Engineering observation data digital twin visual analysis system

Abstract

The application relates to the technical field of engineering data visualization, and discloses an engineering observation data digital twin visualization analysis system. The system comprises a data fusion module, which collects multi-source historical observation data of an engineering structure and performs effectiveness verification to generate a structure historical observation data set; a feature extraction module, which performs multi-dimensional feature extraction on the data set to form a structure topology feature set; a twin space construction module, which establishes a digital twin topology space according to the structure topology feature set and calculates the feature distribution density of historical abnormal events in the space; a real-time mapping module, which obtains real-time monitoring data and extracts real-time topology feature vectors and maps them to the digital twin topology space; and an abnormality prediction module, which generates a structure abnormality probability prediction result according to the spatial correlation between the real-time topology feature vector mapping position and the historical abnormality feature distribution density. The system improves the comprehensiveness and accuracy of engineering observation analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of engineering data visualization, in particular to an engineering observation data digital twin visualization analysis system. BACKGROUND

[0002] In the process of engineering construction and operation, it is crucial to accurately observe and analyze the safety state of engineering structures. With the increasing complexity of engineering structures, traditional engineering observation methods have gradually exposed many limitations. Early engineering observation relies on a single type of sensor data, such as strain sensors or displacement sensors, which can only capture local characteristics of the structure and easily lead to misjudgment of the overall safety state of the structure.

[0003] At the same time, there are also deficiencies in the processing of historical observation data. Traditional methods lack a systematic effectiveness verification process after collecting historical data, and some abnormal or error-prone data directly enters the analysis process, reducing the reliability of conclusions based on these data. In the case of growing data volume, it is difficult to extract valuable information from massive historical data, and traditional manual screening and simple statistical methods are inefficient and difficult to meet the timeliness requirements of engineering observation.

[0004] The application of digital twin technology in the engineering field provides a new way to solve the above problems, but the existing digital twin system still has defects in the construction of topological space. Most of the digital twin space construction systems fail to fully combine the actual topological characteristics of engineering structures, resulting in inaccurate mapping relationship between digital space and physical structure. In addition, the expression of historical abnormal events in the digital twin space is not clear enough, and the effective mining of abnormal feature distribution is lacking, so the reference value of historical data cannot be fully utilized.

[0005] In real-time monitoring, the processing of real-time data by traditional systems is often lagging behind, and the mapping of real-time monitoring data and digital twin space lacks immediacy, making it difficult to dynamically track the state of engineering structures. In the abnormal prediction process, existing methods rely on empirical formulas or simple threshold judgments, and do not fully consider the internal relationship between real-time data and historical abnormal features, resulting in low accuracy of abnormal prediction and easy occurrence of false positives or false negatives, which poses potential risks to the safe operation of engineering structures.

[0006] With the continuous expansion of engineering scale and the increasing demand for safety, there is an urgent need for an engineering observation analysis system that can integrate multi-source data, accurately construct a digital twin space, realize real-time mapping and intelligent abnormal prediction, to overcome the shortcomings of existing technologies. SUMMARY

[0007] The present application aims to provide an engineering observation data digital twin visual analysis system to solve the problems raised in the background.

[0008] To achieve the above-mentioned purpose, the present application provides an engineering observation data digital twin visual analysis system, which comprises:

[0009] A data fusion module is used to collect multi-source historical observation data of an engineering structure and perform validity verification, and generate a structure historical observation data set;

[0010] A feature extraction module is used to perform multi-dimensional feature extraction processing on the structure historical observation data set, and form a structure topology feature set;

[0011] A twin space construction module is used to establish a digital twin topology space according to the structure topology feature set, and calculate the feature distribution density of historical abnormal events in the digital twin topology space;

[0012] A real-time mapping module is used to obtain real-time monitoring data of an engineering structure and extract real-time topology feature vectors, and map the real-time topology feature vectors to the digital twin topology space;

[0013] An abnormality prediction module is used to generate a structure abnormality probability prediction result according to the mapping position of the real-time topology feature vectors in the digital twin topology space and the spatial correlation of the historical abnormal feature distribution density.

[0014] Preferably, the data fusion module comprises:

[0015] A multi-source acquisition unit is used to synchronously acquire sensor time series data, three-dimensional point cloud data and environmental parameter data of an engineering structure;

[0016] An effectiveness verification unit is used to perform abnormal data filtering according to the fluctuation threshold and change continuity of each type of data;

[0017] A data set generation unit is used to integrate the verified multi-source historical observation data into the structure historical observation data set according to the spatio-temporal correlation.

[0018] Preferably, the feature extraction module comprises:

[0019] A topology feature extraction unit is used to extract structure deformation gradient, stress conduction path and material fatigue features from the structure historical observation data set;

[0020] A feature set construction unit is used to classify the extracted topology features according to spatial dimensions, and form the structure topology feature set containing time series;

[0021] The feature matrix generating unit is configured to calculate an evolution track of each feature in the structure topology feature set within a preset time window, and generate a space-time evolution feature matrix.

[0022] Preferably, the twin space construction module comprises:

[0023] The space dimension determining unit is configured to construct the digital twin topology space according to a feature dimension number of the space-time evolution feature matrix.

[0024] The density calculating unit is configured to calculate a gathering density of feature coordinate points corresponding to historical abnormal events in the digital twin topology space.

[0025] The early warning index generating unit is configured to set a density threshold according to the feature distribution density, and screen topology features corresponding to feature coordinate points exceeding the density threshold as an abnormal early warning index set.

[0026] Preferably, the real-time mapping module comprises:

[0027] The real-time feature generating unit is configured to extract a current structure deformation gradient, a stress conduction path and a material fatigue feature from the real-time monitoring data, and generate a real-time topology feature vector consistent with a dimension of the structure topology feature set.

