A bridge health monitoring system, method and storage medium

Abstract

This invention relates to the field of bridge health monitoring technology, and discloses a bridge health monitoring system, method, and storage medium. The method collects multimodal monitoring raw data streams of bridges through a distributed sensor network, and performs intelligent data purification and heterogeneous data integration to generate a unified format bridge health data set. Time series markers, spatial location codes, and monitoring parameter dimension identifiers are decoded from this data set. The data is divided into overlapping time segments based on the time series markers, generating a time segment dataset. Spatial location codes are fused, mapping the spatial information in the time segment datasets to the bridge's global coordinate frame to establish a dynamic health status map. Based on this map and the time segment datasets, multi-dimensional data fusion is performed according to the monitoring parameter dimension identifiers. The state evolution path of bridge components is analyzed through the fusion results, generating health metrics and prediction sequences.

Description

Technical Field

[0001] This invention relates to the field of bridge health monitoring technology, specifically to a bridge health monitoring system, method, and storage medium. Background Technology

[0002] As a critical component of transportation infrastructure, the long-term safe operation of bridges is of paramount importance. Traditional bridge health monitoring mainly relies on periodic manual inspections and simple sensor data threshold alarms. Manual inspections are highly subjective, have long intervals, are difficult to detect potential, gradual structural damage, and cannot achieve continuous monitoring around the clock. Simple sensor threshold alarms can only be triggered when the monitored physical quantity exceeds a preset limit, but the setting of this limit often depends on experience, there is a contradiction between sensitivity and false alarm rate, and it cannot reveal early signs and development trends of damage.

[0003] With the development of sensing technology, it has become possible to deploy multiple sensors on bridges to form a distributed monitoring network, generating massive amounts of multimodal monitoring data. This data varies in type, sampling rate, and data format, i.e., it is heterogeneous data. Current methods typically process and analyze these data independently, lacking an effective integration mechanism, making it difficult to form a unified and collaborative understanding of the overall health status of the bridge. Noise and outliers in the data also affect the accuracy of the analysis.

[0004] At the data analysis level, traditional methods often focus on independent time-series analysis of individual sensor data sequences or local area data, failing to fully explore the spatiotemporal correlations inherent in the data. A bridge is a spatially continuous structure, and the responses of different parts are interconnected; local damage effects can propagate through the structure to affect other parts. Ignoring this spatial correlation and conducting only isolated point analyses makes it difficult to accurately assess the location and extent of damage and its impact on the overall structural performance. Furthermore, the evolution of a bridge's health status is a dynamic process; the occurrence and development of damage are temporal, requiring analytical methods capable of capturing the patterns of state evolution over time.

[0005] Existing monitoring systems often focus on assessing the current state and lack the ability to predict future trends. The lack of effective predictive models means that maintenance management decisions are largely based on the current situation, making it difficult to achieve predictive maintenance and optimize maintenance resources and timing. Therefore, existing bridge health monitoring technologies have significant shortcomings in areas such as multi-source heterogeneous data fusion, spatiotemporal correlation modeling, dynamic state tracking, and trend prediction, limiting the depth and foresight of monitoring results. Summary of the Invention

[0006] The purpose of this invention is to provide a bridge health monitoring system, method, and storage medium to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a bridge health monitoring method, the method comprising:

[0008] The raw data stream of multimodal monitoring of the bridge is collected through a distributed sensor network;

[0009] The original multimodal monitoring data stream is intelligently cleaned and heterogeneous data is integrated to generate a unified format bridge health data set;

[0010] The time series markers, spatial location codes, and monitoring parameter dimension identifiers are decoded from the bridge health dataset; the bridge health dataset is divided into overlapping time segments based on the time series markers to generate a time segment dataset;

[0011] By integrating spatial location coding, spatial information in time-segment datasets is mapped to a global coordinate framework for bridges to establish a dynamic health status map;

[0012] Based on dynamic health status maps and time segment datasets, multi-dimensional data fusion is performed according to the dimension identifiers of monitoring parameters;

[0013] By analyzing the state evolution path of bridge components through fusion results, health metrics and prediction sequences are generated.

[0014] Preferably, the multimodal monitoring raw data stream is subjected to intelligent data purification and heterogeneous data integration to generate a bridge health data set in a unified format, including: applying a deep learning-based noise filtering model to remove outliers and smooth signals in the multimodal monitoring raw data stream to obtain a purified data stream; converting heterogeneous field names in the purified data stream into standardized data patterns to generate a standardized data stream; compensating for and synchronizing time deviations in the standardized data stream to output a bridge health data set in a unified format.

[0015] Preferably, spatial location coding is integrated to map the spatial information in the time segment dataset to the bridge's global coordinate framework to establish a dynamic health status map. This includes: extracting spatial location coding from the time segment dataset; converting local coordinates into a global coordinate point set using a coordinate transformation algorithm; associating the monitoring data within each time segment with the global coordinate point set based on sensor identifiers to generate a spatiotemporal correlation dataset; organizing the coordinate point set in the spatiotemporal correlation dataset into nodes and adding weighted edges based on the data flow direction to form a dynamic health status map.

[0016] Preferably, multi-dimensional data fusion is performed according to the monitoring parameter dimension identifier, including: dividing the time segment dataset into multiple parameter subsets according to the monitoring parameter dimension identifier, and generating a three-dimensional data cube of parameter-node-time segment by combining the node information in the dynamic health status map; applying a feature extraction algorithm to each parameter subset in the three-dimensional data cube to calculate the statistical feature vector of the parameter on the node; and aggregating the statistical feature vectors into a multi-dimensional health feature matrix.

[0017] The statistical feature vectors are aggregated into a multidimensional health feature matrix, including: normalizing the statistical feature vectors of each parameter subset to obtain standardized feature vectors; calculating the weight coefficients of the standardized feature vectors using an attention mechanism; and performing a weighted summation of the standardized feature vectors based on the weight coefficients to generate a multidimensional health feature matrix.

