator corresponds to a preset damage type. That is, for each preset damage type, a structural response spatial distribution template corresponding to the damage is pre-established. These structural response spatial distribution templates are derived from historical damage case data and describe the theoretical distribution pattern that the deviation contribution values ​​at each sensor location should present when a certain damage occurs. The function of the response difference enhancement operator is to match the currently measured spatial deviation distribution field with each damage template and calculate the matching degree. Each response difference enhancement operator internally employs a template-based weighted matching algorithm, outputting a matching degree value between zero and one, representing the similarity between the current deviation distribution and the damage type. The response difference enhancement operator is a computational unit used to match the currently measured spatial deviation distribution field with the theoretical response template of a preset damage type, thereby quantifying the probability of the corresponding damage type occurring at each monitoring location. Then, the spatial deviation distribution field is input into each response difference enhancement operator, and each operator outputs the probability of the corresponding damage type occurring at each monitoring location. That is, for each monitoring location, the response difference enhancement operator compares the measured local residual contribution value at the monitoring location with the theoretical contribution value at the corresponding location in the damage template, using cosine similarity to calculate the similarity between the two. After independently calculating the similarity at each monitoring location, the probability of occurrence of that damage type at that location is obtained. Finally, the occurrence probabilities output by each response difference enhancement operator are competitively normalized to output the structural anomaly confidence distribution for each monitoring location. Specifically, for each monitoring location, the occurrence probabilities output by all response difference enhancement operators are collected to form a probability vector. This probability vector is then normalized by dividing each probability value by the sum of all probability values ​​in the vector to obtain the confidence vector. After normalization, the sum of the confidence scores for all damage types at that monitoring location is one, and the damage type corresponding to the highest confidence score is determined as the most likely damage type at that monitoring location. Finally, the confidence vectors for each monitoring location are organized according to spatial coordinates to output the structural anomaly confidence distribution.

[0085] In step S4, the structural anomaly confidence distribution is accumulated over time and extrapolated spatially to generate a structural health state evolution map, and then health monitoring is performed based on the structural health state evolution map.

[0086] In this embodiment, the structural anomaly confidence distribution is accumulated over time and extrapolated spatially to generate a structural health state evolution diagram. This can be achieved through the following steps:

[0087] Based on the structural anomaly confidence distribution at each monitoring location, the time-series evolution curve of structural anomaly confidence at each monitoring location is determined;

[0088] The time-series evolution curve of the structural anomaly confidence is smoothed, and the smoothed time-series evolution curve of the structural anomaly confidence is extrapolated to the continuous space of the full bridge to generate a spatial anomaly confidence distribution cloud map under each time window.

[0089] The spatial anomaly confidence distribution cloud maps of multiple consecutive time windows are stacked in chronological order to form a structural health state evolution map.

[0090] In practical implementation, firstly, based on the structural anomaly confidence distribution at each monitoring location, the time-series evolution curve of the structural anomaly confidence at each monitoring location is determined. That is, the maximum confidence among all damage types at each monitoring location is taken as the comprehensive anomaly confidence at that location within that time window, characterizing the overall probability of any type of damage occurring at that location. Over time, each sensor location continuously generates new comprehensive anomaly confidence values. These comprehensive anomaly confidence scores are arranged chronologically, forming a curve with the time window number on the horizontal axis and the comprehensive anomaly confidence score on the vertical axis. This curve represents the time-series evolution curve of the structural anomaly confidence score at the monitoring location. Each point on the curve represents the degree of structural anomaly at that location at the corresponding time. An upward curve indicates that the anomaly is developing, while a downward curve indicates that the anomaly is receding. Then, the time-series evolution curve of the structural anomaly confidence score is smoothed. The smoothed curve is then extrapolated to the continuous space of the entire bridge, generating a spatial anomaly confidence score distribution cloud map for each time window. Specifically, the smoothing process uses an exponentially weighted moving average method with a forgetting factor. The forgetting factor is between zero and one, and its value is adaptively adjusted according to the bridge's traffic load cycle. It can be set according to expert recommendations: during peak traffic periods, the forgetting factor is set to a smaller value, making the smoothing result respond more quickly to the current observation value; during off-peak traffic periods, the forgetting factor is set to a larger value, making the smoothing result more dependent on historical values. Then, the radial basis function interpolation method is used to extrapolate the smoothed anomaly confidence values ​​of discrete monitoring points to the continuous space of the entire bridge. The radial basis function selected is the thin plate spline kernel function. During the interpolation process, the parameters of the thin plate spline kernel function are calculated based on the structural distance between sensors rather than the Euclidean distance. The structural distance refers to the actual path length along the main beam or cable surface of the bridge, not the straight-line distance between two points. For each time window, the smoothed anomaly confidence values ​​of each monitoring point in that window are used as known points, and radial basis function interpolation is performed to obtain the anomaly confidence estimate at each spatial coordinate on the entire bridge structural surface. All anomaly confidence estimates are colored according to spatial coordinates, with warm colors used for areas with high anomaly confidence estimates and cool colors used for areas with low anomaly confidence estimates, forming a spatial anomaly confidence distribution cloud map for that time window. Finally, the spatial anomaly confidence distribution cloud maps of multiple consecutive time windows are stacked in chronological order to form a structural health state evolution map, that is, the spatial anomaly confidence distribution cloud maps of multiple consecutive time windows are arranged sequentially in chronological order to form a three-dimensional data structure. The spatial dimension of this three-dimensional data structure corresponds to the geometric coordinates of the bridge, and the temporal dimension corresponds to the time window number. Each voxel stores the anomaly confidence value of the spatial location within the time window. This three-dimensional data structure is used as a structural health status evolution diagram. A structural health status evolution diagram is a three-dimensional data structure that stores the anomaly confidence of each monitoring location changing over time, with spatial coordinates as the planar dimension and time as the vertical axis.

