A structural damage quantification method, device and medium with embedded digital twin

Abstract

The present application discloses a structural damage quantification method, device and medium with embedded digital twins, which relates to the field of structural health monitoring. The method includes determining the influencing parameters that affect the responses of intact structures and damaged structures in a digital twin model based on the domain knowledge of structural integrity; constructing a reduced-order digital twin model based on the influencing parameters using model reduction technology; constructing a physical decoding variational neural network based on the reduced-order digital twin model and a data-driven physical feature extraction model; in the data-driven physical feature extraction model, the feature extraction results are physical features represented by the damage state and the influencing parameters; inputting the measured data of the structure to be measured into the constructed physical decoding variational neural network to achieve damage quantification of the structure to be measured and parameter calibration of the digital twin model. The present application can accurately predict the structural damage state, quantify related uncertainties and achieve model parameter calibration, while ensuring the interpretability of the results.

Description

Technical Field

[0001] The present application relates to the field of structural health monitoring, and in particular to a structural damage quantification method, device and medium with embedded digital twin. Background Art

[0002] Structural damage poses a serious threat to the safety of structural use and may lead to fatal accidents. Damage tolerance is essential to maintaining structural integrity. Under the premise that the structural damage state and its impact on operational safety can be accurately understood, the damage tolerance concept allows the structure to continue to serve despite being damaged. With the development of sensor technology, structural health monitoring technology has received more and more attention. Structural health monitoring relies on a sensor network integrated in the structure to detect, locate and quantify damage inside the structure by collecting various signals (such as strain, temperature, and vibration). The realization of these functions makes it possible to continuously monitor the health of the structure and helps to estimate the remaining life, thereby further supporting the integrity maintenance and predictive maintenance of the structure.

[0003] In recent years, data-driven structural health monitoring methods have been widely studied. These methods are generally based on the correlation between the signal and the structural damage state, directly establishing the mapping relationship between the monitoring signal and the structural damage state. Due to its simple implementation, it has become a common choice in structural health monitoring applications. However, pure data-driven methods lack the integration of physical mechanisms, which may lead to significant limitations on the interpretability of damage quantification results. In addition, factors such as geometry, material properties, and boundary constraints have a significant impact on structural damage quantification, and these influencing factors are usually difficult to handle by data-driven methods. The above limitations weaken the robustness of the structural damage quantification model and cannot effectively realize the auxiliary decision-making function. In order to address the shortcomings of pure data-driven methods, some researchers have explored methods of introducing physics-based constraints or replacing part of the data-driven model with physics-based models. However, there is currently no research that combines physics-based models with data-driven methods specifically for structural damage quantification.

[0004] By leveraging the extensive domain knowledge accumulated over decades of structural integrity research, physical mechanism-based models can be developed to account for the various factors that affect the response of damaged structures. However, running these models is usually very time-consuming. By building a reduced-order digital twin model through model reduction technology, online prediction of the response of intact and damaged structures can be achieved. At present, the combination of physics-based digital twin models and data-driven methods for structural damage quantification still faces two challenges: first, it is still challenging to effectively integrate the aforementioned influencing factors into the damage quantification process and accurately evaluate their specific impact on specific structures, but this integration is crucial to ensure the interpretability of the damage quantification results. Secondly, the inherent uncertainties in the influencing factors and monitoring signals must be quantified in the online stage to ensure the robustness of the results. Summary of the invention

[0005] The purpose of this application is to provide a structural damage quantification method, device and medium with embedded digital twin, which can accurately predict the structural damage state, quantify related uncertainties and realize model parameter calibration, while ensuring the interpretability of the results.

[0006] To achieve the above objectives, this application provides the following solutions.

[0007] In a first aspect, the present application provides a structural damage quantification method with an embedded digital twin, comprising the following steps.

[0008] Based on the domain knowledge of structural integrity, the influencing parameters that affect the responses of intact structures and damaged structures in the digital twin model are determined.

[0009] The model order reduction technology is used to construct a reduced-order digital twin model according to the influencing parameters.

[0010] A physical decoding variational neural network is constructed based on the reduced-order digital twin model and the data-driven physical feature extraction model; in the data-driven physical feature extraction model, the feature extraction result is a physical feature represented by the damage state and the influencing parameters.

[0011] The measured data of the structure to be tested is input into the physical decoding variational neural network to realize online damage quantification of the structure to be tested and parameter calibration of the digital twin model.