[0018] Preferably, the heterogeneous field names in the cleaned data stream are converted into a standardized data pattern to generate a standardized data stream, including: parsing all field names in the cleaned data stream to generate a list of field names; referring to a preset standardized dictionary, mapping synonymous fields in the field name list to unified terms to generate a mapped field set; reorganizing the data structure of the mapped field set according to the standardized data pattern, and outputting the standardized data stream.

[0019] Preferably, a time deviation in the standardized data stream is compensated and synchronized using a timestamp correction algorithm to output a bridge health data set in a unified format. This includes: extracting timestamp fields and sensor identification fields from the standardized data stream and combining them into a synchronization key; detecting and aligning timestamp offsets based on the synchronization key and a preset time window threshold; and using interpolation methods to fill in data at missing time points to generate a time-synchronized dataset as a bridge health data set in a unified format.

[0020] Preferably, the missing time points are filled with data using an interpolation method to generate a time-synchronized dataset, including: identifying gap intervals in the timestamp sequence; calculating missing values ​​using a linear interpolation algorithm based on the trend of data points before and after the gap interval; and inserting the interpolation results into a standardized data stream to form a continuous time-synchronized dataset.

[0021] Preferably, the coordinate point set in the spatiotemporal correlation dataset is organized into nodes, and weighted edges are added based on the data flow direction to form a dynamic health status graph. This includes: mapping each global coordinate point in the spatiotemporal correlation dataset into a graph node and assigning a unique node identifier; determining the connection relationship between nodes according to the transmission path of the monitoring data and using data flow as the edge weight; using a graph database to store node and edge information and constructing a weighted directed graph structure for the dynamic health status graph.

[0022] The connection relationship between nodes is determined based on the transmission path of the monitoring data, and the data flow is used as the edge weight. This includes: analyzing the historical flow pattern of sensor data in the spatiotemporal correlation dataset to identify source and target node pairs; accumulating the data transfer volume between node pairs in each time segment and calculating the standardized flow value; and assigning the standardized flow value as the weight attribute of the corresponding edge.

[0023] Preferably, the present invention also includes a bridge health monitoring system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the bridge health monitoring method described above.

[0024] Preferably, the present invention further includes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the various steps of the bridge health monitoring method described above.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] By acquiring raw multimodal monitoring data streams of the bridge through a distributed sensor network, a comprehensive understanding of the bridge's structural condition was achieved. This distributed acquisition method can cover key parts and important components of the bridge, obtaining monitoring data of various types, including vibration, strain, and displacement. The synchronous acquisition of multimodal data provides a rich source of information for subsequent comprehensive analysis, avoiding the limitations that may arise from a single data type.

[0027] Intelligent data purification and heterogeneous data integration of the raw multimodal monitoring data stream significantly improved data quality. The data purification process effectively eliminated environmental noise and interference introduced by the acquisition equipment, retaining effective information that truly reflects the bridge's condition. Heterogeneous data integration transformed monitoring data from different sources and in different formats into a unified bridge health data set, establishing a standardized data foundation for subsequent analysis and processing.

[0028] By decoding time-series markers, spatial location codes, and monitoring parameter dimension identifiers from the bridge health dataset, multi-dimensional analysis of the monitoring data was achieved. The time-series markers accurately record the time information of data collection, the spatial location codes precisely locate the spatial distribution of monitoring points, and the monitoring parameter dimension identifiers clearly distinguish different types of monitoring parameters. This multi-dimensional analysis provides complete metadata support for in-depth analysis of bridge health status.

[0029] The bridge health data set is divided into overlapping time segments based on time series labels, generating a time segment dataset. This time segmentation method ensures the temporal continuity of the data while facilitating the analysis of the dynamic changes in bridge condition. The design of overlapping time segments avoids the omission of important information, ensuring the completeness of the time series analysis.

[0030] By integrating spatial location coding, spatial information from time-segment datasets is mapped onto a global coordinate framework for the bridge, establishing a dynamic health status map. This spatial mapping method achieves a precise correspondence between monitoring data and the actual spatial location of the bridge, and the dynamic health status map intuitively displays the distribution and changing trends of the health status of various parts of the bridge.

[0031] Based on dynamic health status maps and time-segment datasets, multi-dimensional data fusion is performed according to the dimensional identifiers of monitoring parameters. This multi-dimensional fusion method comprehensively considers multiple factors such as time, space, and parameter type, and can comprehensively reflect the overall health status of the bridge. Collaborative analysis of data from different dimensions helps to discover potential problems that are difficult to identify from a single dimension. Attached Figure Description

[0032] Figure 1 This is a schematic diagram illustrating the working principle of the bridge health monitoring method described in this invention.

[0033] Figure 2 A flowchart for processing the raw data stream of multimodal monitoring;

[0034] Figure 3 Flowchart for establishing a dynamic health status map;

[0035] Figure 4 A standardized data flow heatmap between bridge monitoring nodes;

[0036] Figure 5 A heatmap of the multidimensional health feature matrix of bridge sensor nodes. Detailed Implementation

[0037] 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.