[0028] The space mapping unit is configured to project the real-time topology feature vector to the digital twin topology space.

[0029] The correlation calculating unit is configured to calculate a spatial proximity between a projection position of the real-time topology feature vector and each feature coordinate point in the abnormal early warning index set.

[0030] Preferably, the system further comprises:

[0031] The risk region marking module is configured to identify a structure defect distribution region according to a material microscopic image, and generate a risk marking region in combination with stress concentration region analysis.

[0032] The interactive evaluation module is configured to analyze a coupling relationship between a material grain boundary slip feature and a stress distribution in the risk marking region.

[0033] Preferably, the system further comprises:

[0034] The multi-precision division module is configured to divide a monitoring region into a macroscopic deformation region and a microscopic defect region according to an engineering scene knowledge graph.

[0035] The model coordination module is configured to process a global deformation trend of the macroscopic deformation region by using a first analysis model, and process a local damage feature of the microscopic defect region by using a second analysis model.

[0036] Preferably, the system further comprises:

[0037] A dynamic peeling module is configured to split a to-be-verified sub-region from the macro-deformation region when a topological distortion gradient of the macro-deformation region exceeds a dynamic threshold.

[0038] A feedback optimization module is configured to input a preliminary analysis result of the to-be-verified sub-region by the first analysis model into the second analysis model, and correct parameters of the first analysis model according to a refined output of the second analysis model.

[0039] Preferably, the anomaly prediction module comprises:

[0040] A correlation factor calculation unit is configured to generate a real-time anomaly correlation factor based on the spatial proximity and the feature distribution density;

[0041] A prediction model unit is configured to input the real-time anomaly correlation factor and the spatio-temporal evolution feature matrix into a probability prediction network;

[0042] A decision output unit is configured to generate a multi-precision model collaborative analysis result according to an output of the probability prediction network.

[0043] Preferably, the system further comprises:

[0044] A visual decision module is configured to dynamically superimpose and render the multi-precision model collaborative analysis result and a digital twin topological space to generate an engineering structure maintenance decision suggestion.

[0045] Compared with the prior art, the present application has the following advantages:

[0046] The engineering observation data digital twin visualization analysis system exhibits multiple advantages in the field of engineering observation data processing and analysis through the synergistic effect of various modules. The data fusion module focuses on the collection and effectiveness verification of multi-source historical observation data, changing the limitations of traditional single data sources. The multi-source data covers information of different types of sensors, different monitoring periods and different monitoring positions, and after effectiveness verification, invalid or excessively large error data are eliminated, so that the generated structure historical observation data set is more comprehensive and reliable, and can more completely reflect the state characteristics of the engineering structure at different stages, providing a solid data foundation for subsequent analysis.

[0047] The multi-dimensional feature extraction processing of the feature extraction module breaks through the problem of single dimension in traditional feature extraction. The state of the engineering structure is affected by multiple factors, and single-dimensional features are difficult to capture the complex changes of the structure. Multi-dimensional feature extraction extracts key features from multiple angles such as the geometric shape, mechanical properties and environmental response of the structure, forming a structure topology feature set that can more deeply depict the internal properties of the engineering structure, allowing the structure features hidden in the data to be fully displayed, and providing accurate feature basis for the construction of the digital twin space.

[0048] The digital twin topology space is established according to the structure topology feature set, so that the digital space and the physical properties of the engineering structure are accurately corresponded. This correspondence is not a simple geometric simulation, but a deep mapping based on the structure topology features, which can truly reflect the connection relationship and interaction between different parts of the structure. At the same time, the calculation of the feature distribution density of historical abnormal events in the space allows the feature mode of historical abnormalities to be clearly presented, and these distribution rules can intuitively reflect the occurrence position and influence range of different abnormal types in the structure, providing a reference historical feature benchmark for abnormal identification.

[0049] The real-time mapping module maps the real-time topology feature vector extracted from the real-time monitoring data to the digital twin topology space, realizing the immediate linkage between the real-time state of the engineering structure and the digital space. Real-time data is no longer isolated values, but is converted into specific positions in the digital space through feature vectors, so that engineering personnel can intuitively observe the real-time changes of the structure in the digital twin environment. This real-time mapping breaks the problem of disconnection between data and structure state in traditional monitoring, allowing the value of real-time monitoring data to be immediately realized, and facilitating quick grasp of the current operating state of the structure.

[0050] The abnormal prediction module generates prediction results based on the mapping position of the real-time topology feature vector in the digital twin topology space and the spatial correlation of the historical abnormal feature distribution density, changing the traditional way of relying on single threshold or experience judgment for abnormal prediction. Spatial correlation analysis can comprehensively consider the similarity and correlation of real-time features and historical abnormal features in spatial distribution, and determine whether the real-time state is close to the historical abnormal mode through this correlation, so that the generated structure abnormal probability prediction results can better reflect the actual risk condition of the structure, helping to detect potential risks in advance, and providing more intelligent analysis support for the safe operation and maintenance of engineering structures. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 The timing diagram of the engineering observation data digital twin visualization analysis system described in the present application;

[0052] Figure 2 The working principle diagram of the data fusion module;

[0053] Figure 3 a working principle diagram of a twin space construction module;

[0054] Figure 4 a working principle diagram of a real-time mapping module. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0056] Please refer to Figure 1 The present application provides an engineering observation data digital twin visualization analysis system, which comprises:

[0057] The data fusion module collects multi-source historical observation data of the engineering structure and performs effectiveness verification to generate a structure historical observation data set; the feature extraction module performs multi-dimensional feature extraction processing on the structure historical observation data set to form a structure topological feature set; the twin space construction module establishes a digital twin topological space according to the structure topological feature set, and calculates the feature distribution density of historical abnormal events in the digital twin topological space; the real-time mapping module obtains real-time monitoring data of the engineering structure and extracts a real-time topological feature vector, and maps the real-time topological feature vector to the digital twin topological space; and the abnormal prediction module generates a structure abnormal probability prediction result according to the spatial correlation between the mapping position of the real-time topological feature vector in the digital twin topological space and the historical abnormal feature distribution density.