[0091] In practical implementation, health monitoring is conducted based on the structural health status evolution diagram. Specifically, a dynamic adaptive threshold can be constructed based on the long-range statistical memory and short-term fluctuation inertia of the deviation field under healthy conditions. This dynamic adaptive threshold determines the level of anomaly confidence required to issue an early warning. The dynamic adaptive threshold is dynamically adjusted based on the bridge's historical health statistics. Specifically, it uses the mean value in the long-range statistical memory as a baseline, adds an offset modulated by short-term fluctuation inertia, and multiplies it by a safety factor that increases non-linearly with service time. This safety factor can be set by expert recommendations. Long-range statistical memory refers to using the mean and standard deviation of anomaly confidence at various locations under healthy conditions over the past few weeks or even months as a benchmark reference, reflecting the normal fluctuation range of the structure during long-term operation. Short-term fluctuation inertia refers to the trend of anomaly confidence changes in the recent hours, reflecting the current load environment and response state of the structure. Then, the current value of the anomaly confidence time series curve at each spatial location is compared with the dynamic adaptive threshold. That is, for each time window and each spatial location, it is determined whether the current anomaly confidence at that location exceeds the dynamic adaptive threshold for that location in the current time window. If the dynamic adaptive threshold is not exceeded, the location is considered to be in a healthy state and monitoring continues. If the dynamic adaptive threshold is exceeded, the persistence of the exceedance is further examined. That is, a single exceedance may be caused by instantaneous disturbance or accidental extreme load, and is not a real structural damage. Therefore, it is required that the anomaly confidence level exceeds the threshold for three consecutive time windows before it is determined that there is a structural anomaly at the location.

[0092] Furthermore, in the specific implementation, after a structural anomaly is determined, spatial connectivity analysis based on the anomaly confidence level aggregates adjacent monitoring locations that simultaneously exceed limits into anomaly regions. That is, the spatial slice of the structural health state evolution map within the current time window is discretized into a grid. Grid cells exceeding the threshold are marked with eight-neighbor connectivity; if the cells adjacent to an exceeding-limit cell in the top, bottom, left, right, or four diagonal directions are also exceeding limits, these cells are grouped into the same connectivity domain, with each connectivity domain corresponding to an independent anomaly region. For each anomaly region, the mean anomaly confidence level of all grid cells within that region is calculated as the severity level of the anomaly region, and the spatial extent covered by the anomaly region is calculated as the geometric dimension of the damage. Finally, warning information and corresponding physical location identifiers are output. The warning information includes: the spatial coordinate range of the anomaly region, the severity level, the most likely type of damage, and the timestamp of the first exceedance. This warning information is visualized on the bridge's 3D model, and simultaneously, audible and visual alarms are issued via edge computing terminals and pushed to remote monitoring platforms and mobile terminals of maintenance personnel via communication links. Maintenance personnel conduct on-site verification and make maintenance decisions based on the early warning information, thereby realizing full-link bridge structure health monitoring from data collection, fusion analysis, damage identification to early warning output.