[0038] Please see Figure 1This invention provides a bridge health monitoring system, method, and storage medium. The method includes: continuously collecting multimodal monitoring raw data streams reflecting the structural state of the bridge through a distributed sensor network pre-deployed on key bridge components. The sensor network includes various types of sensors, such as strain gauges, accelerometers, and displacement gauges. Intelligent data purification and heterogeneous data integration operations are performed on the collected multimodal monitoring raw data streams. This process aims to eliminate noise interference in the raw data and unify the data formats from different manufacturers or protocols, ultimately generating a unified format bridge health data set with consistent structure and reliable quality. Three key information dimensions are decoded from this bridge health data set: time series markers, spatial location codes, and monitoring parameter dimension identifiers. Based on the time series markers, the entire bridge health data set is segmented into a series of time segments with overlapping regions, thereby generating a time segment dataset that facilitates time-series analysis. The spatial location code information decoded from the data is fused, and the local spatial information corresponding to each data point in the time segment dataset is mapped to a pre-defined global bridge coordinate framework through coordinate transformation. Based on this, a dynamic health status map reflecting the spatiotemporal correlation of the data is constructed. Supported by this atlas and the aforementioned time-segment dataset, multi-dimensional data fusion analysis is performed on various monitoring parameters from different sensors based on the dimensional identifiers of the monitoring parameters. By analyzing the fusion results, the path of the evolution of the state of each bridge component over time is analyzed, and health metrics for quantitatively assessing the bridge's health status and prediction sequences for predicting future state changes are generated.

[0039] Example 1: See Figure 2 In practical implementation, a deep learning-based noise filtering model performs outlier removal and signal smoothing on the raw multimodal monitoring data stream. This raw data stream originates directly from the distributed sensor network deployed at the bridge site and includes direct measurements of raw voltage signals, frequency signals, and other physical quantities. The deep learning-based noise filtering model is implemented using a one-dimensional convolutional autoencoder (DICO) structure. This DICO is trained unsupervised using a large amount of historical bridge monitoring data under normal operating conditions, learning the characteristic distribution patterns of normal monitoring data. In practice, the raw multimodal monitoring data stream is input into the trained DICO. The encoder part of the DICO extracts the feature representation of the input signal, and the decoder part reconstructs the clean signal based on this feature representation. By calculating the residual between the original input signal and the reconstructed signal, abnormal fluctuation points that significantly deviate from the normal pattern can be identified. These abnormal fluctuation points are judged as outliers and removed. Signal smoothing is achieved through multiple one-dimensional convolutional layers and downsampling operations within the one-dimensional convolutional autoencoder. These operations effectively filter out high-frequency random noise while preserving the overall trend and main features of the signal, ultimately outputting a clean data stream.

[0040] In practice, parsing all field names in the cleaned data stream and generating a list of field names is a crucial step in heterogeneous data integration. The cleaned data stream may contain data records from different manufacturers and sensor models, each with a field name describing its content. The system iterates through each data record in the cleaned data stream, extracting all occurrences of the field name strings and storing them in a temporary set to avoid duplication, ultimately generating a list of field names containing all unique field names. After the field name list is generated, the system loads a predefined standardized dictionary, which is a predefined mapping table storing the mapping relationships between various common but inconsistently expressed synonymous field names and a set of standard terms. The system matches each field name string in the field name list with a key in the predefined standardized dictionary. If a field name string is a synonym or alias of a key in the predefined standardized dictionary, it maps that field name to a unified term defined in the predefined standardized dictionary. This process is performed sequentially on each entry in the field name list, producing a mapped field set. In some embodiments, the pre-defined standardized dictionary supports regular expression matching to handle minor spelling differences or additional suffixes in field names. After generating the mapped field set, the system reorganizes the data according to a predefined standardized data pattern. The standardized data pattern explicitly specifies the field names, data types, and hierarchical structure that the final data set should include. The system checks whether each field in the mapped field set conforms to the definition in the standardized data pattern. For fields whose names are mapped to standard terms, their corresponding data values ​​are extracted and converted and placed according to the format and type specified by the standardized data pattern. For fields defined in the standardized data pattern but missing in the mapped field set, the system fills them with null values ​​or default values. For redundant fields present in the mapped field set but not defined in the standardized data pattern, their data is discarded. After the above reorganization, a standardized data stream is output.

[0041] In implementation, timestamp and sensor identifier fields are extracted from the standardized data stream and combined into a synchronization key. The timestamp field records the specific moment of data acquisition, and the sensor identifier field uniquely identifies the source sensor of the data. The system generates a synchronization key for each data record, which is constructed by concatenating the string value of the sensor identifier field with the numerical value of the timestamp field. This synchronization key serves as the unique identifier for that data record in subsequent time synchronization processes. Based on the synchronization key and a preset time window threshold, timestamp offsets are detected and aligned. The preset time window threshold defines the maximum time deviation allowed between data from different sensors. The system first sorts all data records according to the timestamp component in the synchronization key, and then checks the timestamp differences of data records from different sensors using a reference time axis as a benchmark. If the timestamp difference of the same physical event recorded by different sensors is less than or equal to the preset time window threshold, the system will adjust these timestamps to a unified time point, such as adjusting them to the average of these timestamps. If the timestamp difference exceeds the preset time window threshold, it may indicate a serious clock drift or data error, and the system will record an alarm log and may require manual intervention. Optionally, the timestamp alignment algorithm can employ dynamic time warping techniques to handle nonlinear offsets caused by minute differences in sampling rates between different sensors.

[0042] In practical implementation, interpolation methods are used to fill in missing time points to generate a time-synchronized dataset, and identifying gaps in the timestamp sequence is the primary operation. After completing timestamp alignment, the system checks along the reference time axis for interruptions in the continuous time point sequence; these interruptions are the gaps. Gap identification is achieved by comparing the interval between consecutive time points with the theoretical sampling interval. If the actual interval is significantly larger than the theoretical sampling interval, the interval is determined to be a gap with missing data. Based on the trend of data points before and after the gap, a linear interpolation algorithm is applied to calculate the missing value. The linear interpolation algorithm assumes that data changes linearly between two known points within a short period. For each missing time point within the gap, the system finds the nearest valid data point before and after that missing time point, treating these two valid data points as two points on a straight line. Based on the position of the missing time point between these two points, its corresponding data value is calculated linearly. It can be understood that the linear interpolation algorithm is simple to calculate and highly efficient, suitable for most sensor data showing trends within short time intervals. The interpolation results are inserted into the standardized data stream to form a continuous time-series time-synchronized dataset. The system creates new data records with timestamps representing missing time points, data field values ​​representing the calculated interpolation results, and sensor identification fields determined based on the data stream source of the gap interval. These new records, along with the existing, time-aligned valid data records, are arranged in timestamp order to ultimately form a continuous and complete time-synchronized dataset, which is the bridge health data set in a unified format. In some embodiments, for certain types of parameters with highly non-linear variation patterns, more complex interpolation methods such as spline interpolation can be used as alternative solutions.