[0012] In a second aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the above-described methods for quantifying structural damage of an embedded digital twin.

[0013] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned methods for quantifying structural damage of an embedded digital twin.

[0014] According to the specific embodiments provided in this application, this application discloses the following technical effects.

[0015] First, the influencing parameters that affect the response of intact and damaged structures in the digital twin model are determined based on the domain knowledge of structural integrity. The reduced-order digital twin model is constructed using model reduction technology. The domain knowledge related to structural integrity is embedded in the digital twin model by explicitly representing the damage state and influencing parameters as physical features. Secondly, the reduced-order digital twin model embedded with domain knowledge and the data-driven physical feature extraction model are integrated together. The reduced-order digital twin model is used to effectively constrain and guide the training process of the physical feature extraction model, and a physical decoding variational neural network is constructed. Finally, the variational neural network based on physical decoding is trained through variational inference, and the trained physical feature extraction model is used to realize the online probabilistic damage quantification of the structure and the parameter calibration of the digital twin model. Through variational inference, the physical feature extraction model successfully realizes the quantification of the structural damage state based on the sensor detection signal, and realizes the parameter calibration of the digital twin model within the uncertainty framework.

[0016] Furthermore, this application applies physical decoding variational neural network to the preliminary exploration of structural damage state quantification, which establishes a flexible framework that can provide personalized damage assessment for different individual structures in the same group of equipment. This framework is suitable for integration into the structural digital twin system of state-based maintenance. In addition, it also provides valuable references for research in other fields and promotes the development of reduced-order digital twin models that combine domain-specific knowledge to solve inverse problems faced in different applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 Schematic diagram of the flow of a structural damage quantification method with embedded digital twin in one embodiment of the present application.

[0019] Figure 2 An architectural diagram of a physical decoding variational neural network provided in accordance with an embodiment of the present application.

[0020] Figure 3 A flowchart of structural damage quantification and parameter calibration based on a physical decoding variational neural network is provided in one embodiment of the present application.

[0021] Figure 4 A basic architecture diagram of an artificial neural network provided in one embodiment of the present application.

[0022] Figure 5 A basic flow chart for constructing a random forest is provided in one embodiment of the present application.

[0023] Figure 6 A schematic diagram of the locations and numbers of sensor measurement points of a test piece provided in one embodiment of the present application.

[0024] Figure 7 A comparison chart of crack size quantification results using a physical decoding variational neural network, an artificial neural network, and a random forest provided in an embodiment of the present application; wherein, Figure 7 (a) is a comparison of the crack size quantification results of the No. 1 wall panel using physical decoding variational neural network, artificial neural network and random forest; Figure 7 (b) is a comparison of the crack size quantification results of the No. 2 wall panel using physical decoding variational neural network, artificial neural network and random forest; Figure 7 (c) is a comparison of the crack size quantification results of the No. 3 wall panel using physical decoding variational neural network, artificial neural network and random forest.

[0025] Figure 8 A schematic diagram of the calibration of skin parameters by a physical decoding variational neural network provided in an embodiment of the present application; wherein: Figure 8 (a) is the physical decoding variational neural network for panel 1. t p Schematic diagram of the calibration; Figure 8 (b) is the physical decoding variational neural network for panel 2. t p Schematic diagram of the calibration; Figure 8 (c) is the physical decoding variational neural network for the No. 3 wall panel. t p Schematic diagram of the calibration; Figure 8 (d) is the physical decoding variational neural network for panel 1. E p Schematic diagram of the calibration; Figure 8 (e) is the physical decoding variational neural network for the No. 2 wall panel. E p Schematic diagram of the calibration; Figure 8 (f) is the physical decoding variational neural network for the No. 3 wall panel. E p Calibration diagram of .