[0058] Embodiment 1: Please refer to Figure 2 The data fusion module performs systematic collection and processing of multi-source historical observation data. The multi-source collection unit synchronously obtains three types of observation data of the engineering structure: sensor time series data are continuously captured by a distributed sensor network to respond to the structure dynamics, network deployment covers key nodes and stress concentration areas of the structure, and physical quantities include structure vibration frequency response spectrum, displacement amplitude time series, etc., and the collection process follows a fixed sampling period; three-dimensional point cloud data are non-contact measured by a laser scanning device with a preset accuracy, the scanning device moves according to a planned path to cover the geometric shape of the structure surface, and a spatial coordinate matrix of the structure surface is obtained each time scanning; environmental parameter data are synchronously recorded by a meteorological station and a load monitoring terminal, the parameter types include air temperature and humidity, average wind speed and peak wind speed, and structure bearing load change value, and data transmission adopts a time stamp synchronization mechanism.

[0059] The effectiveness verification unit performs data quality screening based on a preset determination rule. For sensor time series data, a dynamic fluctuation threshold is used for outlier filtering: the standard deviation of the physical quantity value in a fixed time window is calculated, and when the data point deviates from the window mean value by more than a fixed multiple of the standard deviation, it is determined to be an outlier, triggering data rejection operation; for three-dimensional point cloud data, a spatial continuity test algorithm is applied: the point clouds of the common area of the adjacent two scans are compared, the spatial offset error and coincidence degree index of the corresponding point clouds are calculated, and if the index is lower than the preset threshold, the automatic resampling mechanism is triggered, and the scanning path and exposure parameters are optimized during resampling; environmental parameter data is processed by time correlation analysis method: an autoregressive model of each environmental parameter is established, and when the continuous model estimation deviation exceeds the confidence interval, the abnormal data segment is marked, and the invalid data interval is covered by using linear interpolation algorithm.

[0060] The data set generation unit implements data integration in space-time dimension. A unified space-time label system is established: all data sources are labeled with millisecond time stamps by GPS clock, and the spatial coordinate system is converted with the structural reference control point as the origin, and the least squares adjustment method is used for coordinate conversion to realize precision control. Missing data compensation is performed: for data gaps caused by equipment failure, the time and space correlation interpolation method is used to fill in the missing values, the adjacent spatial grid point data is calculated by distance weighted average, and the time series missing segment is filled by ARIMA model prediction. A three-dimensional structured data set is constructed: the time dimension is divided into storage units at fixed intervals, the spatial dimension is indexed by grid encoding, and the data attribute layer is classified to store physical quantity values, environmental parameters and point cloud coordinate matrix. The data set is stored in a distributed database cluster in columnar structure, the database is horizontally fragmented by time range, and the spatial grid is indexed by B-tree to optimize query efficiency.

[0061] The multi-source data acquisition link includes three types of data acquisition processes. Distributed architecture is used for sensor time series data acquisition: strain sensor arrays are deployed at key nodes of the engineering structure, vibration response is obtained through acceleration sensors, all sensor signals are connected to the edge computing gateway after analog-to-digital conversion, and the gateway has a built-in time synchronization protocol to ensure consistency of acquisition time. An automatic scanning scheme is used for three-dimensional point cloud data acquisition: the laser scanner moves at a preset spatial sampling rate, a high-resolution panoramic image is automatically taken after each scanning plane to assist point cloud matching, and the scanning raw data includes three-dimensional position data and reflectivity data in the Cartesian coordinate system. Environmental parameter data is collected by networking devices: temperature and humidity sensors are distributed according to the structure area, wind speed monitors are set up at multiple measurement points on the windward surface of the structure, load data is fed back to the central receiver in real time through pressure sensor nodes, and industrial bus is used for data transmission to ensure low delay.

[0062] The data cleaning process performs a multi-directional verification mechanism. The sensor time series data anomaly detection uses a double-level decision logic: the first level calculates the rate of change of physical quantities in real time, and an alarm is triggered when the rate of change exceeds the theoretical limit of material elastic deformation; the second level analyzes the distribution characteristics of long-period data, and locates the statistically significant abnormal interval through kernel density estimation. The point cloud data quality verification introduces geometric topology verification: calculate the dihedral angle change of adjacent triangles in the point cloud model, and when there is a geometric mutation area that exceeds the maximum deformation angle of the material, start the special area re-scanning, which increases the laser energy density and the number of sampling points. The environmental parameter analysis uses a cross-validation strategy: load data and wind speed data are used to construct a linear regression model, and temperature and humidity parameters are used to establish a correlation matrix with structural strain data. When a single data deviates from the model prediction value by more than a threshold value, the backup device data verification process is started.

[0063] The data set construction uses a hierarchical storage model. The spatial coordinate alignment implementation scheme is to establish a structure control point spatial index network, each spatial grid is assigned a unique code, and the coordinate data of various types of data in the grid is converted to a local rectangular coordinate system. The time dimension processing includes automatic time zone conversion and daylight saving time compensation, and the millisecond level timestamp is stored using UTC standard time. Physical quantity data is stored in time series blocks, each block contains a measurement value matrix for a fixed time period, the matrix row corresponds to the spatial grid number, and the column corresponds to the time point sequence number. Point cloud data is stored using an octree spatial partitioning structure, and each leaf node stores the point cloud slice of the corresponding spatial voxel. Environmental parameters are independently established in a partition set, and real-time association queries with other data are realized through a timestamp association engine. The database cluster is set up with a master-slave replication architecture, the master node handles data write requests, and the slave node is responsible for analysis and query tasks, and the data synchronization uses an incremental log synchronization mechanism.