[0043] Example 2: See Figure 3In practical implementation, extracting spatial location codes from the time-segment dataset is the foundation for establishing spatial mapping. The time-segment dataset originates from the result of dividing a unified format bridge health data set according to time series labels. Each data record contains a spatial location code field. The spatial location code field typically exists in the form of coordinate values ​​in the sensor's local coordinate system, such as three-dimensional coordinates with the sensor installation point as the origin. The system reads each record in the time-segment dataset and parses the values ​​of the spatial location code fields. These values ​​constitute the input for subsequent coordinate transformation. The local coordinates are converted into a global coordinate point set through a coordinate transformation algorithm. The coordinate transformation algorithm requires a set of pre-calibrated transformation parameters, including a rotation matrix and a translation vector. The rotation matrix and translation vector are jointly determined by the bridge design drawings and the sensor's on-site installation measurement data. In practical implementation, for each local coordinate extracted from the data record, the system applies the coordinate transformation algorithm, using a rotation matrix to perform a rotation transformation on the local coordinates, and then using a translation vector to perform a translation transformation, finally calculating the three-dimensional coordinates of the point in the bridge's global coordinate system. The above transformation operations are performed sequentially on the spatial location codes of all records to generate a global coordinate point set that corresponds one-to-one with each time-segment data record. In some embodiments, the coordinate transformation algorithm also needs to consider the scale factor to correct for possible minute scale differences between different measurement units.

[0044] In implementation, monitoring data within each time segment is associated with a global coordinate point set based on sensor identifiers. The sensor identifier is a unique string or numeric code that identifies each sensor and exists in every data record. The system uses the sensor identifier as the association key to match and connect the monitoring data fields in the time segment dataset with the corresponding coordinate points in the global coordinate point set obtained through coordinate transformation. For data records and coordinate points with the same sensor identifier, the system creates a new data record containing the original monitoring data, timestamp information, and the corresponding three-dimensional coordinates in the global coordinate system, generating a spatiotemporal correlated dataset. This spatiotemporal correlated dataset is a collection of data with enhanced spatial location information; each record is explicitly associated with a specific physical location on the bridge structure. It can be understood that the generation of the spatiotemporal correlated dataset gives the data not only a temporal dimension but also a precise spatial dimension under a unified spatial reference system, laying the data foundation for subsequent map construction. Optionally, the association operation can be efficiently implemented in the database using the join query function of a structured query language, especially when processing large-scale time segment datasets.

[0045] In practical implementation, the coordinate points in the spatiotemporal correlation dataset are organized into nodes. Each global coordinate point in the spatiotemporal correlation dataset represents a sensor deployment location on the bridge. The system traverses the spatiotemporal correlation dataset, creating a graph node object for each unique global coordinate point, and assigning each graph node object a unique node identifier across the entire network. The node identifier can be generated based on the hash value of the global coordinate point or using an incrementing sequence number. In addition to the node identifier, the graph node object typically contains attributes such as the node's global coordinate value and associated sensor identifiers. Weighted edges are added based on the data flow direction, which is determined by analyzing the transmission path of monitoring data in the sensor network. The system needs to define the connection relationships between nodes, which reflect the logical or physical transmission path of data from the source node to the target node; for example, data flows from sensor nodes deployed at bridge supports to sensor nodes deployed in the bridge span, or data flows from edge acquisition nodes to regional aggregation nodes and then to the central processing node. After determining the connection relationships, the system creates a directed edge for each pair of source and target nodes with a data flow direction, with the edge pointing from the source node to the target node. The weights of the edges need to be assigned values, which are calculated based on the data flow between node pairs within a certain time segment, forming the prototype of a weighted directed graph structure. In some embodiments, the data flow direction can be preset according to the mechanical force transmission path of the bridge structure, so that the graph can better reflect the physical characteristics of the structure.

[0046] In practical implementation, a graph database is used to store node and edge information, constructing a dynamic health status graph with a weighted directed graph structure. The graph database is a database management system specifically designed for storing and querying graph-structured data. The system persistently stores the created graph node objects and weighted directed edge objects in the graph database. The graph database uses node identifiers as primary keys to store the attributes of each node, and source and target node identifiers as foreign keys to store the relationship and weight attributes of edges. The construction of the dynamic health status graph is not a one-time event, but is dynamically updated with the arrival of new time-segment datasets. When processing new time segments, the system re-executes the coordinate transformation, data association, and edge weight calculation processes, and updates the attributes of the corresponding nodes and edges in the graph database. In particular, the edge weights are refreshed according to the latest data flow. It can be understood that using a graph database for storage facilitates complex graph traversal and graph analysis queries, such as finding the shortest path between two nodes or identifying key nodes in the network. The final constructed dynamic health status graph is a mathematical model that can intuitively express how monitoring data is distributed, associated, and flows on the bridge's spatial structure. Optionally, the visualization of the dynamic health status map can be achieved through the visualization tools built into the graph database or a third-party graphics library to help engineers understand the health status of the bridge.