[0026] Fig. 9 A schematic diagram of calibrating the parameters of a long stringer using a physical decoding variational neural network provided in an embodiment of the present application; wherein: Fig. 9 (a) is the physical decoding variational neural network for panel 1. t s1 Schematic diagram of the calibration; Fig. 9 (b) is the physical decoding variational neural network for panel 2. t s1 Schematic diagram of the calibration; Fig. 9 (c) is the physical decoding variational neural network for the No. 3 wall panel. t s1 Schematic diagram of the calibration; Fig. 9 (d) is the physical decoding variational neural network for panel 1. E s1 Schematic diagram of the calibration; Fig. 9 (e) is the physical decoding variational neural network for the No. 2 wall panel. E s1 Schematic diagram of the calibration; Fig. 9 (f) is the physical decoding variational neural network for the No. 3 wall panel. E s1 Schematic diagram of the calibration; Fig. 9 (g) is the physical decoding variational neural network for panel 1. t s2 Schematic diagram of the calibration; Fig. 9 (h) is the physical decoding variational neural network for the No. 2 wall panel. t s2 Schematic diagram of the calibration; Fig. 9 (i) is the physical decoding variational neural network for wall panel No. 3 t s2 Schematic diagram of the calibration; Fig. 9 (j) is the physical decoding variational neural network for panel 1. E s2 Schematic diagram of the calibration; Fig. 9 (k) is the physical decoding variational neural network for panel 2. E s2 Schematic diagram of the calibration; Fig. 9 (l) is the physical decoding variational neural network for the No. 3 wall panel. E s2 Calibration diagram of .

[0027] Fig.10 A schematic diagram of calibrating the weights of the boundary constraint mode by a physical decoding variational neural network provided in an embodiment of the present application; wherein, Fig.10(a) is a schematic diagram of the calibration of the boundary constraint mode of wall panel No. 1 by the physical decoding variational neural network; Fig.10 (b) is a schematic diagram of the calibration of the boundary constraint mode of wall panel No. 2 by the physical decoding variational neural network; Fig.10 (c) is a schematic diagram of the calibration of the boundary constraint mode of panel No. 3 using the physical decoding variational neural network. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0029] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0030] The embodiment of the present application provides a structural damage quantification method embedded with digital twins, which is executed by a computer device, and can be executed by a computer device such as a terminal or a server alone, or by a terminal and a server together. In the embodiment of the present application, Figure 1 As shown, the method includes the following steps 101-104.

[0031] Step 101: Determine influencing parameters that affect the responses of the intact structure and the damaged structure in the digital twin model based on the domain knowledge of structural integrity.

[0032] Step 102: Using model reduction technology, a reduced-order digital twin model is constructed according to the influencing parameters.

[0033] Step 103: construct a physical decoding variational neural network based on the reduced-order digital twin model and the data-driven physical feature extraction model; in the data-driven physical feature extraction model, the feature extraction result is a physical feature represented by the damage state and the influencing parameters.

[0034] Step 104: input the measured data of the structure to be measured into the physical decoding variational neural network to realize online damage quantification of the structure to be measured and parameter calibration of the digital twin model.

[0035] In an exemplary embodiment, step 102 may be replaced by the following steps 201 to 203 .

[0036] Step 201: Select the distribution range of the influencing parameter, and perform sampling in the parameter space to determine multiple sample points; each sample point represents a set of influencing parameters for the simulation model; there are many influencing parameters, each influencing parameter represents a dimension, and they together constitute a multi-dimensional parameter space, each influencing parameter takes a value on a dimension in the parameter space, and these value ranges are the distribution ranges of the influencing parameters.

[0037] Furthermore, the distribution range of the influencing parameters is selected and sampling is performed in the parameter space to cover all possible states and changes of the influencing parameters; this process generates enough sample points for subsequent batch simulations.

[0038] Step 202: input the sample points into a full-order simulation model for batch simulation to generate a simulation data set; the simulation data set includes damage state parameters, influencing parameters and sensor detection signals obtained by simulation.

[0039] Furthermore, the sampled parameters are input into the full-order simulation model The Monte Carlo simulation method is generally used to perform batch simulation. The structural response is extracted at the sensor network measurement point location and combined with the corresponding input parameters to form a simulation data set. .in, , and They respectively represent the damage state parameters, influencing parameters and sensor detection signals obtained by simulation.

[0040] Step 203: Use the simulation data set to train a reduced-order digital twin model to replace the full-order simulation model.

[0041] Furthermore, using a simulated dataset Training a Reduced-Order Digital Twin Model , in order to establish the mapping between damage state and influencing parameters and sensor detection signals, which can be used as a full-order simulation model Alternative.

[0042] It is important to note that reduced-order digital twin models can be built using a variety of machine learning methods, such as neural networks and Gaussian process regression. When the input and output of the model are complex, deep learning techniques are required. For example, convolutional neural networks can be used for image-type inputs, recurrent neural networks can be used for time series inputs, and graph neural networks can be used for graph-type inputs. When the digital twin model is complex and contains multiple reduced-order models, co-simulation methods are needed to effectively integrate these models.

