Test risk digital twinborn early warning method based on multi-domain cooperative monitoring

Abstract

The invention provides a test risk digital twinning early warning method based on multi-domain cooperative monitoring, and belongs to the technical field of virtual-real fusion test and digital twinning, and the method comprises the steps: firstly, building a fine finite element simulation model which comprises a digital tool system, a digital sensor and a test piece and considers nonlinearity; secondly, performing nonlinear finite element simulation analysis, and constructing a multi-level mechanical response field inversion reduced-order model; thirdly, completing the construction of a complete sensor data set through a data filling algorithm, and carrying out the failure judgment of the first hierarchical structure based on the complete sensor data set; and finally, carrying out future loading level sensor data prediction and completing failure judgment of a second hierarchical structure. Carrying out full-field mechanical response inversion and online real-time correction; and performing response inversion of the region of interest to realize failure judgment of the third hierarchical structure. According to the invention, real-time dynamic monitoring and early warning of the structure test risk can be realized, the real-time performance, the robustness and the accuracy are high, and a powerful guarantee is provided for the safety and the reliability of the structure test.

Description

Technical Field

[0001] The present invention belongs to the field of virtual-reality fusion experiments and digital twin technology, and relates to a digital twin early warning method for experimental risks with multi-domain collaborative monitoring. Background Art

[0002] As modern engineering structures become increasingly complex and large-scale, structural testing plays an irreplaceable role in validating design theories and evaluating safety performance. In recent years, structural digital twin monitoring has become a key technology for ensuring the safety and reliability of structural testing. However, with the increasing complexity of engineering structures and increasing safety requirements, traditional structural digital twin monitoring methods have gradually exposed numerous problems.

[0003] Under complex loads or extreme conditions, structures often exhibit strong nonlinear behavior due to material nonlinearity, geometric nonlinearity, or local damage accumulation, and may even cause sudden failure. Traditional structural test monitoring methods often rely on linear assumptions or empirical thresholds, making it difficult to accurately capture the nonlinear evolution process, resulting in delayed or misjudgment of risk warnings, which seriously threatens test safety and data reliability. At the same time, although nonlinear finite element simulation is an effective means of evaluating the nonlinear behavior of structures, it faces significant challenges in practical applications: First, idealized structural loading methods often make it difficult to accurately simulate actual test conditions, especially under strong nonlinear conditions. This simplification will lead to continuous accumulation and significant amplification of errors; second, high-fidelity nonlinear analysis requires fine meshing and complex constitutive models, resulting in an exponential increase in computational complexity; third, convergence issues in the iterative solution process further increase the computational cost. These bottlenecks seriously restrict the application effect of a single evaluation method of nonlinear finite element simulation in real-time monitoring.

[0004] At the same time, sensors in the test are susceptible to environmental interference or hardware failures, resulting in missing or abnormal data at the measurement points, which also poses a challenge to the effective implementation of structural digital twin monitoring. The core goal of structural digital twin monitoring is to collect real-time structural response data through sensors and evaluate the health status of the structure based on this data. However, in practical applications, problems such as sensor failure, missing or abnormal data often lead to a decrease in the reliability of the monitoring system. Especially in large-scale tests of complex structures, the number of sensors is huge, and the failure of some sensors or abnormal data may affect the risk assessment of the entire structure. Traditional methods lack real-time identification and repair mechanisms, which seriously affects the reliability of predictions.

[0005] Li Ying et al. (Chinese invention patent CN118940569A) proposed a multi-source data fusion virtual testing method based on support vector machine classification and regression training for target ship operating conditions. Although this method trains a proxy model by fusing batches of simulation data with a small amount of test data, the proxy model only has high prediction accuracy for the test conditions used for model training, and its accuracy for unknown conditions remains to be verified. Fang Ming et al. (Chinese invention patent CN119004594A) reduced the computational workload by dividing large, complex structures into local nonlinear substructures and simultaneously achieved rapid online acquisition of analysis results using order reduction methods. However, this method focuses on nonlinear seismic response analysis and does not incorporate real test data. Furthermore, none of these approaches consider the construction of digital loading systems or the real-time identification and repair of faulty sensors.

[0006] To address these challenges, a new early warning method is urgently needed that integrates the advantages of physical mechanisms and data-driven approaches. By combining high-fidelity nonlinear modeling with efficient order reduction techniques, it can achieve dynamic and accurate prediction of complex structural test risks. This paper proposes a digital twin early warning method for test risks with multi-domain collaborative monitoring, aiming to overcome these technical bottlenecks and provide an innovative solution for the safety and controllability of structural testing. Summary of the Invention

[0007] To address the challenges of existing technologies, the present invention provides a digital twin early warning method for test risks with multi-domain collaborative monitoring. By combining physical strain gauge and displacement sensor data measured during structural testing with full-field response reconstruction results, real-time test risk warnings are performed. Simultaneously, based on the measured data, dynamic updates and evolution of the prediction model are achieved. Ultimately, a digital twin early warning system for test risks with multi-domain collaborative monitoring is established, enabling precise monitoring and risk warning for strength tests.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] A digital twin early warning method for test risks with multi-domain collaborative monitoring includes the following steps:

[0010] The first step is to establish a detailed finite element simulation model that takes nonlinearity into account, including the digital tooling system, digital sensors, and test pieces. Specifically:

[0011] Step 1.1: Construct a finite element simulation model based on the geometric model of the structure, and perform nonlinear area division according to the location of the hazardous area.

[0012] Specifically, the nonlinear region should be located in a position that meets the following requirements: ① The interface between the linear and nonlinear regions should not be too close to the dangerous area where nonlinearity occurs (the minimum distance between the interface and the dangerous area should be greater than 2*local grid size or 3*structure thickness); ② The interface between the linear and nonlinear regions should not be located in the area where the mechanical behavior of the structure changes significantly (stress gradient >20% / mm). Common areas that should be excluded include stress concentration areas, areas where material properties change, contact surfaces or connection parts.

[0013] Furthermore, for complex structures with too many grids, simplified finite element models can also be constructed using the substructure method and the submodel method. The specific analysis process of the substructure method is as follows: First, it is necessary to use simplified shell modeling for the non-dangerous linear areas of the structure, and for the structural parts including the dangerous nonlinear areas, a local detailed finite element model is established according to the geometric model of the structure, and fine mesh division is performed. The specific analysis process of the substructure method is as follows: the global shell model and the local detailed model of the structure are connected through binding, solid-shell coupling, and MPC beam interaction to achieve effective load transfer. The specific analysis process of the submodel method is as follows: based on the simulation results of the global shell model, the simulation loading is achieved by applying submodel boundaries to the local detailed model of the structure. If there are multiple levels of submodels, the submodels of each level and the global model can be gradually added based on this idea.

[0014] Step 1.2: Based on the nonlinear region obtained in step 1.1, fine meshing is performed, and linear attributes are applied to the linear region and nonlinear attributes are applied to the nonlinear region.

[0015] The linear properties include material linear properties, and the nonlinear properties include material nonlinear properties and contact nonlinear properties. Material nonlinear properties include material plasticity and damage criteria. Material plasticity can be set using the yield criterion, hardening model, flow law, or porous metal plasticity model, and the damage criterion can be set using the ductile metal damage criterion, Johnson-Cook criterion, or Hashin damage initiation criterion. Contact nonlinear properties can be set using the hard contact method, penalty function method, Lagrange multiplier method, or augmented Lagrangian method.