[0047] In practical implementation, the historical flow patterns of sensor data in the spatiotemporal correlation dataset are analyzed to identify source and target node pairs. These historical flow patterns are obtained by querying the data source and destination information recorded in the spatiotemporal correlation dataset. The system analyzes a large number of data records over a period of time, statistically analyzing which sensor node each record indicates from which data originated and which sensor node or processing node ultimately received it. Through aggregation analysis, the system can identify frequently occurring node combinations with stable data transmission relationships; these combinations are the source and target node pairs. The data transfer volume between node pairs within each time segment is accumulated, quantified as bytes, data packets, or valid records. The system maintains a counter for each identified source and target node pair. When a data record belonging to the data flow of that node pair is found in the spatiotemporal correlation dataset, the counter value is increased based on the data size or unit count of that record. After each time segment is processed, the accumulated counter value represents the data transfer volume between that node pair within that time segment. Calculating standardized flow values ​​eliminates the influence of absolute numerical values, ensuring comparability of edge weights between different node pairs. The system standardizes the accumulated data transfer volume, for example, by using min-max normalization to linearly transform the data transfer volume to a range between zero and one; or by using the proportion relative to the total network traffic as the standardized traffic value. This standardized traffic value is assigned as the weight attribute of the corresponding edge. When creating or updating a directed edge from the source node to the target node, the system stores the calculated standardized traffic value as an attribute field of that edge in the graph database. It can be understood that the edge weight attribute dynamically reflects the strength of data interaction between nodes; a larger weight indicates a tighter connection and more frequent data flow, which may mean a stronger correlation in the structural responses of these two locations in subsequent analysis.

[0048] See Figure 4This is a heatmap, using source and target nodes as dimensions. It visually presents the standardized data flow distribution among eight sensor nodes through color gradients. Numerical labels quantify the intensity of data interaction between nodes, and the color bars on the right clearly define the range of standardized data flow values. This diagram corresponds to a core step in constructing a dynamic health status graph: after fusing spatial location encoding and generating a spatiotemporal correlated dataset, the connection relationships between nodes need to be determined based on the transmission paths of the monitoring data, using data flow as edge weights. The flow distribution among nodes in the diagram reflects the flow pattern of sensor data on the bridge structure; nodes with close structural connections interact more frequently. This diagram is a visual representation of constructing a weighted directed graph structure for dynamic health status graphs: by quantifying data flow between nodes, the system can identify key data transmission paths, providing topological support for subsequent multi-dimensional data fusion and health status assessment based on the graph, fully demonstrating the mining and application of data spatial flow characteristics in spatiotemporal correlation modeling.

[0049] Example 3: In specific implementation, the time-segment dataset is divided into multiple parameter subsets based on the monitoring parameter dimension identifiers. These identifiers are labels that identify the physical meaning of the monitoring parameters, such as strain parameters, acceleration parameters, and displacement parameters. The system reads each record in the time-segment dataset and groups them according to the monitoring parameter dimension identifier field value. Records with the same monitoring parameter dimension identifier value are grouped into the same group, forming a parameter subset for a single monitoring parameter. Combining this with node information in the dynamic health status map, where nodes correspond to sensor physical locations, the system associates each record in the parameter subset with a specific node in the dynamic health status map using sensor identifiers. This ensures that each data point is clearly assigned to a specific node location and monitoring parameter category. A three-dimensional data cube of parameters, nodes, and time segments is generated. This three-dimensional data cube is a logical data model, with its three dimensions representing different monitoring parameter types, different node locations, and different time segment indices, respectively. The system constructs a multidimensional array structure in memory. The first dimension index of the array corresponds to different subsets of monitoring parameters, the second dimension index corresponds to different node identifiers, and the third dimension index corresponds to different time segment numbers. Each cell of the array stores the original monitoring data sequence or preliminary summary value.

[0050] In practical implementation, a feature extraction algorithm is applied to each parameter subset in the 3D data cube. This algorithm aims to extract representative statistical features from the original data sequence. The system traverses each parameter category of the 3D data cube. For each node within each parameter category, and further for each time segment within each node, it extracts all data points of the monitored parameter at that node within that time segment to form a time series segment. The system calculates a statistical feature vector for the parameter at each node. For each extracted time series segment, the system calculates a set of statistics, including arithmetic mean, standard deviation, maximum value, minimum value, skewness, and kurtosis. These statistics are sequentially arranged to form an ordered list of values, which is the statistical feature vector. The statistical feature vector characterizes the behavioral characteristics of the monitored parameter at a specific node and within a specific time segment. Essentially, the statistical feature vector converts the variable-length original data sequence into a fixed-length feature representation, facilitating subsequent fusion and comparative analysis. In some embodiments, the feature extraction algorithm may also include calculating richer time-domain features such as root mean square value, waveform factor, and impulse factor to more comprehensively describe the data characteristics.

[0051] In practical implementation, statistical feature vectors are aggregated into a multidimensional health feature matrix, which is the final fusion feature representation used for health status assessment. The statistical feature vectors of each parameter subset are normalized to eliminate differences in units and numerical ranges between different statistical features. The system iterates through all statistical feature vectors for the same monitoring parameter, calculating the maximum and minimum values ​​of each feature dimension across all nodes and time segments. A min-max normalization method is used to linearly transform all values ​​in that dimension to the interval [0,1], resulting in a standardized feature vector. An attention mechanism is used to calculate the weight coefficients of the standardized feature vectors. The attention mechanism is an algorithm that automatically allocates attention based on feature importance. The system inputs the standardized feature vectors into an attention network, which calculates the relative importance of each feature element using learned parameters. The calculation process of the weight coefficients can be expressed in the following mathematical form:

[0052]