[0043] In an exemplary embodiment, step 103 may be replaced by the following steps 301 - 302 .

[0044] Step 301: Input the sensor detection signal into the data-driven physical feature extraction model, probabilistically characterize the physical features through the mean and standard deviation, and sample the physical features; the data-driven physical feature extraction model is defined according to the number of the sensor detection signals and the number of physical features.

[0045] Step 302: Input the sampled physical features into the reduced-order digital twin model, output the reconstructed sensor signal, and complete the construction of the physical decoding variational neural network.

[0046] In an exemplary embodiment, step 103 further includes: training the physical decoding variational neural network through variational inference to determine the trained physical decoding variational neural network; wherein the physical feature is modeled as a probability distribution, and the posterior distribution is estimated by Bayesian theorem based on the sensor detection signal.

[0047] In an exemplary embodiment, the loss function of the physical decoding variational neural network is ;in, is the loss function, is the expected value of the log-likelihood of the data under the variational posterior, is the true posterior distribution, For sensor detection signal, For physical characteristics, is the approximate posterior distribution of the physical characteristics, is the prior distribution of the physical characteristics.

[0048] In an exemplary embodiment, the physical decoding variational neural network is trained by variational inference to determine the trained physical decoding variational neural network, specifically including: adding noise to the simulation signal to form a hypothetical noise signal; integrating the reduced-order digital twin model and the data-driven physical feature extraction model into the physical decoding variational neural network, and training the physical decoding variational neural network according to the hypothetical noise signal and the loss function to determine the trained physical decoding variational neural network.

[0049] Furthermore, the architecture of the physical decoding variational neural network is as follows Figure 2 As shown, the left part of the architecture represents the data-driven physical feature extraction model , which consists of parameterized, whose structure depends on the number of sensor detection signals and the number of physical features The middle part represents the physical characteristics. ,in, is the damage state parameter, are influencing parameters; these physical characteristics are expressed by the mean and standard deviation The right side of the architecture consists of a signal reconstruction model based on digital twins. (i.e., reduced-order digital twin model), which takes the sampling of physical features as input and the reconstructed sensor signal as output.

[0050] In the physical decoding variational neural network, the physical features Modeled as a probability distribution , rather than a definitive estimate.

[0051] Given the sensor detection signals in the training dataset, the posterior distribution It can be estimated by Bayes' theorem: .

[0052] in, is the true posterior distribution, which is a likelihood function; probability distribution Physical characteristics The prior distribution of Is the sensor detection signal The prior distribution of , also called model evidence.

[0053] Due to the complexity of the model, the true posterior distribution To solve this problem, we introduce a parameter The approximate posterior distribution of the physical characteristics represented by to approximate the true posterior distribution. Its goal is to minimize the approximate posterior (i.e., the approximate posterior distribution of the physical characteristics) ) and the true posterior distribution The KL (Kullback-Leibler) divergence between . However, directly minimizing this divergence is difficult because Involving marginal likelihood (i.e. model evidence ), which is also not tractable. Therefore, the physical decoding variational neural network chooses to optimize the evidence lower bound (ELBO) instead of the minimization process.

[0054] The expression of ELBO is: .

[0055] Among them, the first represents the expected value of the log-likelihood of the data under the variational posterior, which is usually called the mean square error loss in traditional neural networks, and is the variational posterior and the prior distribution The KL divergence between .

[0056] Physically Decoding the Loss Function of Variational Neural Networks by Maximizing the ELBO It can be expressed as: .

[0057] In this way, the physics-decoding variational neural network simultaneously learns to infer the parameters of the model and the approximate posterior distribution of the physical features in a probabilistic framework. .

[0058] Defining the loss function After that, the physical decoding variational neural network can be trained. However, in actual usage scenarios, the signals detected by the sensor often contain noise, which may affect the robustness of the model. In order to mitigate this effect, this application adds noise to the simulation signal to form a hypothetical noise signal: .

[0059] in, and denote the hypothetical noise signal and the simulated signal respectively. The hypothetical noise signal is a noisy signal, the added noise , represents a normal distribution, is the standard deviation determined by the noise level of the structural health monitoring sensor. The model assumes that the noisy signal Conduct training.