[0064] The data maintenance system implements full life cycle management. The storage node sets an automatic backup period, and the hot and cold data separation storage strategy stores historical observation data in different media according to the time dimension. The data access interface realizes two access modes: batch processing mode supports large-scale spatio-temporal data query, and stream processing mode provides real-time monitoring data subscription service. The data audit module records all operation logs, supports data version backtracking and integrity verification at any time point. The system deployment data quality monitoring board dynamically displays the collection completeness, spatial coverage and time series continuity indicators of the three types of data, and triggers maintenance work orders to the on-site device management terminal when the indicators are abnormal.

[0065] The above implementation process constructs a structural historical observation data set, providing high-quality basic data in a unified format for subsequent processing of the system. The entire process includes five standard links: data acquisition, quality verification, spatial calibration, abnormal compensation, and structured storage. Each link is configured with a dedicated algorithm unit and resource management strategy. The system can adapt to the observation needs of structures of different scales by adjusting the spatial grid size and time sampling interval to balance data accuracy and storage cost.

[0066] Example 2: Refer to Figure 3 The feature extraction module processes the structural historical observation data set and performs multi-dimensional topological feature extraction. The topological feature extraction unit analyzes three core features: structural deformation gradient calculation based on the displacement observation values of adjacent measuring points in space, obtaining the gradient value through the proportional relationship between the displacement difference and the reference length; stress conduction path analysis relies on the finite element grid model, and according to the stress value distribution of each grid node, the path optimization algorithm is used to determine the main stress transmission direction, and the path output includes node sequence and path weight; material fatigue feature evaluation uses strain energy accumulation model, combined with load cycle number record and material fatigue characteristic curve, calculates micro-damage accumulation factor.

[0067] The feature set construction unit implements spatial dimension classification and time sequence integration of features. Macro-scale features cover overall structural response indicators, including structural overall deformation rate calculation value and main stress direction angle change trajectory; micro-scale features focus on local structure state, extracting local curvature change and micro-crack density distribution value. All features are attached with timestamp sequence, which is strictly synchronized with the original observation data. Feature storage uses a hierarchical structure: the first layer divides features into large categories according to spatial scale, the second layer divides sub-categories according to physical properties, and the third layer stores continuous feature value sequence with timestamp.

[0068] The feature matrix generation unit calculates the evolution trajectory of features based on the preset time window mechanism. The time window length is set to a fixed period, and the sliding window method is used for processing. The structural deformation gradient evolution trajectory calculates the change trend of feature values within the window, and obtains the gradient change slope through linear regression analysis; the stress conduction path evolution analysis extracts the spatial displacement vector of the path inflection point, records the migration distance and direction of the inflection point within the window period; the material fatigue feature statistics the damage factor increment within the window period, and calculates the proportional relationship between the increment and the load cycle number. The evolution trajectory output uses a tensor data structure to construct a three-dimensional spatio-temporal evolution feature matrix: the matrix row index corresponds to the spatial grid code, the column index corresponds to the time window number, and the depth dimension stores the quantitative values of different features' evolution trajectories. The matrix elements contain a two-element group of feature type identifier and evolution value.

[0069] The twin space construction module constructs an analysis space according to the space-time evolution characteristic matrix. The space dimension determination unit counts the number of effective dimensions of the characteristic matrix, and establishes a digital twin topology space corresponding to the dimensions. The space coordinate system is defined by orthogonal basis vectors, and each historical observation time point is represented as a characteristic vector point in the space, with the vector component value being the quantitative value of the characteristic evolution trajectory of the time period. The space initialization process performs characteristic vector standardization to eliminate the dimensional differences of different characteristics.

[0070] The density calculation unit performs historical abnormal event distribution analysis in the digital twin topology space. The historical abnormal event library stores the records of abnormal time periods verified by engineering, and each abnormal event is associated with the characteristic vector point coordinates of its occurrence period. A spatial clustering method is used to establish an analysis area centered on each abnormal event coordinate, and the number of characteristic vector points in the unit volume of the area is calculated. The shape of the analysis area adopts a hypersphere structure, and the radius parameter is adaptively adjusted according to the dispersion degree of the characteristic space distribution. The clustering degree calculation result is recorded as the distribution density value of each abnormal event coordinate, and the density value is mapped to a spatial position function.

[0071] The early warning index generation unit sets a dynamic judgment threshold based on the characteristic distribution density. The threshold is set by referring to the density distribution characteristics of the historical normal operation state, and a certain multiple of the density peak value of the normal operation state is taken as the baseline. The space scanning process identifies all density areas that exceed the baseline, and merges adjacent high-density areas using a spatial clustering algorithm. For each clustering area, the core coordinate point is extracted, the corresponding characteristic vector group is backtracked, and the high-frequency appearing topological feature type is identified. The output result constructs an abnormal early warning index set, including spatial coordinates, density values, related feature types and historical abnormal event association records. The index set is stored as a spatial index structure, supporting fast range queries.

[0072] The feature space construction process implements dimension reduction visualization support. While maintaining the analysis accuracy of the original high-dimensional space, a three-dimensional projection subspace is created for visualization. The projection conversion uses a feature preservation algorithm, which selects the main dimensions with high abnormal distribution discrimination contribution for coordinate axis mapping. The distribution of historical abnormal event coordinate points, density contour surfaces, and early warning index set positions are displayed simultaneously in the projection space. The space interaction interface supports rotation and scaling operations, allowing users to view the feature distribution state in any cross section.

[0073] A version management mechanism is established for the digital twin topology space. When new historical observation data is added, the space incremental update process is triggered: the new characteristic vector point is imported into the space coordinate system, and the distribution density values of the affected areas are recalculated; when the density changes exceed the threshold, the early warning index set is dynamically updated. The space backup system retains historical version snapshots, supporting comparative analysis of space states at different periods. The space data uses a distributed storage architecture, and the space coordinate index is built in a memory database to realize fast space queries.