[0093] like Figure 2 As shown in the diagram, this solution provides a schematic of a bridge structural health monitoring system based on multi-source sensor fusion. It includes multi-source sensor units deployed at key parts of the bridge to collect various types of raw data, such as strain, acceleration, temperature, and deflection. The multi-source data acquisition module synchronizes data acquisition under a unified clock source. The collected raw data is transmitted to edge computing nodes, which perform computational tasks such as time alignment, anomaly removal, cross-modal correlation feature extraction, health status comparison, damage sensitivity analysis, and health status evolution diagram generation. Finally, the edge computing nodes upload the processing results and early warning information to a remote cloud platform for long-term model updates and data storage management.

[0094] Therefore, this application firstly generates a standardized data sequence with synchronized time and unified dimensions by performing time alignment and anomaly removal on the raw data from multi-source sensing, providing a data foundation for the accurate extraction of subsequent cross-modal correlation features; secondly, by fusing and analyzing the standardized data sequence and extracting cross-modal correlation features, the extracted correlation features contain the physical essence of the structural response and are highly sensitive to cross-modal coupling anomalies caused by damage, providing robust feature inputs for subsequent structural state characterization and damage identification; and thirdly, by constructing a multi-parameter coupled state vector and aligning it with a stable attractor manifold reference... By comparing domains, high-dimensional cross-modal correlation features are compressed into local deviation fields, enabling a quantitative description of the overall structural health status. This allows structural deviations to be mapped to the probability of different damage types occurring at each monitoring location, achieving damage location and type identification, and providing an intuitive decision-making basis for subsequent health monitoring and early warning. Finally, by accumulating the confidence distribution of structural anomalies over time and extrapolating it spatially, the anomaly information of discrete monitoring points is expanded into a dynamic evolution map of the entire bridge in continuous space, enabling accurate location and type identification of bridge structural damage and improving the anti-interference capability of health monitoring against environmental disturbances.

[0095] In summary, the technical solution adopted in this application can achieve accurate location and type identification of bridge structural damage through multi-source sensor fusion and cross-modal correlation feature extraction, thereby improving the anti-interference capability of health monitoring against environmental disturbances.

[0096] Example 2: This application provides a bridge structural health monitoring system based on multi-source sensor fusion, referring to... Figure 3 As shown in the figure, this is a modular structure diagram of a bridge structure health monitoring system based on multi-source sensor fusion according to this embodiment of the present application. The health monitoring system includes:

[0097] The data acquisition module 100 is used to acquire multi-source sensor raw data of key parts of the bridge, perform time alignment and anomaly removal on the multi-source sensor raw data, and generate a standardized data sequence.

[0098] The feature extraction module 200 is used to perform fusion analysis on the standardized data sequence, and then extract cross-modal correlation features between different physical quantities;

[0099] The confidence output module 300 is used to construct a multi-parameter coupled state vector based on the cross-modal correlation features, compare it with the health state reference domain, generate a structural deviation characterization, perform damage sensitivity analysis on the structural deviation characterization, and output the structural anomaly confidence distribution at each monitoring location.

[0100] The health monitoring module 400 is used to perform time-series accumulation and spatial extrapolation of the structural anomaly confidence distribution, generate a structural health state evolution map, and then perform health monitoring based on the structural health state evolution map.

[0101] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0102] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0103] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. A method for monitoring the structural health of bridges based on multi-source sensor fusion, characterized in that, The health monitoring method includes the following steps: Acquire multi-source sensor raw data of key parts of the bridge, perform time alignment and anomaly removal on the multi-source sensor raw data, and generate a standardized data sequence; The standardized data sequence is fused and analyzed to extract cross-modal correlation features between different physical quantities; Based on the cross-modal correlation features, a multi-parameter coupled state vector is constructed and compared with the health state benchmark domain to generate a structural deviation characterization. Damage sensitivity analysis is performed on the structural deviation characterization, and the structural anomaly confidence distribution at each monitoring location is output. The structural anomaly confidence distribution is accumulated over time and extrapolated spatially to generate a structural health status evolution map, and then health monitoring is carried out based on the structural health status evolution map.

2. The bridge structural health monitoring method based on multi-source sensor fusion as described in claim 1, characterized in that, The multi-source sensing raw data includes strain data, acceleration data, temperature data, and deflection data.

3. The bridge structural health monitoring method based on multi-source sensor fusion as described in claim 1, characterized in that, The fusion analysis of the standardized data sequence to extract cross-modal correlation features between different physical quantities specifically includes: The standardized data sequence is time-scale aligned to establish a synchronous mapping relationship between multiple data sources; Based on the propagation hysteresis characteristics of different physical quantities in the structural response, a time delay correlation matrix across physical quantities is constructed; Principal component analysis was performed on the time delay correlation matrix to determine the initial cross-modal correlation characteristics; By introducing a transfer entropy operator, the initial cross-modal correlation features are nonlinearly enhanced to determine the information flow intensity between each pair of physical quantities, thereby obtaining the cross-modal correlation features between different physical quantities.