[0016] In step 1.3, a parametric script is used to assemble the digital tooling system, digital sensors, and digital specimen, creating a detailed finite element model of the digital tooling system, digital sensors, and specimen. Connections within the digital tooling system and between the digital tooling system, digital sensors, and specimen are achieved through binding, MPC beams, or coupling. Specifically:

[0017] Furthermore, the digital tooling system includes a loading lever, a joint, and a fixing fixture, and the digital sensors include strain sensors, displacement sensors, temperature sensors, force sensors, and acceleration sensors. The binding between the internal parts of the digital tooling system includes the loading lever and the loading lever, and the loading lever and the joint; the binding between the digital tooling system and the specimen includes the joint and the specimen, the loading lever and the specimen, and the fixing fixture and the specimen; the parameterized script can move the digital sensor to the patch position and complete the binding with the specimen according to the test patch scheme, and the patch position coordinates and angle can be changed arbitrarily. The test patch scheme includes the sensor type, number, patch position, and patch angle. The load of the finite element model is applied to the loading coupling point of the above-mentioned digital tooling system through the parameterized script.

[0018] The second step is to conduct nonlinear finite element simulation analysis based on the fine finite element simulation model considering nonlinearity constructed in the first step, sample based on the simulation results, extract the mechanical response results and form the corresponding data set, and build a multi-level mechanical response field inversion and reduction model. Specifically:

[0019] Step 2.1: Perform simulation analysis on the refined finite element simulation model constructed in the first step, taking into account nonlinearity. Specifically, calculations are performed using statics, explicit dynamics, implicit dynamics, or Riks method analysis steps to obtain finite element simulation analysis results for the model.

[0020] Step 2.2: Based on the obtained finite element simulation analysis results, sensitivity analysis is performed based on the model parameters.

[0021] Specifically, the specific process of sensitivity analysis is: for the model shown in formula (1) ,have:

[0022] (1)

[0023] in, are model parameters, is the total number of parameters, To use parameters The calculated finite element results. Then the local sensitivity of the parameters for:

[0024] (2)

[0025] Using the central difference method to approximate formula (2), we get:

[0026] (3)

[0027] in, Represents model parameters The increase in .

[0028] Specifically, the model parameters include material elastic modulus, Poisson's ratio, plasticity parameters, load size, position and direction.

[0029] Step 2.3, select the one with the highest sensitivity The model parameters are set to generate several sampling points in the design space, and interval simulation sampling is performed to obtain the finite element model sampling results.

[0030] Furthermore, the design space is a load design domain based on multiple test conditions, and the interval simulation sampling method includes Latin hypercube sampling, uniform random sampling or stratified sampling.

[0031] In step 2.4, a sample set is constructed based on the finite element model sampling results obtained in step 2.3, and a snapshot matrix is ​​assembled.

[0032] The constructed snapshot matrix , the number of rows in the matrix Represents the total number of finite element nodes of the specimen and the number of matrix columns Represents the total number of frames of all simulation results, the input matrix , the number of rows in the matrix is the number of digital sensor channels. At the same time, during the strength static verification test, the material is in the elastic stage at low loading levels, and the strength mechanical response of the specimen changes linearly. However, as the loading level increases, the material enters the plastic stage and even fails. At this time, the strength mechanical response of the specimen shows obvious nonlinear changes. Therefore, it is necessary to construct the snapshot sub-matrix hierarchically to reduce the nonlinear effect. Specifically:

[0033] Methods for constructing hierarchical snapshot matrices include: conducting simulation analysis based on the test conditions to be conducted, setting output frames based on the test loading step, extracting the mechanical response field of each frame to construct a snapshot matrix. For each loading level, the output of all frames within the window is expanded 5%-15% forward and backward, starting from the loading moment, and reorganizing the output of all frames within the window into snapshot submatrices. Alternatively, snapshot matrix segmentation is performed based on correlation: first, the Pearson correlation coefficient matrix between the results of each frame of all sampling results is calculated, then hierarchical clustering or K-means clustering algorithms are used to cluster based on the Pearson correlation coefficient, and finally, the corresponding columns are assigned to each submatrix based on the clustering results to create snapshot submatrices. Alternatively, snapshot matrix segmentation is performed based on principal component analysis: first, the snapshot matrix is ​​centered and normalized, then principal component analysis is performed on the processed matrix to obtain a load matrix, and then the load matrix is ​​grouped using simple thresholding, K-means clustering, or rotated principal component analysis. Finally, snapshot submatrices are created based on the grouping results.

[0034] At the same time, based on the snapshot matrix layering results, the input matrix The same layering is performed according to the corresponding frame number to obtain the snapshot sub-matrix of the input matrix.

[0035] In step 2.5, the reduced-order method is used to perform eigendecomposition on each snapshot submatrix after stratification, extracting the corresponding reduced-order basis vectors and their reduced-order basis coefficients. The reduced-order basis vectors and their corresponding coefficients are then truncated and screened by setting an energy threshold, ultimately obtaining the reduced-order basis and their corresponding reduced-order basis coefficients. Simultaneously, the reduced-order basis coefficient prediction proxy model for each reduced-order model is trained based on the input matrix and the reduced-order coefficient set. Specifically:

[0036] The constructed multi-level mechanical response field inversion and reduction model includes a linear domain reduction model and a nonlinear domain reduction model. The linear domain reduction model takes as input the digital displacement and strain sensor simulation results, and outputs the full-field mechanical response in the linear domain. The nonlinear domain reduction model takes as input the digital displacement, strain sensor simulation results, and the mechanical response at the interface in the linear domain, and outputs the full-field mechanical response in the nonlinear domain.

[0037] Furthermore, the order reduction method includes proper orthogonal decomposition, Krylov method, dynamic mode decomposition, random projection, non-negative matrix decomposition or canonical multivariate decomposition.

[0038] Furthermore, the method for screening the reduced-order basis vectors and corresponding coefficients is that the energy threshold used for truncation screening is not less than 99%.

[0039] Furthermore, methods for constructing reduced-order basis coefficient prediction proxy models include Gaussian process regression, random forest, support vector machine, deep neural network or gradient boosting tree.

[0040] The third step is to establish data feature judgment criteria to screen sensor data, and fill in abnormal measurement point data through data filling algorithm to obtain a complete sensor data set; at the same time, the first-level structure failure judgment is performed based on the complete sensor data set. Specifically:

[0041] Step 3.1: During the test, use the data feature judgment criteria to screen sensor data failures to screen and eliminate faulty sensors. Specifically:

[0042] The data feature judgment criteria include physical rationality judgment, outlier judgment, strain change trend judgment, data stability judgment, white noise fault judgment, and zero drift judgment; the data set used for judgment includes the historical sensor data set composed of historical test data and the current sensor data set composed of the current test data. The details are as follows:

[0043] (1) Physical rationality criterion:

[0044] The physical rationality criterion means that the strain value of a single strain gauge channel should be within the reasonable range of the material's mechanical properties, which can be expressed as:

[0045] (4)

[0046] in, is the strain value of the current sensor data set; is the allowable strain value of the material.