[0074] The system runs, feature extraction and space construction form a closed loop processing flow. The feature matrix generation unit outputs the evolution features of the new time window regularly, triggering the incremental update operation of the twin space construction module. The space density distribution calculation result is fed back to the feature extraction module to optimize the feature selection strategy. When a specific type of feature appears frequently in the early warning indicator set, the feature extraction module will improve the sampling accuracy and analysis frequency of this type of feature.

[0075] The entire implementation process constructs a dynamically evolving feature analysis space, which converts the historical state of the engineering structure into a quantifiable spatial distribution model. The feature extraction link realizes the transformation from raw data to abstract features, and the space construction link establishes the mapping relationship between features and engineering anomalies. The system keeps the digital twin topological space in synchronization with the evolution of the entity structure through continuous space state updating, providing a historical state reference benchmark for real-time monitoring and analysis.

[0076] Example 3: refer to Figure 4 The real-time mapping module periodically processes the real-time monitoring data stream of the engineering structure. The real-time feature generation unit obtains the latest readings of the sensor network through the edge computing gateway, and the data sampling interval uses a pre-set fixed period. The unit performs three types of feature extraction operations: the current structure deformation gradient calculation is based on the instantaneous displacement difference of adjacent monitoring points, combined with the structure reference size for dimensionless processing; the real-time stress conduction path update relies on the dynamic finite element model, and the latest stress measurement value is used to recalculate the stress transmission path between nodes, and the main path node sequence is output; the material fatigue feature evaluation calls the online damage accumulation model, inputs the real-time data stream of the strain sensor and the load change count, and outputs the current damage factor increment value. The generated feature vector dimension is strictly consistent with the historical structure topological feature set, and each feature component is converted through standardization to eliminate dimension differences.

[0077] The space mapping unit projects the real-time topological feature vector into the digital twin topological space. The projection process uses a pre-trained principal component analysis model for dimension reduction: the model parameters are trained and fixed in the twin space construction stage; after the real-time feature vector is input into the model, the three-dimensional space coordinates are output. The projection algorithm retains the main variation direction of the original space, and the projection error is controlled within the pre-set acceptable range. The projection result includes coordinate points and confidence evaluation values, and the confidence reflects the information retention ratio of the feature vector in the dimension reduction process.

[0078] The correlation calculation unit analyzes the spatial relationship between the real-time projection point and the historical anomaly early warning indicator set. The calculation uses a distance measurement method with time decay effect:

[0079]

[0080] Where: d M represents the spatial proximity quantitative value; xreal is the projection coordinate of real-time feature vector; y hist is the coordinate of specific historical point in abnormal early warning index set;∑ -1 is the inverse matrix of feature space covariance matrix; ω t is the time decay factor (the farther the historical point is from the current time, the smaller the factor value is)

[0081] The system establishes an independent time decay function for each historical abnormal point, and the function parameters are dynamically adjusted according to the historical recurrence period of the abnormal event. Finally, the spatial proximity list of each early warning point is output, and the high-risk associated objects are identified by value sorting.

[0082] The risk area marking module processes the material micrograph data in parallel. The electron microscope collects high-resolution metallographic images of the key area at a preset period, and the images are transmitted to the image processing server. Defect recognition uses a deep neural network architecture: the network input is the original grayscale image, the crystal boundary profile is reconstructed through the convolution layer to extract the grain boundary features and the up-sampling layer, and the binary mask containing micro-cracks, holes and other defects is output. The recognition result is post-processed by morphology to eliminate noise and generate a defect distribution heat map.

[0083] The stress-defect coupling analysis engine processes two inputs: real-time stress cloud data from the finite element analysis system, containing the structure surface stress distribution matrix; and defect distribution heat map containing defect position and density information. After spatial coordinate alignment, the superposition algorithm calculates the stress concentration coefficient and defect density product in each spatial unit. When the product value exceeds the preset critical threshold value in the material property library, the unit is marked as a risk area. The marking result is rendered and displayed in the three-dimensional structure model: the risk boundary is depicted with dynamic contour lines, the inside is filled with semi-transparent warning color, and the risk level is expressed through the color saturation gradient.

[0084] The interactive evaluation module performs coupling analysis of material micro-characteristics and macro-stress. In the marked risk area, extract the material characteristic data of each analysis point: electron backscattering diffraction data provides the lattice orientation matrix; X-ray diffraction data provides the residual stress value. Establish a grain boundary slip evaluation model: based on the crystal plasticity theory, calculate the Schmidt factor of each grain slip system, and count the proportion of grain boundaries whose factor exceeds the activation threshold. The coupling analysis process includes three steps: first, establish the conversion matrix of stress tensor and lattice orientation; second, calculate the decomposed shear stress on each slip system; finally, predict the grain boundary slip accumulation through the slip strain integral model.

[0085] The stress gradient and slip amount correlation modeling adopts a multivariate regression method. The independent variables include the stress gradient modulus of each point in the risk area, the principal stress direction change rate, and the grain size distribution variance; the dependent variable is the grain boundary slip amount measured at the corresponding position. The model training uses a historical observation data set, and the online update uses the recursive least squares algorithm. The model output is a slip amount prediction curve, which is superimposed with the stress distribution cloud chart, and the deviation between the two is visualized through the contour line offset.

[0086] Real-time monitoring and risk analysis form a closed-loop control. The edge computing layer deploys a lightweight risk prediction model: when the real-time feature vector projection point enters the high-density area of the early warning index set, the microscope focusing scanning instruction is automatically triggered; when the risk marking area area expansion rate exceeds the threshold, the grid resolution of the stress analysis model is improved. All analysis results are written into a distributed database, with strict timestamp alignment with the original monitoring data, supporting full-cycle state backtracking analysis. The system is configured with a visual control console, dynamically displaying the real-time projection point trajectory, risk area evolution process, and coupling analysis parameter change curve, and the operator can manually adjust the analysis area focus.