[0060] Extracting physical features from the model Compared with the pre-trained reduced-order digital twin model Integrate into the physical decoding variational neural network, using As a data set, the model is trained using the data set and loss function to obtain the trained model for structural damage quantification and model parameter calibration. .

[0061] This application also provides an application scenario to illustrate the structural damage quantification method of the embedded digital twin involved in this application.

[0062] like Figure 3 As shown, the present application proposes a structural damage quantification method of an embedded digital twin based on a physical decoding variational neural network, which includes the following steps.

[0063] Step 1) Determine the image parameters that affect the structural response based on the domain knowledge of structural integrity, and determine the influencing parameters and their distribution ranges of the intact structure and damaged structure responses that need to be considered in the digital twin model, as shown in Table 1.

[0064] Table 1 Influencing parameters and their distribution range

[0065]

[0066] The following are the weights of the boundary constraint modes in Table 1 Give a brief explanation.

[0067] Complex boundary constraints encountered in actual use scenarios, state superposition is an effective method to characterize constraints by weighted averaging of different constraint modes, superimposed simulation strain signals The calculation is as follows.

[0068]

[0069] in, Indicates The weights corresponding to the constraint modes, is the simulated strain for a single constraint mode.

[0070] For the current example, this application also considers three constraint modes, see Table 2 for details. For the three constraint modes, the degrees of freedom of the lower end of the wall panel are all fixed.

[0071] At the upper end of the panel, the constraints are as follows: the first constraint mode restricts out-of-plane translation and rotation; the second constraint mode only allows vertical translation; the third constraint mode only allows out-of-plane translation, which is a relaxed form compared to the first constraint mode.

[0072] It is worth noting that the sum of the weights of the three constraint modes is always equal to 1, which indicates that only two of the constraint modes are independent. Therefore, only the weights of the first two constraint modes are used in the influencing parameters.

[0073] Table 2 Constraints of different constraint modes on the degrees of freedom of the clamping end

[0074]

[0075] in, and In the Cartesian coordinate system, The translational and rotational degrees of freedom of the nodes in all directions.

[0076] The final set of influencing parameters Includes the following 8 parameters: .

[0077] Step 2) Build a reduced-order digital twin model of the damaged structure.

[0078] Step 2.1) Sampling of influencing parameters.

[0079] According to the influencing parameters and their distribution range determined in step 1), sampling is performed in the parameter space to cover all possible states and changes of the influencing parameters. This process generates enough sample points for subsequent batch simulations, and each sample point represents a set of influencing parameters of the input simulation model. Latin hypercube sampling is used to generate a sufficient number of data points (10,000 in this example) in the parameter space spanned by the crack size and influencing parameters to ensure that the sample distribution in the parameter space is statistically significant.

[0080] Step 2.2) Generate data sets through batch simulation.

[0081] First, a finite element model is constructed in the structural simulation software as a full-order simulation model according to the actual test conditions. The sampled crack sizes and influencing parameters are then incorporated into the script file using the parametric modeling function to generate a corresponding input file for each sampling point. The solver is then used to solve the full-order simulation model specified in the input file. Then, the strain of the simulation model is extracted at the actual layout of the fiber Bragg grating (FBG) sensor network and combined with the corresponding input parameters to form a simulation data set. Finally, outliers are removed from the simulation results to obtain a simulation data set containing 9862 samples. .

[0082] Step 3) Use a neural network to train a reduced-order digital twin model based on the noise signal .

[0083] Model is not based on a simulated dataset Instead of training, the training is based on a hypothetical noise signal, and the simulation data set is divided into training and validation subsets in a ratio of 0.8 / 0.2. Training is performed to establish a mapping between damage states and influencing parameters and detection signals and use it as a full-order simulation model The input layer of the network contains 9 neurons, representing the number of combinations of damage states and influencing parameters, while the output layer contains 20 neurons, corresponding to the number of FBGs. The trained network consists of two hidden layers with 13 and 27 neurons, respectively.

[0084] Step 4) Reduce the order digital twin model Integrate into the physical decoding variational neural network and implement model training based on variational inference.

[0085] Step 4.1) Physical feature extraction model in physical decoding variational neural network Two hidden layers are also set, but the size of the input and output of this model is opposite to that of the pre-trained reduced-order digital twin model. Similar to the operation in step 3), according to the noise level of the sensor used for structural health monitoring, the simulation data set Add noise to obtain the hypothetical noise signal for training .