[0047] (2) Outlier criteria:

[0048] The outlier criterion uses the 3σ principle to detect outliers for a single sensor, where σ is the standard deviation of the data. This principle is based on the normal distribution assumption and assumes that the probability of a data point falling within the range of ±3 times the standard deviation of the mean is 99.7%. Data points outside this range are considered outliers. The criterion formula is:

[0049] (5)

[0050] in, is the response value of the current sensor data set; is the mean of the historical sensor dataset; is the standard deviation of the historical sensor dataset.

[0051] (3) Strain change trend criterion:

[0052] The strain change trend criterion means that the strain change trend between two adjacent loading steps should be consistent with the theoretical or historical data, and should not show an increasing or decreasing phenomenon that is unrelated to the external load size. The criterion formula uses the least squares regression slope between the response data and the time series. :

[0053] (6)

[0054] in, is the loading step sequence of the current sensor dataset; Response data sequence for the current sensor data set; for The number of data points in the.

[0055] (4) Data stability criteria:

[0056] The data stability criterion means that the strain fluctuation between two adjacent loading steps should be within a reasonable range. Stable data should have a small fluctuation range, while data with excessive fluctuations may indicate a sensor failure or external interference. The criterion formula is:

[0057] (7)

[0058] in, for The response value of the current sensor data set under the loading level; for The response value of the current sensor data set under the loading level; is the standard deviation of the historical sensor dataset; is the threshold coefficient, which is usually in the range of 2-3.

[0059] (5) White noise fault criterion:

[0060] White noise fault means that the sensor reading consists only of background noise, that is, the sensor value fluctuates around 0. The judgment formula is:

[0061] (8)

[0062] in, is white noise, that is, a random signal with unknown mean and covariance.

[0063] (6) Zero drift criterion:

[0064] Zero drift means that the zero strain value of a single sensor should not drift significantly. The criterion formula is:

[0065] (9)

[0066] in, The zero-point sensor value of the current sensor data set; The maximum value allowed for the zero point sensor.

[0067] In the process of judging sensor failure, structural failure judgment is also carried out at the same time. Through the strain change trend judgment and data stability judgment in the judgment criteria, the sudden slope mutation during the loading process is further judged to complete the first-level failure judgment.

[0068] In step 3.2, the sensor data filling model is constructed offline through the data filling algorithm. Online, the actual sensor data is collected for mapping, the missing data of the failed sensor is reconstructed, the data of the faulty sensor is filled, and a complete sensor data set is constructed. At the same time, the first-level failure judgment is performed.

[0069] Furthermore, the data filling algorithm refers to the process of reasonably estimating and reconstructing the missing parts of the overall data using mathematical modeling or statistical methods. The data filling algorithms used include linear interpolation, singular value decomposition, margin eigenorthogonal decomposition, random forest method, or generative adversarial network.

[0070] The fourth step is to perform full-field mechanical response inversion and online real-time correction through the multi-level mechanical response field inversion reduction model established in the second step and the complete sensor data set established in the third step; the time series prediction model established in combination with historical loading data is used to predict the sensor data of future loading levels, perform second-level structural failure judgment, and perform response inversion of the focus area. By calculating the damage factor of the nonlinear area, future risk warnings are carried out in real time to achieve third-level structural failure judgment. Specifically:

[0071] In step 4.1, the complete sensor data set obtained in step 3.2 is input online. The reduced-order basis coefficients of each constructed reduced-order model are obtained based on the reduced-order basis coefficient prediction proxy model constructed in step 2.5. The reduced-order basis coefficients are multiplied by the reduced-order basis to complete the raised-order inversion of the mechanical response field and obtain the full-field mechanical response field result.

[0072] Step 4.2: Based on the complete sensor data set and the residuals of the numerical calculation of the sensor measurement point positions, a full-field response correction proxy model is constructed to perform online real-time correction on the mechanical response field of the specimen after the upgraded-order inversion.

[0073] Based on the sensor coordinates in the test patch scheme, extract the response values ​​at the corresponding coordinates in the specimen's mechanical response field predicted in step 4.1. Calculate the deviation between the predicted sensor response values ​​and the actual sensor values ​​measured during the test. A global response correction proxy model is constructed based on the relationship between each sensor's deviation value and the corresponding coordinates. By superimposing the resulting global correction deviation value with the specimen's mechanical response field obtained in step 4.1, online dynamic correction of the global mechanical response field is achieved.

[0074] Furthermore, methods for training the full-field response correction agent model include radial basis functions, Gaussian process regression, random forests, transfer learning neural networks, or incremental learning gradient boosting trees.

[0075] Step 4.3: By establishing a historical loading database, training the sensor timing prediction model, and predicting the sensor data of future loading steps, the second-level failure determination is completed.

[0076] The historical loading database includes simulation sample data at the measurement points and historical working condition loading sensor data. The sensor time series data model training method uses a long short-term memory network, autoregressive and moving average model, recurrent neural network, convolutional neural network, Transformer model, support vector machine, or random forest algorithm.

[0077] During the training process of the sensor time series prediction model, based on the time series data from the historical load database, a time series prediction method is used to capture the time series characteristics of the impact of load changes on the response of the measurement point, and then predict the time series as the load changes. The time series prediction algorithm uses the time series dependency characteristics to identify short-term and long-term dependencies in the time series data, enabling the sensor time series prediction model to maintain accurate predictions of the measurement point response even in the face of complex load changes.

[0078] In the online stage, based on the established sensor time series prediction model, the sensor data set consisting of the collected normal sensor data and the replaced faulty sensor data in step 3.2 is input to realize real-time sensor future prediction. The second-level failure judgment is realized by combining the predicted value with the first-level failure judgment method in step 3.1.

[0079] Furthermore, after the test is completed, the normal sensor data collected and the faulty sensor data replaced in step 3.2 are added to the historical loading database as historical working condition loading sensor data, and the sensor timing prediction model is retrained and updated.

[0080] In step 4.4, based on the multi-level mechanical response field inversion reduction model constructed in step 2, the predicted future loading step measurement point data is input to achieve real-time inversion of the future mechanical response field. The measurement point data includes strain and displacement data, and the mechanical response field includes strain, stress, plastic strain field, and displacement field.

[0081] Furthermore, after the test is completed, the input measurement point data and the inverted mechanical response field are used as snapshot matrix elements to update the snapshot matrix, and the order reduction is performed again to achieve dynamic updating of the reduced-order model.

[0082] Step 4.5: Based on the inverted mechanical response field, calculate the full-field damage factor value corresponding to the adopted damage criterion, obtain the full-field damage situation of the structure, perform the third-level failure judgment, and predict the loading level at which the damage factor reaches the failure threshold. Specifically:

[0083] Common damage criteria for calculating damage factors include the ductile metal damage criterion, the Johnson-Cook criterion, and the Hashin damage onset criterion. The damage factor ranges from 0 to 1, with the failure threshold typically set at 1. When the damage factor is equal to 0, the structure is considered undamaged, with no stiffness degradation. When the damage factor is greater than 0 and less than 1, the structure is considered partially damaged, with stiffness degradation beginning. When the damage factor is greater than 1, the material completely fails and loses its load-bearing capacity.