[0053] Where: symbol Represents the first in the standardized eigenvector Each feature element It is a learnable weight matrix. It is a learnable bias vector. It is the transpose of a learnable attention weight vector. It is the hyperbolic tangent activation function. It is the total number of feature elements in the standardized feature vector. It is the calculated number of Each feature element corresponds to a weight coefficient. The standardized feature vector is weighted and summed based on these weight coefficients to generate a multidimensional health feature matrix. The system multiplies each feature element in the standardized feature vector by its corresponding weight coefficient. Then, all weighted feature elements are summed to obtain a fused feature value representing the health status of that node-time segment unit. This weighted summation operation is performed once for each node in each time segment, ultimately generating a two-dimensional matrix. The rows of the matrix correspond to node indices, the columns correspond to time segment indices, and the matrix elements are the fused feature values. This matrix is ​​the multidimensional health feature matrix. It can be understood that the attention mechanism enables the model to autonomously focus on features more relevant to the health status, improving the quality of feature representation. In some embodiments, the weighted summation operation can be replaced by other feature aggregation methods such as weighted averaging or weighted concatenation. Optionally, the learnable parameters in the attention mechanism... , , The weights can be obtained through supervised training using historical data and bridge health status labels. In some embodiments, the calculation of weight coefficients can incorporate a multi-head attention mechanism to capture feature information from different subspaces, thereby enhancing the model's expressive power.

[0054] See Figure 5 This is a heatmap, using sensor nodes and monitoring time as two dimensions. It visually presents the changes in health characteristic values ​​of each node at different times through color gradients, with the color bars on the right clearly indicating the range of health characteristic values. This diagram corresponds to the core step of multi-dimensional data fusion and health feature aggregation: based on a dynamic health status map and time segment dataset, the system first generates a three-dimensional data cube of parameters-nodes-time segments according to the monitoring parameter dimension, then extracts statistical feature vectors from each parameter subset, and finally aggregates them into a multi-dimensional health feature matrix through an attention mechanism. The temporal evolution of node health characteristics in the diagram intuitively depicts the dynamic changing trend of bridge component status, providing core data support for subsequent analysis of component status evolution paths, generation of health measurements and prediction sequences. It fully demonstrates the ability to fuse three-dimensional features of time, space, and parameters in multi-source data collaborative analysis, serving as a key visual bridge from data fusion to health assessment.

[0055] Example 4: In specific implementation, the timestamp field and sensor identifier field are extracted from the standardized data stream and combined into a synchronization key. The standardized data stream is a data set after field name standardization. The system reads data records in the standardized data stream one by one, parsing out the timestamp field of the recorded data acquisition time and the sensor identifier field identifying the data source device from each record. The system concatenates the string value of the sensor identifier field with the value of the timestamp field to form a composite string as the synchronization key for that record. The synchronization key uniquely identifies the data record generated by a specific sensor at a specific time in the entire system. Based on the synchronization key and a preset time window threshold, the system detects and aligns the timestamp offset. The preset time window threshold is a time length parameter that defines the maximum time deviation allowed between data from different sensors. The system first sorts all data records in ascending order according to the timestamp component in the synchronization key, and then scans the sorted record sequence based on a unified reference time axis to check the timestamp differences of data records from different sensors. When multiple data records representing the same or adjacent physical events are found to have timestamp differences less than or equal to a preset time window threshold, the system corrects the timestamps of these records to a consistent value, such as the average or minimum value of the set of timestamps, thus completing timestamp alignment. If the timestamps of some records cannot be correlated with those of surrounding records within the preset time window threshold, these records may be considered abnormal and require special handling. It can be understood that the timestamp alignment operation eliminates time base differences caused by asynchronous clocks within individual sensors.

[0056] In practice, the first step is to use interpolation methods to fill in missing time points and identify gaps in the timestamp sequence. After aligning the timestamps, the system checks the continuity of the timestamp sequence against a reference time axis. The continuity is determined by calculating the interval between adjacent timestamps and comparing it with the theoretical sampling interval. If the actual time interval between two consecutive valid timestamps is greater than the theoretical sampling interval plus a tolerance range, it is determined that there are one or more missing time points between these two timestamps, and this missing time period is identified as a gap. The system records the start time, end time, and number of missing time points for each gap. Based on the trend of data points before and after the gap, a linear interpolation algorithm is applied to calculate the missing values. The linear interpolation algorithm is based on the assumption that data changes linearly between two known points within a short period of time. For each identified gap, the system obtains the most recent valid data record before the start time and the most recent valid data record after the end time, treating these two records as two known data points. The linear interpolation algorithm uses the numerical and temporal information of the two known data points to calculate the corresponding data value for each missing time point within the gap interval. The calculation principle is based on the assumption that the data value changes linearly with time, and the missing value is obtained by proportional calculation. Refer to Table 1 for the process of calculating missing values ​​using the linear interpolation algorithm.

[0057] Table 1: Linear Interpolation Calculation Table

[0058] Time point Data status Sensor A recorded value Calculation method description t1 valid data points 10.5 The last valid data point before the gap interval t2 Missing time points To be calculated The time points that need interpolation t3 Missing time points To be calculated The time points that need interpolation t4 valid data points 25.1 The first valid data point after the gap interval

[0059] For the example in Table 1, assuming time point t2 is located at one-third of the time span between t1 and t4, the interpolation calculation at t2 can be performed based on the values ​​of t1 and t4 in a time-proportional manner. The linear interpolation algorithm performs interpolation calculations independently for each parameter channel of each sensor. The interpolation results are inserted into the standardized data stream to form a time-synchronized dataset of continuous time series. The system creates a new data record for each calculated missing value. The timestamp field of the new record is set to the corresponding missing time point, the sensor identification field is inherited from the data stream to which the gap interval belongs, and the values ​​calculated by the linear interpolation algorithm are filled into each data field. These newly generated records are merged with the original, time-aligned valid data records and sorted according to the timestamp field, ultimately forming a continuous and complete time-synchronized dataset in the time dimension. This time-synchronized dataset is the final bridge health data set in a uniform format required. In some embodiments, for certain monitoring parameters exhibiting strong nonlinear characteristics, the system can provide configuration options to select different interpolation algorithms, such as spline interpolation. Optionally, interpolation operations can be performed in batches to reduce memory usage. The system divides the long-term series into multiple overlapping batches, performs gap identification and interpolation independently on each batch, and finally merges the results. It is understood that the generation of the time-synchronized dataset ensures that all subsequent analysis modules are based on unified, continuous time-series data, which is fundamental for accurate time-series analysis and state assessment. Optionally, the system records all location markers where interpolation operations have been performed, so that these interpolation points can be excluded in analyses requiring accurate raw data.