[0087] The data processing pipeline implements resource optimization strategies. Computing tasks are scheduled according to their timeliness: real-time feature extraction tasks are deployed on edge nodes for priority execution; spatial projection and correlation calculation are allocated to cloud computing platforms; risk marking and coupling analysis run on high-performance computing clusters. Data transmission uses a differential compression algorithm, and when network bandwidth is insufficient, the microscope image resolution is automatically reduced. Analysis results are cached in time series, and when the system detects similar historical working conditions, cached results are directly called to speed up the response.

[0088] Embodiment 4: The multi-precision division module defines spatial partitioning rules based on an engineering knowledge graph. The knowledge graph stores design parameters, material properties, and historical maintenance records of the engineering structure, including a topological relationship network. The macro-deformation region division standard is the region covered by the principal eigenvector of the structure stiffness matrix, which reflects the overall load transmission path; the micro-defect region is defined based on the material metallographic atlas, identifying the grain size analysis area, and the region boundary is determined according to the grain boundary distribution density. Spatial division uses a quadtree index structure, and the division process includes three levels of recursion: the first level is divided according to the structure function, the second level is divided according to the material uniformity, and the third level considers the sensor deployment density. The minimum partition size is set to a fixed value, and each partition records the stiffness characteristic value, the average grain size, and the monitoring equipment number.

[0089] The model coordination module configures a double-model processing pipeline. The first analysis model adopts a time series analysis architecture, with the input being a macro-region deformation rate dataset containing time series collected synchronously by multiple measuring points. The model structure includes a feature extraction layer and a trend prediction layer: the feature extraction layer calculates the statistics (mean, variance, autocorrelation coefficient) of the deformation rate within the window; the trend prediction layer generates the deformation trend envelope for the future period through a recurrent neural network unit, outputting the maximum deformation value and the location. The second analysis model adopts a graph neural network architecture, with the input being a topological adjacency graph of the micro-defect region: the nodes represent defect units (such as single micro-cracks), and the node attributes include defect size and orientation angle; the edges represent the spatial relationship between defects, and the edge weights are calculated according to the distance between defects. The network aggregates neighborhood information through the message passing mechanism, outputting a probability distribution map of the local damage propagation path.

[0090] The dynamic peeling module monitors the topological distortion gradient of the macro-region in real time. The distortion gradient calculation is based on the strain rosette sensor array data: three-direction strain values are collected, and the principal strain and its direction are solved through coordinate transformation. The trigger threshold is set to a fixed proportion of the material yield parameter. When the gradient of a certain sub-region continuously exceeds the limit, boundary segmentation is performed: taking the distortion peak point as the center, the principal distortion direction is determined by Hankel matrix eigenvalue decomposition; a circular sub-region to be verified is generated along the principal direction with a fixed radius, and the radius value is set according to the structural feature size. The segmentation results record the spatial coordinates of the region and the associated macro-model output values.

[0091] The feedback optimization module performs cross-model parameter correction. The deformation prediction value of the sub-region to be verified is extracted from the first analysis model output cache, including the deformation value prediction and the confidence interval. After inputting this data into the second analysis model, the model performs refined analysis: the grid resolution is improved to the grain scale; the electron microscope historical image of the region is loaded; the local stress concentration coefficient is calculated in combination with the real-time strain data. The output includes the micro-stress peak position and the defect propagation tendency score. The parameter correction process is as follows: compare the deformation prediction value of the first model with the stress concentration coefficient mapping value of the second model, calculate the correlation coefficient of the spatial distribution; when the correlation coefficient is lower than the set threshold, start the backpropagation algorithm: take the stress distribution of the second analysis model as the reference, adjust the convolution kernel weight of the feature extraction layer of the first analysis model, and focus on optimizing the kernel function sensitive to local distortion. Bridge structure application scenario: assume that the monitoring data of a certain steel box girder bridge triggers the dynamic peeling process. The multi-precision division module divides the bridge deck into 20 macro-regions (box girder units) and 35 micro-regions (weld heat affected zones). During the operation of the model coordination module, the first analysis model outputs the deformation trend in box girder unit B3, as shown in Table 1.

[0092] Table 1: Deformation trend output by the first analysis model in box girder unit B3 during the operation of the model coordination module.

[0093] Time window number Mean deformation rate (mm / h) Predicted deformation (mm) Confidence interval (± mm) T-120 0.15 1.82 0.21 T-119 0.18 2.15 0.24 T-118 0.23 2.61 0.29

[0094] When the distortion gradient detection value of the T-118 window exceeds the threshold value, the dynamic stripping module divides a circular verification area (covering the top plate weld of the box girder) with a diameter of 50 cm centered at the coordinates (102.3, 45.7). The second analysis model loads the area data: the microscopic image shows that the weld fusion line has a group of microcracks; the real-time strain data calculates the local stress concentration coefficient as 3.2. The feedback optimization module detects that the spatial correlation between the macro model prediction deformation of 2.61 mm and the microscopic stress distribution is lower than the threshold value, and accordingly adjusts the first analysis model: enhances the weight coefficient of the top plate measurement point data; increases the high-frequency vibration feature input channel. After optimization, the next cycle prediction value converges to 2.38 mm (confidence interval ± 0.18 mm), and the correlation with the microscopic analysis result is improved to the effective range.