[0084] In step 4.6, after the damage factor value is calculated, the failure is determined by comparing it with the failure threshold. Based on the time-series predicted loading-strain curve, the loading level at which the damage factor value reaches the failure threshold is predicted, and an early warning is issued.

[0085] The beneficial effects of the present invention are:

[0086] (1) The present invention can realize real-time monitoring of the test process through nonlinear modeling and data-driven technology.

[0087] (2) To address the problem of difficulty in accurately simulating actual test conditions, a digital tooling system was established. To address the large amount of nonlinear analysis calculations, the dangerous area was refined and nonlinear properties were assigned to ensure the calculation accuracy while effectively controlling the calculation scale. To address the problem of sensor failure that may occur during the test, a data reconstruction mechanism for abnormal measurement points was established based on the data filling method. At the same time, a multi-level failure judgment method was constructed from the entire field to the dangerous area, from the current loading state to the future loading condition, combining non-invasive reduced-order models and time series prediction algorithms, realizing the digital twin warning of test risks with multi-domain collaborative monitoring.

[0088] In summary, the present invention integrates the advantages of nonlinear modeling and data-driven, improves the ability to characterize the nonlinear behavior of structures while ensuring computational efficiency, provides a technical approach for continuous updating and optimization of the model after the test, and provides a new solution for the safety monitoring of complex structural tests. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 This is a flowchart for implementing a digital twin early warning method for test risks with multi-domain collaborative monitoring;

[0090] Figure 2 This is a schematic diagram of the assembly of the reinforced tube structure test system;

[0091] Figure 3 This is a schematic diagram of the patch scheme for the reinforced tube structure test;

[0092] Figure 4 Schematic diagram of digital sensor assembly effect and simulation effect; Figure 4 a in the figure is the digital sensor assembly effect diagram. Figure 4 b is a schematic diagram of the digital sensor simulation effect;

[0093] Figure 5 Schematic diagram of the full-field von Mises stress inversion results of the reinforced tube structure;

[0094] Figure 6 Schematic diagram of the inversion results of the plastic strain field in the nonlinear region of the reinforced tube structure; Figure 6 a in the figure is the full-field cloud map of PE11. Figure 6 The b in the figure is the full-field cloud map of PE22. Figure 6 The c in the figure is the full-field cloud image of PE33. Figure 6 The d in the figure is the full-field cloud image of PE12. Figure 6 The e in the figure is the full-field cloud map of PE13. Figure 6 The f in the figure is the full-field cloud image of PE23;

[0095] Figure 7 This is a schematic diagram of early warning of failure determination at future loading levels in the reinforced tube structure test; Figure 7 a in the figure is a schematic diagram of the displacement-load curve of the reinforced tube structure test. Figure 7 b in the figure is a schematic diagram of sensor data prediction for the future loading step of the reinforced tube structure;

[0096] Figure 8 This is a schematic diagram of the calculation results of the full-field damage factor in the nonlinear region of the reinforced tube structure;

[0097] In the figure: 1 loading joint, 2 reinforcement tube, 3 loading lever, 4 electric push rod, 5 base platform, 6 tooling beam, 7 foundation platform. DETAILED DESCRIPTION

[0098] The following will be combined with a specific implementation example to clearly and completely explain the technical solutions involved in the present invention. The specific embodiments described here are only used to explain the present invention, not to limit the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0099] Figure 1 This is a flow chart of the technical route for implementing the present invention. According to the steps in the flow chart, the technical solution of the present invention includes:

[0100] The first step is to establish a detailed finite element simulation model that takes nonlinearity into account, including the digital tooling system, digital sensors, and test pieces. Specifically:

[0101] Step 1.1: Construct a finite element simulation model based on the geometric model of the structure, and perform nonlinear area division according to the location of the danger area.

[0102] A digital tooling system, digital sensors and a finite element model of the reinforced tube specimen were established. The digital tooling system includes a loading and clamping rod, a loading joint, an electric push rod, a base, a tooling beam and a foundation. The digital sensor is a digital strain sensor. ,high , base platform radius ,high The digital tooling system is modeled using C3D8 units. The material is high-strength alloy steel with an elastic modulus of 210,000 MPa, a Poisson's ratio of 0.3, and a density of 7.8 g / cm 3 The specimen material is 2024 aluminum alloy, with an elastic modulus of 85413 MPa, a Poisson's ratio of 0.33, and a density of 2.78 g / cm 3 The digital strain sensor is modeled as a shell S4R unit, with a size of 2×1mm and a thickness of 0.2mm. The material properties are consistent with those of the patch location. For local hazardous details, the material linear-nonlinear interface and the material and contact nonlinear regions are divided.

[0103] Step 1.2: Based on the nonlinear region obtained in step 1.1, fine meshing is performed, and linear attributes are applied to the linear region and nonlinear attributes are applied to the nonlinear region. In this embodiment:

[0104] For the contact nonlinear region, contact settings are applied to the screw hole surface and the screw surface, where the tangential behavior is set to a penalty function, the friction coefficient is set to 0.1, and the normal behavior is set to hard contact. For the material nonlinear region, in addition to applying metal elastic properties (consistent with the linear region of the specimen), the material nonlinear properties are obtained based on the uniaxial tensile test of a standard 2024 aluminum alloy rod specimen with a diameter of 5.991mm at room temperature, and the ductile metal damage criterion is selected for application. The specific properties are as follows: the yield stress is 336MPa, the ultimate stress is 550MPa, the corresponding plastic strain is 0.04933mm, the stress triaxiality is 0.33, the strain ratio is 0.0067, and the fracture strain is 0.04933mm.

[0105] Step 1.3: Use a parameterized script to complete the assembly of the digital tooling system, digital sensor, and digital test piece, and establish a detailed finite element model including the digital tooling system, digital sensor, and test piece. In this embodiment:

[0106] For the assembly of the digital tooling system and the test piece, the connection between the loading and clamping rod and the loading and clamping rod, the loading and clamping rod and the loading joint, the loading joint and the test piece, the loading joint and the tooling beam, the tooling beam and the foundation, the base platform and the test piece, the loading and clamping rod and the test piece, the base platform and the foundation, the electric push rod and the loading and clamping rod, and the electric push rod and the foundation is completed. The connection form is binding, and a surface set is established at the top of the loading lever. After assembly, the test system is as follows Figure 2 shown.

[0107] For the parametric assembly of digital strain sensors, according to Figure 3The test patch scheme shown is based on the getByBoundingSphere detection parameterized algorithm secondary developed by Abaqus to locate the target unit. Specifically, with the coordinates of the sensor in the global coordinate system as the center, the search radius is gradually expanded until the patch unit is found. Subsequently, the FaceToFace fitting algorithm is applied to parameterize and establish a geometric constraint relationship to ensure that the sensor and the target unit surface are accurately fitted. Then, the RadialInstancePattern rotation and copy function is used to generate a three-dimensional digital strain sensor. Finally, the connection relationship between the digital strain sensor and the specimen is established. By establishing a binding between the strain sensor and the surface of the specimen unit and specifying the binding tolerance size. According to the experimental measurement point layout plan, repeat the above process to parameterize and assemble all digital strain sensors in turn. Taking strain gauge No. 6 as an example, the maximum principal strain cloud diagram of the digital strain sensor after parameterized assembly and the digital strain sensor and the unit where the patch is located is shown as follows Figure 4 shown.