[0060] Example 5: In specific implementation, each global coordinate point in the spatiotemporal correlation dataset is mapped to a graph node. The spatiotemporal correlation dataset is a collection of data that has already associated monitoring data with coordinate points in the bridge's global coordinate system. The system reads each record in the spatiotemporal correlation dataset, extracts the global coordinate point information contained in the record, and each unique global coordinate point corresponds to a specific physical location on the bridge. The system creates a graph node object for each unique global coordinate point encountered and assigns a unique node identifier to each graph node object across the entire network. The node identifier can be generated using a hash algorithm based on coordinate values ​​or an auto-incrementing sequence number initialized at system startup. In addition to storing the node identifier, the graph node object also stores the corresponding three-dimensional coordinate values ​​of the global coordinate point, the associated sensor identifier, and sensor type, among other attribute information. A unique node identifier is assigned, ensuring that each node can be accurately referenced and identified in subsequent graph operations.

[0061] In practical implementation, the connection relationships between nodes are determined based on the transmission path of the monitoring data. This transmission path reflects the actual flow direction of data within the sensor network. The system needs to analyze historical data or rely on a preset network topology configuration to define the connection relationships between nodes. These connections are typically represented as directed data flows from a source node to a target node. For example, data might flow from sensor nodes located at bridge supports to sensor nodes located in the middle of the main beam span, or from edge data acquisition nodes to regional data aggregation nodes. After determining the connection relationships, the system creates a directed edge for each pair of source and target nodes with a data flow direction, pointing from the source node to the target node. Data flow is used as the edge weight, reflecting the intensity of data interaction between nodes. The system needs to quantify the actual amount of data transmitted over the defined connection relationships within each time segment. This data volume can be quantified as bytes, data packets, or valid data records. The system accumulates the data transmission volume between node pairs within each time segment and maintains a data counter for each pair of source and target nodes with a connection relationship. When processing spatiotemporally correlated datasets, for each data record, if the system determines that the data in that record logically originates from the source node and is transmitted to the target node, the system increments the data counter for the corresponding node pair based on the data size or unit count of that record. At the end of each time segment, the accumulated value in the data counter represents the amount of data transferred between this pair of nodes within that time segment. The system calculates a standardized flow value to eliminate differences in absolute values, making the edge weights between different node pairs comparable. The system standardizes the accumulated data transfer volume, for example, by using a min-max normalization method to find the maximum and minimum data transfer volumes for all node pairs in the entire network within that time segment, and then linearly transforms the data transfer volume of each node pair to the interval between zero and one; alternatively, the system can calculate the proportion of each node pair's data transfer volume to the total data transfer volume of the entire network as the standardized flow value. The standardized flow value is assigned as the weight attribute of the corresponding edge. When creating or updating a directed edge from the source node to the target node, the system stores the calculated standardized flow value as an attribute field of that edge. It is understandable that the edge weight attribute enables the dynamic health status graph to not only reflect the connection relationship between nodes, but also dynamically reflect the connection strength. The larger the weight of the edge, the more frequent the data interaction, which usually means that the correlation between the two positions may be stronger in structural analysis.

[0062] In practical implementation, a graph database is used to store node and edge information. A graph database is a database management system specifically designed for processing graph-structured data. The system persistently stores all created graph node objects and directed edge objects with weight attributes in the graph database. The graph database uses node identifiers as primary keys to store the attribute information of each node, and source node identifiers and target node identifiers as foreign keys to store the relationship and weight attributes of edges. A dynamic health status graph with a weighted directed graph structure is constructed. The construction of the dynamic health status graph is a continuously updated process. After the system processes a spatiotemporal correlation dataset for a time segment, it updates the node attributes and edge weights in the graph database according to the data of that time segment, so that the dynamic health status graph can reflect the latest monitoring data flow patterns and spatial correlation status. The final constructed dynamic health status graph is a weighted directed graph containing a set of nodes, a set of edges, and edge weight attributes. Optionally, the system can add timestamp tags to graph nodes to record the time of the last update of node attributes. Optionally, the system can periodically optimize the graph structure, such as merging redundant nodes or pruning inactive edges, to maintain the efficiency of the graph. Optionally, the dynamic health status graph can be exposed as a graphical query interface, allowing other analysis modules to query information such as the neighbors and paths of nodes through graph traversal algorithms.

[0063] In practical implementation, the historical flow patterns of sensor data in the spatiotemporal correlation dataset are analyzed to identify source and target node pairs. These historical flow patterns reveal stable data transmission paths within the sensor network. The system queries the historical records of the spatiotemporal correlation dataset to analyze the source and destination information of each data record. Source information is typically indicated by sensor identifiers, while destination information may be implicit in the data record type or processing logic. Through statistical analysis, the system identifies sensor node combinations that frequently co-occur in historical data and exhibit stable data transmission relationships; these combinations are identified as potential source and target node pairs. The system accumulates the data transfer volume between node pairs within each time segment, maintaining a data transfer volume accumulator for each identified source and target node pair. When processing new time segment data, for each data record in the spatiotemporal correlation dataset, the system determines which source and target node pair it belongs to based on its sensor identifier and predefined flow logic. Then, it increments the corresponding node pair's data transfer volume accumulator based on the record's data size (e.g., number of bytes) or simply by the number of records. A normalized flow value is calculated to convert absolute data volume into relative weights. The system employs a min-max normalization method. First, it finds the maximum and minimum data transfer rates among all node pairs within the current time segment. Then, for each node pair, its data transfer rate is substituted into the normalization formula for calculation, ensuring that the final standardized flow rate falls within the range of zero to one. The standardized flow rate is assigned as the weight attribute of the corresponding edge. When constructing or updating the dynamic health status graph, the system creates or updates a directed edge for each pair of connected source and target nodes, setting the calculated standardized flow rate as the weight attribute value of this edge. In some embodiments, data flow pattern identification can utilize clustering algorithms from machine learning to automatically discover stable data flow patterns. In some embodiments, the accumulation of data transfer rate can consider a time decay factor, placing greater emphasis on recent data flow. It can be understood that the dynamic edge weight update mechanism based on data flow enables the dynamic health status graph to adaptively reflect changes in the spatial correlation of bridge monitoring data, providing dynamic topological information for health status assessment.