4. The bridge structural health monitoring method based on multi-source sensor fusion as described in claim 3, characterized in that, The construction of the time delay correlation matrix across physical quantities specifically includes: The response propagation time lag between each pair of sensing sequences is estimated using a cross-correlation function, and then a time delay matrix is ​​constructed. Construct a confidence matrix based on the magnitude of the cross-correlation peaks; The time delay correlation matrix is ​​determined based on the time delay matrix and the confidence matrix.

5. The bridge structural health monitoring method based on multi-source sensor fusion as described in claim 1, characterized in that, Based on the cross-modal correlation features, a multi-parameter coupled state vector is constructed and compared with the health state baseline domain to generate a structural deviation representation, specifically including: The cross-modal correlation features are flattened and normalized within a time window, and then spliced ​​together to form a multi-parameter coupled state vector. During the calibration window period in the early stage of bridge operation, a set of multi-parameter coupled state vectors under no abnormal conditions is collected, the stable attractor manifold of the set of multi-parameter coupled state vectors in the state space is learned, and the stable attractor manifold is used as the reference domain for the healthy state. The multi-parameter coupled state vector at the current moment is projected onto the stable attractor manifold to determine the projection residual, which is then used as a characterization of structural deviation.

6. The bridge structural health monitoring method based on multi-source sensor fusion as described in claim 5, characterized in that, The learning method for the stable attractor manifold specifically includes: A local linear embedding algorithm is used to perform dimensionality reduction manifold learning on the multi-parameter coupled state vector within the calibration window to obtain low-dimensional embedding coordinates. In a low-dimensional embedding space, the convex hull surface of all calibrated sample points is determined, and the convex hull surface is extended outward by an adaptive tolerance distance to determine the boundary of the stable attractor manifold.

7. The bridge structural health monitoring method based on multi-source sensor fusion as described in claim 1, characterized in that, Damage sensitivity analysis is performed on the structural deviation characterization, and the structural anomaly confidence distribution at each monitoring location is output, specifically including: The structural deviation characterization is organized according to the spatial location of the sensor to form a spatial deviation distribution field; Establish a set of parallel response difference enhancement operators, each corresponding to a preset damage type; The spatial deviation distribution field is input into each response difference enhancement operator, and each response difference enhancement operator outputs the probability of occurrence of the corresponding damage type at each monitoring location; The probability of occurrence of the output of each response difference enhancement operator is competitively normalized, thereby outputting the structural anomaly confidence distribution at each monitoring location.

8. The bridge structural health monitoring method based on multi-source sensor fusion as described in claim 1, characterized in that, The process of accumulating and spatially extrapolating the confidence distribution of structural anomalies to generate a structural health state evolution map specifically includes: Based on the structural anomaly confidence distribution at each monitoring location, the time-series evolution curve of structural anomaly confidence at each monitoring location is determined; The time-series evolution curve of the structural anomaly confidence is smoothed, and the smoothed time-series evolution curve of the structural anomaly confidence is extrapolated to the continuous space of the full bridge to generate a spatial anomaly confidence distribution cloud map under each time window. The spatial anomaly confidence distribution cloud maps of multiple consecutive time windows are stacked in chronological order to form a structural health state evolution map.

9. The bridge structural health monitoring method based on multi-source sensor fusion as described in claim 1, characterized in that, The structural health status evolution diagram refers to a three-dimensional data structure that stores the change of anomaly confidence at each monitoring location over time, with spatial coordinates as the planar dimension and time as the vertical axis.

10. A bridge structural health monitoring system based on multi-source sensor fusion, used to execute a bridge structural health monitoring method based on multi-source sensor fusion as described in any one of claims 1 to 9, characterized in that, The health monitoring system includes: The data acquisition module is used to acquire multi-source sensor raw data of key parts of the bridge, perform time alignment and anomaly removal on the multi-source sensor raw data, and generate a standardized data sequence. The feature extraction module is used to perform fusion analysis on the standardized data sequence, and then extract cross-modal correlation features between different physical quantities; The confidence output module is used to construct a multi-parameter coupled state vector based on the cross-modal correlation features, compare it with the health state reference domain, generate a structural deviation characterization, perform damage sensitivity analysis on the structural deviation characterization, and output the structural anomaly confidence distribution at each monitoring location. The health monitoring module is used to perform temporal accumulation and spatial extrapolation of the structural anomaly confidence distribution, generate a structural health state evolution map, and then perform health monitoring based on the structural health state evolution map.

Images