[0108] The second step is to conduct nonlinear finite element simulation analysis based on the fine finite element simulation model considering nonlinearity constructed in the first step, sample based on the simulation results, extract the mechanical response results and form the corresponding data set, and build a multi-level mechanical response field inversion and reduction model. Specifically:

[0109] Step 2.1: Perform simulation analysis on the refined finite element simulation model constructed in the first step, taking into account nonlinearity. Specifically, calculations are performed using statics, explicit dynamics, implicit dynamics, or Riks method analysis steps to obtain finite element simulation analysis results for the model.

[0110] In this example, the static analysis step is used to simulate and analyze the fine finite element simulation model considering nonlinearity constructed in the first step, and the finite element simulation analysis results of the model are obtained.

[0111] Step 2.2: Based on the obtained finite element simulation analysis results of the model, sensitivity analysis is performed based on the model parameters.

[0112] The sensitivity analysis method is used to analyze the elastic modulus, Poisson's ratio, plasticity parameters and load parameters of the material.

[0113] Step 2.3, select the one with the highest sensitivity The model parameters are set to generate several sampling points in the design space, and interval simulation sampling is performed to obtain the finite element model sampling results.

[0114] Through the sensitivity calculation in step 2.2, the two most sensitive model parameters, load magnitude and Young's modulus, were determined as sampling parameters. Using the Latin hypercube sampling method, random samples were drawn from the Young's modulus and load magnitude ranges of 0.7 to 1.2 times, with intervals of 0.1, for five test conditions. Loading was performed using a parameterized loading procedure, and simulation analysis was performed using Abaqus finite element simulation software, resulting in a total of 36 sets of simulation results.

[0115] In step 2.4, a sample set is constructed based on the finite element model sampling results obtained in step 2.3, and snapshot matrices are assembled to obtain the layered snapshot sub-matrices.

[0116] The multi-level mechanical response field inversion reduction model constructed in this example includes a linear domain reduction model and a nonlinear domain reduction model. The linear domain reduction model takes digital strain sensor results as input and outputs the full-field von Mises stress response in the linear domain. The nonlinear domain reduction model takes digital strain sensor results and von Mises stress results at the interface between the linear and nonlinear domains as input and outputs the full-field plastic strain values ​​PE11, PE22, PE33, PE12, PE23, and PE13.

[0117] According to the experimental loading step, the number of frames of the simulation results is divided into 0-50% (10% interval division, a total of 5 frames), 50%-70% (5% division, a total of 4 frames), and 70%-100% (2% division, a total of 15 frames). Based on the snapshot matrix layering method①, the partition window is expanded by 10% before and after each loading step. The full-field plastic strain PE11, PE22, PE33, PE12, PE23, and PE13 values ​​of the nonlinear region, the digital strain sensor results, the full-field von Mises stress response in the linear region, and the von Mises stress results at the interface between the linear and nonlinear regions of all frames in the window are extracted. The layered snapshot sub-matrices are constructed according to the number of divided frames.

[0118] In step 2.5, the eigendecomposition of each snapshot submatrix after stratification is performed using an order reduction method to extract the corresponding reduced-order basis vectors and their reduced-order basis coefficients. The reduced-order basis vectors and their corresponding coefficients are truncated and screened by setting an energy threshold, ultimately obtaining the reduced-order basis and their corresponding reduced-order basis coefficients. Simultaneously, a proxy model for predicting the reduced-order basis coefficients of each reduced-order model is trained based on the input matrix and the reduced-order coefficient set.

[0119] The Proper Orthogonal Decomposition (POD) technique is selected to perform non-invasive order reduction on the input and output snapshot sub-matrices of each reduced-order model, and the reduced-order basis and the corresponding reduced-order basis coefficients are obtained. The specific principle is:

[0120] Assemble the sample set into a snapshot matrix:

[0121] (10)

[0122] in, is the snapshot matrix, is the snapshot vector.

[0123] After obtaining the snapshot matrix, use the POD model reduction method to process the snapshot matrix constructed in the first step , obtain the basis and calculate the reduced-order basis coefficients corresponding to the data at each stage Based on the snapshot matrix Calculate the correlation matrix :

[0124] (11)

[0125] Solving the correlation matrix The eigenvalue of and eigenvectors , where the number of eigenvalues ​​is , and arrange the eigenvalues ​​from large to small:

[0126] (12)

[0127] Select Satisfaction of The minimum value of is the cutoff threshold, which is usually 99% or a number closer to 1. The energy occupied by the eigenvalue of the POD mode.

[0128] (13)

[0129] According to the previous Reduced basis matrix is ​​established by using eigenvectors of order :

[0130] (14)

[0131] Where, are the reduced basis vectors that make up the reduced basis matrix.

[0132] Based on the reduced basis matrix Calculate the reduced-order basis coefficients corresponding to the data at each stage .

[0133] (15)

[0134] Finally, based on the input matrix and the reduced-order coefficient set, the Gaussian process regression method is used in this example to perform the reduced-order basis coefficients of each reduced-order model. Training of prediction agent models.

[0135] The third step is to establish data feature judgment criteria to screen sensor data, and at the same time, fill in the abnormal measurement point data through the data filling algorithm to obtain a complete sensor data set; at the same time, perform the first-level structure failure judgment based on the complete sensor data set. Step 3.1: During the test, use the data feature judgment criteria to screen sensor data failures to screen and eliminate faulty sensors. Specifically:

[0136] Secondly, in the online stage, sensor data were screened according to the established criteria. The screening results showed that acquisition anomalies occurred at measuring points 1, 5, and 8. Among them, measuring points 1 and 5 did not meet the physical rationality criteria, measuring point 8 did not meet the white noise fault criteria, and no measuring point did not meet the strain change trend criteria and data stability criteria.

[0137] In step 3.2, the sensor data filling model is constructed offline through the data filling algorithm. Online, the actual sensor data is collected for mapping, the missing data of the failed sensor is reconstructed, the data of the faulty sensor is filled, and a complete sensor data set is constructed. At the same time, the first-level failure judgment is performed.

[0138] Based on the sampled digital strain sensor measurement values, a sensor data filling model is constructed based on the data filling method mentioned in step 3.2. In this example, the interval intrinsic orthogonal decomposition technology is used for construction. The construction principle is as follows:

[0139] Get sensor measurement value simulation data based on digital sensors and construct a snapshot matrix consisting of all digital sensor simulation data :

[0140] (16)

[0141] in, For the The vector of all sensor readings under each working condition, is the number of strain sensors, is the total number of sampled working conditions. In the offline phase, based on the POD reduction method mentioned above, the snapshot matrix Perform POD decomposition and extract by truncation screening The reduced order basis matrix is ​​constructed based on the main modes , forming a reduced-order model.