[0086] Step 4.2) When the loss function is given by the formula When defined, the physical decoding variational neural network can be trained. Compared with the pre-trained reduced-order digital twin model Integrate into the physical decoding variational neural network, using As a data set and formula The loss function defined in is used to train the physical decoding variational neural network. In this stage, the simulation dataset is also divided into training and validation subsets with a ratio of 0.8 / 0.2, and then the Adam optimizer is used to optimize the network. The specific hyperparameters are detailed in Table 3.

[0087] Table 3 Parameter settings of physical decoding variational neural network for panel damage quantification

[0088]

[0089] After training, we finally get the trained model for structural damage quantification and model parameter calibration. ,Should There are two layers of the network, and they have 23 and 12 neurons respectively.

[0090] Step 5) In order to verify the superiority of this method, in this example, the physical decoding variational neural network of this application is compared with two classic pure data-driven methods: artificial neural network (ANN) and random forest (RF). Both methods are configured with parameter settings equivalent to the physical decoding variational neural network, such as Figure 4 and Figure 5 shown.

[0091] Step 5.1) Artificial neural networks are often used to develop surrogate models to construct a mapping from detection signals to damage states. However, traditional artificial neural network models neither contain physical mechanisms nor can quantify uncertainty. The artificial neural network architecture used in this example contains 2 hidden layers, which are fully connected layers with 29 and 17 neurons respectively. The front-connected input layer is used to process 20 FBG measurements, and the back-connected output layer consists of a neuron representing the crack size. In order to account for the inherent uncertainty in initialization and training, a total of 20 artificial neural network models were trained in this example, and their predictions were averaged to obtain the final result.

[0092] Step 5.2) Random forests consist of multiple decision trees, each acting as a separate model. Random forest models are constructed by randomly selecting a subset of variables and samples from the training dataset and building decision trees using a deterministic algorithm. In this example, two-thirds of the training dataset was used to develop the regression function, and the remaining one-third was used as out-of-bag samples. The optimized random forest model consists of 100 trees, each with a depth of 10 layers and 4 samples per leaf node.

[0093] Step 5.3) Generate a test data set. In this application, a set of fatigue test data of helicopter rear fuselage panel specimens is used for verification. The panel specimens are #1, #2 and #3. The panel specimen simulates a piece of skin connected to the truss by rivets. 20 FBGs are configured on the specimen, such as Figure 6 As shown, they are arranged along the rivet line to collect the static strain at peak load.

[0094] The fatigue test was performed with a sinusoidal load and an amplitude of , the stress ratio is The experimental data set contains 20 FBG strain measurement values ​​and the corresponding crack lengths. The three panels have 29, 15, and 13 measurement values ​​recorded at the same time. In order to reduce the influence of load changes on strain measurement results during fatigue crack growth, this example defines a damage index based on the detection signal. .

[0095] Assume that the rivets are arranged along A sensor is used to collect the static strain under peak load. Then, Damage index of each sensor According to its collection value Defined as: .

[0096] in, Indicates that a certain detection signal The collected values ​​of the sensors are averaged.

[0097] Then, the obtained test data set It can be expressed as: ;in, Corresponding to 3 crack growth tests, It is The crack length set of the wall panel, It is The damage index set of the wall panels.

[0098] Step 5.4) Use the test data of three panels to verify the ability of the physical decoding variational neural network to quantify the damage of the structure. The results of the artificial neural network, random forest and physical decoding variational neural network for crack size quantification are shown in Figure 5. Figure 7 shown.

[0099] The artificial neural network and random forest models were trained using a simulation data set consisting of 101 samples, where the crack sizes were uniformly distributed from 0 mm to 100 mm, and the geometric, material, and constraint parameters were set to nominal values. The measured crack sizes of the three panels were very consistent with the predictions of the physical decoding variational neural network, basically within the 95% confidence interval. In contrast, the prediction errors of the artificial neural network and random forest models were larger. This difference should be attributed to the limitations of pure data-driven methods, which fail to consider the influencing parameters and therefore have difficulty resolving the differences between the simulation model and the actual structure. In addition, over the entire damage process, the average root mean square error (RMSE) of the artificial neural network and random forest was 6.978 mm and 7.319 mm, respectively, while the RMSE of the physical decoding variational neural network was 4.652±1.870 mm. This shows that the physical decoding variational neural network is superior to the pure data-driven methods represented by artificial neural networks and random forests in terms of damage quantification accuracy.