[0095] The system maintains a real-time updating mechanism for the knowledge graph. After each model collaborative analysis, key parameters are recorded: macro / micro area correspondence table, model weight adjustment record, verification area size statistical value. These data generate new association rules through the graph reasoning engine, for example, when the weld area of a specific material (such as Q345 steel) frequently triggers verification, the macro partition size of this type of area is automatically reduced. Graph update triggers the re-partition operation of the multi-precision division module, forming a closed-loop optimization from data analysis to knowledge accumulation.

[0096] The computing resources implement a dynamic allocation strategy. Under normal conditions, the first analysis model is deployed on a cloud computing platform, and the second analysis model runs on an edge computing node. When the verification sub-area is generated, the system automatically promotes the computing priority of the area: the second analysis model task is migrated to a GPU accelerated cluster; the microscopic image processing task is allocated additional bandwidth resources. The resource scheduler dynamically adjusts the number of parallel computing nodes according to the number of verification areas to ensure that multi-precision analysis is completed within a limited time. All analysis processes retain version snapshots, supporting backtracking of model state and partition structure at any time point.

[0097] The implementation process builds a progressive analysis system from macro trend grasping to microscopic mechanism verification. Multi-precision division establishes a spatial analysis benchmark, a dual-model architecture realizes state assessment at different scales, a dynamic stripping mechanism captures local anomalies, and feedback optimization forms a collaborative correction between models. Through continuous knowledge accumulation and parameter adjustment, the system improves the abnormal response sensitivity and analysis accuracy of complex engineering structures.

[0098] Example 5: Abnormal prediction module handles fusion analysis of real-time abnormal correlation factors and historical feature data. The correlation factor calculation unit receives two inputs: the spatial proximity list of real-time projection points and historical abnormal points output by the spatial mapping module, and the feature distribution density dataset provided by the twin space construction module. The calculation process performs weighted fusion: the spatial proximity data is given a higher weight coefficient, and the feature distribution density data is given a lower weight coefficient, and the two are linearly superimposed to generate a scalarized real-time abnormal correlation factor. Each correlation factor corresponds to a specific historical abnormal type label, and the factor value size reflects the similarity between the current state and the pattern of this type of abnormality.

[0099] The prediction model unit constructs a three-layer processing architecture of the probability prediction network. The input layer receives heterogeneous data: real-time abnormal correlation factors are input into the first channel in vector form, and the spatio-temporal evolution feature matrix is input through the second channel, with the same matrix dimension as in the historical training phase. The hidden layer uses a nonlinear transformation function: performs kernel transformation of the feature dimension on the spatio-temporal feature matrix, and establishes the spatio-temporal correlation pattern through basis function expansion; the real-time abnormal correlation factor and the transformed feature tensor perform dot product operation to generate a spatio-temporal correlation feature vector. The output layer implements multi-classification mapping: the feature vector is input into the fully connected layer, and the normalized function is processed to output a probability distribution vector, and each element of the vector represents the probability value of a specific type of abnormality occurring in the future period.

[0100] The decision output unit integrates the analysis conclusions of the multi-precision model. The input sources include three parts: the abnormal probability distribution output by the probability prediction network, the structural overall trend prediction of the macroscopic deformation model, and the local defect evolution path of the microscopic damage model. The integration logic adopts a hierarchical decision rule: when the macroscopic model detects that the overall deformation exceeds the limit and the microscopic model identifies that the damage is accelerating and expanding, a high-risk warning is triggered regardless of the probability value; when the probability prediction value is consistently high but the double-model conclusion does not trigger the threshold, a confidence assessment process is started: perturbation samples are generated in the feature space using a random sampling method, and the stability distribution of the prediction results is counted. The final output includes a structured analysis report containing the abnormal type, occurrence probability, impact range, and confidence level.

[0101] The visual decision module realizes dynamic superposition rendering of multi-source data. The three-dimensional digital twin topology space is loaded as the rendering base, and the spatial coordinate system is aligned with the engineering structure real scene model. The abnormal probability distribution rendering uses heat map technology: the prediction probability value of the spatial grid point is mapped to a continuous color spectrum, and the area with a probability value higher than the set interval is displayed in warm colors, and the area with a probability value lower than the interval is displayed in cold colors. The rendering engine supports dynamic adjustment of transparency, and the high-probability area uses opaque rendering to highlight the warning effect. The risk marked area is displayed as a geometric superposition layer in synchronization: the polygon boundary line is updated in real time with the evolution of the microscopic defect, the internal filling has a pulsing flashing effect, and the flashing frequency is positively correlated with the damage expansion rate.

[0102] The maintenance decision suggestion generation adopts a rule-based reasoning engine. The input parameters include the anomaly prediction results, the risk area distribution, and the structure design parameter library. The rule base defines three types of maintenance instruction trigger conditions: the "immediate repair" instruction requires that the probability value exceeds the upper limit and the risk area area expansion rate exceeds the limit; the "enhanced monitoring" instruction is activated when the probability value continuously fluctuates in the middle or the macro / micro model conclusions conflict; the "normal operation and maintenance" instruction requires that all indicators are within the safety threshold. The instruction output is accompanied by a detailed parameter table: the repair instruction lists the high-risk component number and the detection priority; the monitoring instruction specifies the list of positions that need to be encrypted and sensor deployment.

[0103] The visualization interface implements multi-view collaborative display. The main window presents a three-dimensional rendering scene, and the left panel superimposes real-time monitoring data curves: the deformation rate curve and the macroscopic prediction envelope line are displayed side by side, and the damage factor curve is synchronized with the micro-analysis path. The right information panel displays decision information in layers: the top layer displays the current maintenance instruction icon, the middle layer scrolls to update the top three type labels of the abnormal probability value, and the bottom layer displays the risk area dynamic parameters (including position coordinates, area change rate, and defect density). User interaction functions support clicking on risk areas to retrieve material micro images and sliding time axes to play back historical state evolution processes.