[0142] When online, the data of measuring points 1, 5, and 8 in the collected data are replaced with null values, and the data of the remaining normal test points are used as the input of the interval intrinsic orthogonal decomposition model to obtain the replacement values ​​of measuring points 1, 5, and 8 in real time. The specific principle of interval intrinsic orthogonal decomposition online data filling is as follows:

[0143] When online, define the corresponding normal sensor vector Contains all normal sensor indices, mask matrix Each line is a unit vector, corresponding to the normal sensor index. Normal sensor data is used as input:

[0144] (17)

[0145] in is the normal number of sensors. First, centralize the data:

[0146] (18)

[0147] Extracting reduced-order basis matrices , solve the reduced-order basis coefficients by the least squares method :

[0148] (19)

[0149] We can get:

[0150] (20)

[0151] Using the obtained reduced-order basis coefficients Reconstructing data:

[0152] (twenty one)

[0153] Decentralize to get predicted sensor data:

[0154] (twenty two)

[0155] Replace missing sensor data with , combined with the normally collected sensor data to form a complete sensor data set, completing the process of predicting and replacing missing sensor data based on interval intrinsic orthogonal decomposition. The results show that compared with the normally collected data, the average relative error of the replacement value is only 3.5%.

[0156] In the experiment, the first-level structural failure judgment was performed based on the complete sensor data set obtained above, and the maximum sensor strain value was , which is less than the material failure strain determined by the uniaxial tensile test of the standard 2024 aluminum alloy rod specimen before the test. The result of the first-level structural failure is judged as no failure.

[0157] The fourth step is to perform full-field mechanical response inversion and online real-time correction through the multi-level mechanical response field inversion and reduction model established in the second step and the complete sensor data set established in the third step; combine the time series prediction model established with historical loading data to predict the sensor data of future loading levels, perform second-level structural failure judgment, and perform response inversion of the area of ​​interest. Through the calculation of the nonlinear area damage factor, future risk warning is carried out in real time to achieve third-level structural failure judgment. Step 4.1, input the complete sensor data set obtained in step 3.2 online, obtain the reduced-order basis coefficients of each reduced-order model constructed based on the reduced-order basis coefficient prediction proxy model constructed in step 2.5, and use the reduced-order basis coefficients to multiply the reduced-order basis to complete the increased-order inversion of the mechanical response field of the specimen and obtain the full-field mechanical response field results. Specifically:

[0158] The complete sensor data set (30% loading level) obtained in the third step is input into the established full-field reduced-order model, and the full-field von Mises stress results are obtained by upgrading the order.

[0159] Step 4.2: Based on the complete sensor data set and the residuals of the numerical calculation of the sensor measurement point positions, a full-field response correction proxy model is constructed to perform online real-time correction on the mechanical response field of the specimen after the upgraded-order inversion.

[0160] The RBF bridge function is used to establish the full-field von Mises stress response correction proxy model online, and the full-field von Mises stress response is corrected online, such as Figure 5 The full-field von Mises stress response prediction results were compared with the experimental results using the leave-one-out validation method, that is, 26 of the 27 test points were used to construct the proxy model each time, and the data of the remaining point were predicted, and the cycle was repeated 27 times. The relative root mean square error (RRMSE) and the coefficient of determination (R) were used to calculate the predicted values. 2 As the accuracy criterion, the formula is as follows:

[0161] (twenty three)

[0162] (twenty four)

[0163] The leave-one-out validation results show that the RRMSE is 0.096 and R 2 It is 0.9557.

[0164] Step 4.3: Establish a historical loading database offline, train the sensor time series prediction model, and predict the sensor data of future loading steps online to complete the second-level failure judgment.

[0165] Before the test, a historical loading database consisting of digital strain sensor measurement values ​​at each loading level was established offline based on simulation data. The time series relationship between the loading level and the digital strain sensor measurement values ​​was trained using a time series prediction algorithm. In this example, the time series prediction algorithm uses the LSTM algorithm. The specific principles are as follows:

[0166] For the time series data in the historical loading database, the LSTM network training will go through the following process at each time step:

[0167] (1) Forget gate: determines which information is forgotten.

[0168] (25)

[0169] in, is the sigmoid activation function, is the weight of the forget gate, is the bias term, is the output at the previous moment, is the current input.

[0170] (2) Input gate and candidate cell states: decide which new information to write into the cell state.

[0171] (26)

[0172] (27)

[0173] Among them, tanh is the hyperbolic tangent function, and are the weights of the input gate and candidate cell state, respectively.

[0174] (3) Update cell status : Update the cell state, forget the state of the previous moment through the forget gate, and update the current state through the input gate.

[0175] (28)

[0176] (4) Output gate : Determines how much information to output for the current cell state.

[0177] (29)

[0178] Current output Can be calculated as:

[0179] (30)

[0180] Then, based on the established sensor time series prediction model, the complete sensor data set (30% loading level) obtained in the third step is input online to predict the sensor data of 80% loading level. Taking the No. 14 strain channel as an example, the prediction effect is as follows: Figure 7 The predicted maximum von Mises stress is 252.1 MPa, which is less than the material yield limit, and the second-level structural failure is determined to be non-failure. Compared with the data obtained by normal loading, the average error of the 81 strain channels is 5.1%.

[0181] Step 4.4: Based on the multi-level mechanical response field inversion reduction model constructed in the second step, the predicted future loading step measurement point data is input to achieve real-time inversion of the future mechanical response field. Specifically:

[0182] Then, based on the established multi-level mechanical response field inversion reduction model, the sensor data of the predicted 80% loading level are input online to obtain the predicted values ​​of the full-field plastic strains PE11, PE22, PE33, PE12, PE23, and PE13 in the nonlinear region under the 80% loading level, as shown in Figure 2. Figure 6 As shown, the time taken is 0.0073s.

[0183] Step 4.5: Based on the inverted mechanical response field, calculate the full-field damage factor value corresponding to the adopted damage criterion, obtain the full-field damage situation of the structure, perform the third-level failure judgment, and predict the loading level at which the damage factor reaches the failure threshold. Specifically:

[0184] This example uses the damage factor calculation method corresponding to the ductile metal damage criterion to calculate the full-field damage factor based on the full-field plastic strain inversion results. The calculation process is as follows:

[0185] First, calculate the equivalent plastic strain. The plastic strain tensor satisfies volume incompressibility. The equivalent plastic strain increment calculation formula is:

[0186] (31)

[0187] in, Plastic strain increment components in each direction, is the plastic shear strain increment in the xy plane.

[0188] Secondly, calculate the cumulative equivalent plastic strain :

[0189] (32)

[0190] Finally, the damage factor is calculated based on the accumulated equivalent plastic strain and fracture strain:

[0191] (33)

[0192] in, is the fracture strain of the material.

[0193] Finally, the full-field damage factor of the nonlinear region of the local detail model under the 80% loading level is calculated, as follows: Figure 8 shown.

[0194] In step 4.6, after the damage factor value is calculated, the failure is determined by comparing it with the failure threshold. Based on the time-series predicted loading-strain curve, the loading level at which the damage factor value reaches the failure threshold is predicted, and an early warning is issued.

[0195] The calculated maximum value of the full-field damage factor is 0.963, which indicates that the structure has entered a partially damaged state at this loading level, but has not yet failed. The third failure is considered to be non-failure. According to the prediction, at the 86% loading level, the maximum value of the damage factor reaches the threshold of 1, and the structure will fail. The failure judgment warning effect is as follows: Figure 7 shown.