[0064] 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 method for monitoring bridge health, characterized in that, The method performs the following operations: The raw data stream of multimodal monitoring of the bridge is collected through a distributed sensor network; The original multimodal monitoring data stream is intelligently cleaned and heterogeneous data is integrated to generate a unified format bridge health data set; Decode the time series markers, spatial location codes, and monitoring parameter dimension identifiers from the bridge health dataset; The bridge health dataset is divided into overlapping time segments based on time series labels to generate a time segment dataset. By integrating spatial location coding, spatial information in time-segment datasets is mapped to a global coordinate framework for bridges to establish a dynamic health status map; Based on dynamic health status maps and time segment datasets, multi-dimensional data fusion is performed according to the dimension identifiers of monitoring parameters; By analyzing the state evolution path of bridge components through fusion results, health metrics and prediction sequences are generated. By integrating spatial location coding, spatial information in time-segment datasets is mapped to a global coordinate framework for bridges to establish a dynamic health status map. This includes: extracting spatial location codes from time-segment datasets and converting local coordinates into a global coordinate point set using a coordinate transformation algorithm; associating monitoring data within each time segment with the global coordinate point set based on sensor identifiers to generate a spatiotemporal correlated dataset; organizing the coordinate point set in the spatiotemporal correlated dataset into nodes and adding weighted edges based on the data flow direction to form a dynamic health status map. Multi-dimensional data fusion is performed according to the dimension identifier of the monitoring parameters, including: dividing the time segment dataset into multiple parameter subsets according to the dimension identifier of the monitoring parameters, and generating a three-dimensional data cube of parameters-nodes-time segments by combining the node information in the dynamic health status map; applying a feature extraction algorithm to each parameter subset in the three-dimensional data cube to calculate the statistical feature vector of the parameter on the node; and aggregating the statistical feature vectors into a multi-dimensional health feature matrix. The statistical feature vectors are aggregated into a multidimensional health feature matrix, including: normalizing the statistical feature vectors of each parameter subset to obtain standardized feature vectors; calculating the weight coefficients of the standardized feature vectors using an attention mechanism; and performing a weighted summation of the standardized feature vectors based on the weight coefficients to generate a multidimensional health feature matrix. The coordinate point set in the spatiotemporal correlation dataset is organized into nodes, and weighted edges are added based on the data flow direction to form a dynamic health status graph. This includes: mapping each global coordinate point in the spatiotemporal correlation dataset to a graph node and assigning a unique node identifier; determining the connection relationship between nodes according to the transmission path of the monitoring data and using data flow as the edge weight; using a graph database to store node and edge information and constructing a weighted directed graph structure for the dynamic health status graph. The connection relationship between nodes is determined based on the transmission path of the monitoring data, and the data flow is used as the edge weight. This includes: analyzing the historical flow pattern of sensor data in the spatiotemporal correlation dataset to identify source and target node pairs; accumulating the data transfer volume between node pairs in each time segment and calculating the standardized flow value; and assigning the standardized flow value as the weight attribute of the corresponding edge.

2. The bridge health monitoring method according to claim 1, characterized in that, The process involves intelligent data purification and heterogeneous data integration of the original multimodal monitoring data stream to generate a unified format bridge health data set. This includes: applying a deep learning-based noise filtering model to remove outliers and smooth signals in the original multimodal monitoring data stream to obtain a purified data stream; converting heterogeneous field names in the purified data stream into standardized data patterns to generate a standardized data stream; and compensating for and synchronizing time deviations in the standardized data stream to output a unified format bridge health data set.

3. The bridge health monitoring method according to claim 2, characterized in that, The process involves converting heterogeneous field names in the cleaned data stream into a standardized data pattern, generating a standardized data stream, including: parsing all field names in the cleaned data stream to generate a list of field names; mapping synonymous fields in the field name list to unified terms with reference to a pre-defined standardized dictionary to generate a mapped field set; and reorganizing the data structure of the mapped field set according to the standardized data pattern to output the standardized data stream.

4. The bridge health monitoring method according to claim 3, characterized in that, The time deviation in the standardized data stream is compensated and synchronized by a timestamp correction algorithm, and a bridge health data set in a unified format is output. This includes: extracting timestamp fields and sensor identification fields from the standardized data stream and combining them into a synchronization key; detecting and aligning timestamp offsets based on the synchronization key and a preset time window threshold; and using interpolation methods to fill in data with missing time points to generate a time-synchronized dataset as a bridge health data set in a unified format.

5. A bridge health monitoring method according to claim 4, characterized in that, To fill in missing time points in the data using interpolation methods, a time-synchronized dataset is generated. This includes: identifying gaps in the timestamp sequence; calculating missing values ​​using a linear interpolation algorithm based on the trend of data points before and after the gaps; and inserting the interpolation results into a standardized data stream to form a continuous time-series time-synchronized dataset.

6. A bridge health monitoring system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the bridge health monitoring method according to any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the various steps of the bridge health monitoring method as described in any one of claims 1 to 5.

Images