[0104] The data processing pipeline implements time-sensitive hierarchical control. The probability prediction network runs on a high-performance computing cluster, with batch computation started at fixed intervals. The visualization rendering task is assigned to a graphics workstation, and the rendering frame rate is automatically adjusted according to the scene complexity: the basic refresh rate is maintained in normal state, and the rendering priority is automatically increased when abnormal probability mutation is detected. The decision rule engine uses an event-driven mechanism, and any input parameter update triggers rule re-evaluation. All output results are written to a time series database, with data packets containing timestamp, spatial coordinates, and analysis results triplets, supporting full life cycle backtracking analysis.

[0105] The system maintenance module implements self-optimization functions. Record the deviation data of each anomaly prediction and the actual situation, and establish the spatio-temporal distribution map of prediction error. When a specific area frequently appears prediction deviation, trigger feature extraction strategy adjustment: increase the input of new monitoring features in this area, and optimize the associated factor weight coefficient. The visualization configuration supports user-defined scheme saving: the parameters adjusted by the operator, such as the color scale mapping scheme and the risk area display style, can be stored as templates for quick calling in similar engineering scenarios.

[0106] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0107] While the embodiments of the application have been shown and described herein, it is to be understood that the scope of the application, jointly pointed out in the appended claims, is not to be limited to the above-described embodiments but can be otherwise variously changed, modified, replaced, and altered within the principles and spirit of the present application.

Claims

1. A digital twin visualization and analysis system for engineering observation data, characterized in that, include: The data fusion module is used to collect multi-source historical observation data of engineering structures and verify their validity, generating a structural historical observation dataset. The feature extraction module is used to perform multi-dimensional feature extraction processing on the historical observation dataset of the structure to form a structural topological feature set; The twin space construction module is used to establish a digital twin topological space based on the structural topological feature set, and to calculate the feature distribution density of historical abnormal events in the digital twin topological space. The real-time mapping module is used to acquire real-time monitoring data of the engineering structure and extract real-time topological feature vectors, and map the real-time topological feature vectors to the digital twin topological space. The anomaly prediction module is used to generate a structural anomaly probability prediction result based on the spatial correlation between the mapping position of the real-time topological feature vector in the digital twin topological space and the distribution density of historical anomaly features. The feature extraction module includes: The topology feature extraction unit is used to extract structural deformation gradient, stress transmission path and material fatigue features from the structural history observation dataset. The feature set construction unit is used to classify the extracted topological features according to the spatial dimension to form the structural topological feature set containing the time series. The feature matrix generation unit is used to calculate the evolution trajectory of each feature in the structural topology feature set within a preset time window and generate a spatiotemporal evolution feature matrix. The twin space construction module includes: The spatial dimension determination unit constructs the digital twin topological space based on the number of feature dimensions of the spatiotemporal evolution feature matrix; The density calculation unit is used to calculate the cluster density of feature coordinate points corresponding to historical anomalous events within the digital twin topological space. The early warning indicator generation unit sets a density threshold based on the feature distribution density and selects the topological features corresponding to feature coordinate points that exceed the density threshold as an abnormal early warning indicator set. The real-time mapping module includes: The real-time feature generation unit is used to extract the current structural deformation gradient, stress transmission path and material fatigue features from the real-time monitoring data, and generate the real-time topological feature vector that is consistent with the dimension of the structural topological feature set. A spatial mapping unit is used to project the real-time topological feature vector onto the digital twin topological space; The correlation calculation unit is used to calculate the spatial proximity between the projection position of the real-time topological feature vector and each feature coordinate point in the anomaly warning index set; The anomaly prediction module includes: A correlation factor calculation unit is used to generate real-time anomaly correlation factors based on the spatial proximity and the feature distribution density. A prediction model unit is used to input the real-time anomaly correlation factor and the spatiotemporal evolution feature matrix into a probability prediction network; The decision output unit is used to generate multi-precision model collaborative analysis results based on the output of the probability prediction network.

2. The digital twin visualization analysis system for engineering observation data according to claim 1, characterized in that, The data fusion module includes: The multi-source acquisition unit is used to simultaneously acquire sensor time-series data, 3D point cloud data, and environmental parameter data of the engineering structure; The validity verification unit is used to filter abnormal data based on the fluctuation threshold and the continuity of change of each type of data. The dataset generation unit is used to integrate the verified multi-source historical observation data into the structured historical observation dataset according to the spatiotemporal correlation.

3. The digital twin visualization analysis system for engineering observation data according to claim 1, characterized in that, Also includes: The risk area marking module is used to identify the distribution area of ​​structural defects based on the material's microscopic image and generate risk marking areas by combining stress concentration area analysis. The interactive evaluation module is used to analyze the coupling relationship between the grain boundary slip characteristics and stress distribution of the material within the risk-marked region.

4. The digital twin visualization analysis system for engineering observation data according to claim 3, characterized in that, Also includes: The multi-precision segmentation module is used to divide the monitoring area into macroscopic deformation areas and microscopic defect areas based on the engineering scenario knowledge graph. The model collaboration module is used to process the overall deformation trend of the macroscopic deformation region using a first analysis model and to process the local damage characteristics of the microscopic defect region using a second analysis model.

5. The digital twin visualization analysis system for engineering observation data according to claim 4, characterized in that, Also includes: The dynamic stripping module is used to segment out the sub-region to be verified from the macro-deformation region when the topological distortion gradient of the macro-deformation region exceeds a dynamic threshold. The feedback optimization module is used to input the preliminary analysis results of the first analysis model on the sub-region to be verified into the second analysis model, and to correct the parameters of the first analysis model according to the refined output of the second analysis model.

6. The digital twin visualization analysis system for engineering observation data according to claim 1, characterized in that, Also includes: The visualization decision module is used to dynamically overlay and render the results of the multi-precision model collaborative analysis with the digital twin topology space to generate engineering structure maintenance decision suggestions.

Images