[0196] Finally, it is particularly noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit its scope of protection. Those skilled in the art, based on a full understanding of the above embodiments, may reasonably adjust the technical solutions or replace some of the technical features with equivalents. Such modifications or replacements shall still fall within the scope of protection defined by the claims of the present invention.

Claims

1. A digital twin early warning method for test risks with multi-domain collaborative monitoring, characterized by: The test risk digital twin early warning method includes the following steps: The first step is to establish a detailed finite element simulation model that takes nonlinearity into account, including the digital tooling system, digital sensors, and test pieces. In the second step, based on the refined finite element simulation model constructed in the first step, nonlinear finite element simulation analysis is performed. Based on the simulation results, sampling is performed, the mechanical response results are extracted and the corresponding data sets are compiled to construct a multi-level mechanical response field inversion and reduction model. The third step is to establish data feature judgment criteria to screen sensor data, and reconstruct abnormal measurement point data through data filling algorithm to obtain a complete sensor data set. At the same time, the first-level structure failure judgment is performed based on the complete sensor data set. In the fourth step, full-field mechanical response inversion and online real-time correction are performed through the multi-level mechanical response field inversion and reduction model established in the second step and the complete sensor data set established in the third step. The second-level full-field dangerous area judgment is performed based on the full-field inversion data; and the time series prediction model established in combination with historical loading data is used to predict the sensor data of future loading levels, perform response inversion of the area of ​​concern, and perform real-time future risk warnings through the calculation of nonlinear area damage factors to achieve the third-level structural failure judgment.

2. The digital twin early warning method for test risk with multi-domain collaborative monitoring according to claim 1 is characterized in that: The first step is specifically as follows: Step 1.1: Construct a finite element simulation model based on the geometric model of the structure, and perform nonlinear area division according to the location of the hazard area. Step 1.2: Based on the nonlinear region obtained in step 1.1, fine meshing is performed, and linear attributes are applied to the linear region and nonlinear attributes are applied to the nonlinear region; In step 1.3, a parametric script is used to complete the assembly of the digital tooling system, digital sensor, and digital specimen, and a detailed finite element model including the digital tooling system, digital sensor, and specimen is established. Connections are achieved within the digital tooling system and between the digital tooling system, digital sensor, and specimen through binding, MPC beams, or coupling.

3. The digital twin early warning method for test risk with multi-domain collaborative monitoring according to claim 2 is characterized in that: In the first step: In step 1.1, the nonlinear region should be located in a location that meets the following requirements: ① The linear-nonlinear region interface should not be too close to the dangerous area where nonlinearity occurs; ② The linear-nonlinear region interface should not be located in an area where the mechanical behavior of the structure changes significantly, with a stress gradient of >20% / mm; For complex structures with multiple meshes, finite element models can also be constructed using the substructure method and submodel method. In step 1.2, the linear properties include material linear properties, and the nonlinear properties include material nonlinear properties and contact nonlinear properties; the material nonlinear properties include material plasticity and damage criteria, wherein the material plasticity is set by a yield criterion, a hardening model, a flow law, or a porous metal plasticity model, and the damage criterion is set by a ductile metal damage criterion, a Johnson-Cook criterion, or a Hashin damage initiation criterion; the contact nonlinear properties are set by a hard contact method, a penalty function method, a Lagrange multiplier method, or an augmented Lagrangian method.

4. The digital twin early warning method for test risk with multi-domain collaborative monitoring according to claim 1 is characterized in that: The second step is specifically as follows: Step 2.1: Perform simulation analysis on the refined finite element simulation model considering nonlinearity constructed in the first step; specifically, perform calculations through static, explicit dynamics, implicit dynamics, or Riks method analysis steps to obtain the finite element simulation analysis results of the model; Step 2.2, based on the obtained finite element simulation analysis results, sensitivity analysis is performed based on the model parameters; Step 2.3: Select the two sets of model parameters with the highest sensitivity to generate several sampling points in the design space, perform interval simulation sampling, and obtain the finite element model sampling results; Step 2.4: construct a sample set based on the finite element model sampling results obtained in step 2.3, assemble the snapshot matrix, and obtain the layered snapshot sub-matrices; In step 2.5, the reduced-order method is used to perform eigendecomposition on each snapshot sub-matrix after stratification, and the corresponding reduced-order basis vectors and their reduced-order basis coefficients are extracted. The reduced-order basis vectors and corresponding coefficients are truncated and screened by setting the energy threshold, and finally the reduced-order basis and its corresponding reduced-order basis coefficients are obtained. At the same time, the reduced-order basis coefficient prediction agent model of each reduced-order model is trained based on the input matrix and the reduced-order coefficient set to construct a multi-level mechanical response field inversion reduced-order model.

5. The digital twin early warning method for test risk with multi-domain collaborative monitoring according to claim 4 is characterized in that: In the second step: In step 2.2: The specific process of sensitivity analysis is as follows: For the model shown in formula (1) ,have: （1） in, are model parameters, is the total number of parameters, To use parameters The calculated finite element results; then the local sensitivity of the parameters for: （2） Using the central difference method to approximate formula (2), we get: （3） in, Represents model parameters the increase in The model parameters include material elastic modulus, Poisson's ratio, plasticity parameters, load size, position and direction; In step 2.3, the design space is a load design domain based on multiple test conditions, and the interval simulation sampling method includes Latin hypercube sampling, uniform random sampling, or stratified sampling; In step 2.4, the snapshot matrix constructed is , the number of rows in the matrix Represents the total number of finite element nodes of the specimen and the number of matrix columns Represents the total number of frames of all simulation results, the input matrix , the number of rows in the matrix is the number of digital sensor channels; at the same time, the snapshot matrix is ​​layered, and based on the snapshot matrix layering results, the input matrix Perform the same layering according to the corresponding frame number to obtain the snapshot sub-matrix of the input matrix; In step 2.5, the order reduction method includes proper orthogonal decomposition, Krylov method, dynamic mode decomposition, random projection, non-negative matrix decomposition or canonical multivariate decomposition; In step 2.5, the multi-level mechanical response field inversion reduction model includes: a linear region reduction model and a nonlinear region reduction model; the sample input of the linear region reduction model is the digital displacement and strain sensor simulation results, and the output is the full-field mechanical response of the linear region; the sample input of the nonlinear region reduction model is the digital displacement, strain sensor simulation results, and the mechanical response of the linear region on the interface, and the output is the full-field mechanical response field of the nonlinear region; In step 2.5, the method for screening the reduced-order basis vectors and corresponding coefficients is that the energy threshold used for the truncation screening is not less than 99%; In step 2.5, the construction method of the reduced-order basis coefficient prediction proxy model includes Gaussian process regression, random forest, support vector machine, deep neural network or gradient boosting tree.