[0100] Figure 8 and Fig. 9 The calibration results of the influencing parameters for each of the three panels are shown, which have the same nominal values ​​for geometry and material properties. Based on the physical features extracted from the test signals, the differences from the nominal values ​​of the skin and stringer thicknesses and Young's modulus are obvious. Fig.10 The calibration of the boundary constraint modes of the three panels is described. It can be seen that the weights of modes #2 and #3 are much lower than mode #1, which indicates that although the panels encountered complex constraints during the test, they are very close to the ideal constraints represented by mode #1.

[0101] The above three figures show that this application improves the interpretability of the structural damage quantification model through effective parameter calibration. Based on the calibrated parameters, a more accurate digital twin model can be obtained to produce simulation results that are very close to the actual structural behavior. In addition, it will also support personalized damage quantification of similar structures in the same group of equipment, which is a capability that traditional feature extraction methods that lack domain knowledge embedding do not have.

[0102] In an exemplary embodiment, a computer device is provided, which may be a server or a terminal. The computer device includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store structural damage quantification data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a structural damage quantification method with an embedded digital twin is implemented.

[0103] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the above method is implemented when the processor executes the computer program.

[0104] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and the computer program implements the above method when executed by a processor.

[0105] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0106] In this application, all actions to obtain signals, information or data are carried out in compliance with the relevant data protection laws and policies of the country where they are located and with the authorization given by the owner of the corresponding device.

[0107] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0108] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0109] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A structural damage quantification method with embedded digital twins, characterized in that: The structural damage quantification method of the embedded digital twin includes: Determine the influencing parameters that affect the response of intact and damaged structures in the digital twin model based on the domain knowledge of structural integrity; Using the model reduction technology, a reduced-order digital twin model is constructed according to the influencing parameters, specifically including: Selecting the distribution range of the influencing parameters, and sampling in the parameter space to determine a plurality of sample points; each sample point represents a set of influencing parameters for the simulation model; Inputting the sample points into a full-order simulation model for batch simulation to generate a simulation data set; the simulation data set includes damage state parameters, influencing parameters and sensor detection signals obtained by simulation; Using the simulation data set to train a reduced-order digital twin model to replace the full-order simulation model; According to the reduced-order digital twin model and the data-driven physical feature extraction model, a physical decoding variational neural network is constructed, which specifically includes: Inputting the sensor detection signal into the data-driven physical feature extraction model, probabilistically characterizing the physical feature through the mean and standard deviation, and sampling the physical feature; the data-driven physical feature extraction model is defined according to the number of the sensor detection signals and the number of physical features; The sampled physical features are input into the reduced-order digital twin model, and the reconstructed sensor signals are output to complete the construction of the physical decoding variational neural network; in the data-driven physical feature extraction model, the feature extraction results are physical features represented by damage states and influencing parameters; The measured data of the structure to be tested is input into the physical decoding variational neural network to realize online damage quantification of the structure to be tested and parameter calibration of the digital twin model.

2. The structural damage quantification method with embedded digital twin according to claim 1 is characterized in that: According to the reduced-order digital twin model and the data-driven physical feature extraction model, a physical decoding variational neural network is constructed, which then includes: The physical decoding variational neural network is trained by variational inference to determine the trained physical decoding variational neural network; wherein the physical feature is modeled as a probability distribution, and the posterior distribution is estimated by Bayesian theorem based on the sensor detection signal.

3. The structural damage quantification method of embedded digital twin according to claim 2 is characterized in that: The loss function of the physical decoding variational neural network is ; in, is the loss function, is the expected value of the log-likelihood of the data under the variational posterior, is the true posterior distribution, For sensor detection signal, For physical characteristics, is the approximate posterior distribution of the physical characteristics, is the prior distribution of the physical characteristics.

4. The structural damage quantification method of embedded digital twin according to claim 3 is characterized in that: Training the physical decoding variational neural network by variational inference and determining the trained physical decoding variational neural network specifically includes: Adding noise to the simulation signal to form a hypothetical noise signal; The reduced-order digital twin model and the data-driven physical feature extraction model are integrated into the physical decoding variational neural network, and the physical decoding variational neural network is trained according to the assumed noise signal and the loss function to determine the trained physical decoding variational neural network.

5. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the structural damage quantification method for the embedded digital twin according to any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the structural damage quantification method of the embedded digital twin described in any one of claims 1-4 is implemented.

Images