6. The digital twin early warning method for test risk with multi-domain collaborative monitoring according to claim 5 is characterized in that: In step 2.4, the method for constructing the snapshot submatrix includes: Conduct simulation analysis based on the test conditions to be carried out, set the output frame according to the test loading step, extract the strength mechanical response field of each frame to construct a snapshot matrix; for each loading level, take the loading moment as the benchmark, expand 5% of the frames forward and backward, and reorganize all the output frames in the window into a snapshot sub-matrix; Alternatively, snapshot matrix segmentation can be performed based on correlation. First, the Pearson correlation coefficient matrix between all frame results of all sampling results is calculated. Then, hierarchical clustering or K-means clustering algorithm is used to perform clustering based on the Pearson correlation coefficient. Finally, the corresponding columns are assigned to each sub-matrix according to the clustering results to create a snapshot sub-matrix. Or snapshot matrix segmentation is performed based on principal component analysis. First, the snapshot matrix is ​​centered and standardized. Then, principal component analysis is performed on the processed matrix to obtain the loading matrix. Then, the loading matrix is ​​grouped using the simple threshold method, K-means clustering method, or rotated principal component method. Finally, snapshot sub-matrices are created based on the grouping results.

7. The digital twin early warning method for test risk with multi-domain collaborative monitoring according to claim 6 is characterized in that: The third step is specifically as follows: Step 3.1: During the test, use the data feature judgment criteria to screen sensor data failures to screen and eliminate faulty sensors; The data feature judgment criteria include physical rationality judgment, outlier judgment, strain change trend judgment, data stability judgment, white noise fault judgment, and zero drift judgment. The data sets used for judgment include the historical sensor data set composed of historical test data and the current sensor data set composed of the current test data. In step 3.2, the sensor data filling model is constructed offline through the data filling algorithm. Online, the actual sensor data is collected for mapping, the missing data of the failed sensor is reconstructed, the data of the faulty sensor is filled, and a complete sensor data set is constructed. At the same time, the first-level failure judgment is performed.

8. The digital twin early warning method for test risk with multi-domain collaborative monitoring according to claim 7 is characterized in that: In the third step: The step 3.1 is as follows: (1) The formula for physical rationality criterion is: （4） in, is the strain value of the current sensor data set; is the maximum allowable strain value of the material; is the minimum allowable strain value of the material; (2) The criterion formula for outlier criterion is: （5） in, is the response value of the current sensor data set; is the mean of the historical sensor dataset; is the standard deviation of the historical sensor dataset; (3) The criterion formula for strain change trend is: （6） in, is the loading step sequence of the current sensor dataset; Response data sequence for the current sensor data set; for The number of data points in the (4) The criterion formula for data stability is: （7） in, for The response value of the current sensor data set under the loading level; for The response value of the current sensor data set under the loading level; is the standard deviation of the historical sensor dataset; is the threshold coefficient, usually ranging from 2 to 3; (5) The criterion formula for white noise fault criterion is: （8） in, is white noise, that is, a random signal with unknown mean and covariance; (6) The criterion formula for zero drift is: （9） in, The zero-point sensor value of the current sensor data set; The maximum value allowed for the zero point sensor; In the process of judging sensor failure, structural failure judgment is also carried out at the same time. Through the strain change trend judgment and data stability judgment in the judgment criteria, the sudden slope mutation during the loading process is further judged to complete the first-level failure judgment; In step 3.2, the data filling algorithm refers to the process of reasonably estimating and reconstructing the missing parts of the overall data using mathematical modeling or statistical methods; the data filling algorithms used include linear interpolation, singular value decomposition, interval intrinsic orthogonal decomposition, random forest method or generative adversarial network.

9. The digital twin early warning method for test risk with multi-domain collaborative monitoring according to claim 8 is characterized in that: The fourth step is specifically as follows: Step 4.1: Input the complete sensor data set obtained in step 3.2 online. Based on the reduced-order basis coefficient prediction proxy model constructed in step 2.5, obtain the reduced-order basis coefficients of each constructed reduced-order model. Multiply the reduced-order basis coefficients by the reduced-order basis to complete the raised-order inversion of the strength mechanical response field and obtain the full-field mechanical response field result. Step 4.2: Based on the complete sensor data set and the residuals of the numerical calculation of the sensor measurement point positions, a full-field response correction proxy model is constructed to perform online real-time correction on the strength mechanical response field of the specimen after the upgraded-order inversion. Specifically: Based on the coordinates of the sensors in the test patch scheme, extract the response values ​​at the corresponding coordinates in the specimen strength mechanical response field predicted in step 4.1, and calculate the deviation between the sensor predicted response value and the actual value measured by the test sensor. Construct a global response correction proxy model based on the relationship between the deviation value of each sensor and the corresponding coordinate. By superimposing the obtained global correction deviation value with the specimen strength mechanical response field obtained in step 4.1, an online dynamic correction of the global strength mechanical response field is achieved. Step 4.3: Establish a historical loading database, train a sensor timing prediction model, and predict sensor data for future loading steps to complete the second-level failure determination; the historical loading database includes simulation sample data at the measurement point and historical working condition loading sensor data; After the test, the normal sensor data collected and the faulty sensor data replaced in step 3.2 are added to the historical loading database as historical working condition loading sensor data, and the sensor time series prediction model is retrained and updated; Step 4.4: Based on the multi-level mechanical response field inversion reduction model constructed in step 2, the predicted future loading step measurement point data are input to achieve real-time inversion of the future mechanical response field. The measurement point data includes strain and displacement data, and the mechanical response field includes strain, stress, and plastic strain fields. After the test, the input measurement point data and the inverted mechanical response field are used as snapshot matrix elements to update the snapshot matrix, and the order is reduced again to achieve dynamic update of the reduced-order model; Step 4.5: Based on the inverted mechanical response field, calculate the full-field damage factor value corresponding to the adopted damage criterion, obtain the full-field damage situation of the structure, perform the third-level failure judgment, and predict the loading level at which the damage factor reaches the failure threshold; In step 4.6, after the damage factor value is calculated, the failure is determined by comparing it with the failure threshold. Based on the time-series predicted loading-strain curve, the loading level at which the damage factor value reaches the failure threshold is predicted, and an early warning is issued.

10. The digital twin early warning method for test risk with multi-domain collaborative monitoring according to claim 9 is characterized in that: In the fourth step: In step 4.2, the method for training the full-field response correction agent model includes radial basis function, Gaussian process regression, random forest, transfer learning neural network or incremental learning gradient boosting tree; The step 4.3 is specifically as follows: During the training process of the sensor time series prediction model, based on the time series data of the historical load database, the time series prediction method is used to capture the time series characteristics of the impact of load changes on the response of the measurement point, and time series prediction with load changes is performed. The time series dependency characteristics of the time series prediction algorithm are used to identify the short-term and long-term dependencies in the time series data, so that the sensor time series prediction model can still accurately predict the response of the measurement point when facing complex load changes. In the online phase, based on the established sensor time series prediction model, the sensor data set consisting of the collected normal sensor data and the replaced faulty sensor data in step 3.2 is input to achieve real-time sensor future prediction. The predicted value is combined with the first-level failure determination method in step 3.1 to achieve the second-level failure determination. In step 4.5, the damage criteria for calculating the full-field damage factor include: ductile metal damage criterion, Johnson-Cook criterion, and Hashin damage initiation criterion; The damage factor ranges from 0 to 1, and the failure threshold is usually 1. When the damage factor is equal to 0, the structure is considered to be undamaged and the stiffness has not degraded. When the damage factor is greater than 0 and less than 1, the structure is considered to be partially damaged and the stiffness begins to decrease. When the damage factor is greater than 1, the material fails completely and loses its bearing capacity.

Images