Method for diagnosing and predicting digital twinning injury of fleet by considering structural characteristics

Abstract

The invention discloses a fleet digital twin injury diagnosis and prediction method considering structural features, and relates to the technical field of structural health monitoring, and the method comprises the steps: inputting a collected monitoring signal of an individual structure in a fleet into a physical feature extraction model, and obtaining the physical features of the individual structure; the physical feature extraction model is obtained by training a physical decoding variational neural network; establishing similarity measurement among different individual structures according to the physical characteristics of the different individual structures; determining joint probability distribution of damage states among different individual structures according to similarity measurement by adopting a connection function; and taking the individual structure with the current monitoring signal output as a target individual, updating the damage state of the target individual according to the damage state of the target individual in a particle filtering-based structure damage state prediction process, and carrying out collaborative updating on the damage states of other individuals according to the damage state of the target individual and the joint probability distribution. The accuracy of damage prediction can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of structural health monitoring, and in particular to a fleet digital twin damage diagnosis and prediction method that takes structural characteristics into consideration. Background Art

[0002] In aviation structures, alternating loads during flight can cause cracks to initiate and propagate in key parts, seriously affecting the structural integrity of the aircraft. Due to differences in service load history and airframe structure manufacturing, different aircraft in the same fleet have different damage states. Therefore, the relevant unified lifespan and regular inspection model can no longer effectively guarantee safety during service, and condition-based inspection and maintenance are urgently needed. Summary of the invention

[0003] The purpose of this application is to provide a fleet digital twin damage diagnosis and prediction method that takes structural characteristics into account, which can improve the accuracy of damage prediction.

[0004] To achieve the above-mentioned purpose, the present application provides the following scheme: The present application provides a fleet digital twin damage diagnosis and prediction method considering structural characteristics, including: collecting monitoring signals of individual structures in the fleet; inputting the monitoring signals into a physical feature extraction model to obtain physical characteristics of the individual structures; the physical characteristics include damage status and structural characteristics; the physical feature extraction model is obtained by training a physical decoding variational neural network using a training set; the physical decoding variational neural network includes a feature extraction network and a signal reconstruction network connected in sequence; the feature extraction network is used to input monitoring signals and output damage status parameters and structural characteristics; the signal reconstruction network is used to input damage status parameters and structural characteristics The signal reconstruction network is a digital twin model constructed by the order reduction method; according to the physical characteristics of different individual structures, a similarity measure between different individual structures is established; a link function is used to determine the joint probability distribution of damage states between different individual structures according to the similarity measure; the individual structure with a monitoring signal output at the current moment is taken as the target individual, and in the structural damage state prediction process based on particle filtering, the damage state of the target individual is updated according to the damage state of the target individual, and the damage state of other individuals is collaboratively updated according to the damage state of the target individual and the joint probability distribution; the other individuals are individual structures that do not output monitoring signals at the current moment.

[0005] According to the specific embodiments provided by the present application, the present application discloses the following technical effects: The present application provides a fleet digital twin damage diagnosis and prediction method that takes structural characteristics into consideration, inputs the monitoring signal into a physical feature extraction model to obtain the physical characteristics of the individual structure, and takes the individual structure that has a monitoring signal output at the current moment as the target individual. In the structural damage state prediction process based on particle filtering, the damage state of the target individual is updated according to the damage state of the target individual, which effectively utilizes the monitoring signal of the target individual, reduces the uncertainty of the damage state update, and improves the accuracy of damage prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] 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.

[0007] Figure 1 A flowchart of a fleet digital twin damage diagnosis and prediction method considering structural characteristics is provided in one embodiment of the present application.

[0008] Figure 2 A principle framework diagram of a fleet digital twin damage diagnosis and prediction method considering structural characteristics provided in one embodiment of the present application.

[0009] Figure 3 A schematic diagram of the physical decoding variational neural network structure provided in one embodiment of the present application.

[0010] Figure 4 A schematic diagram of the locations and numbers of sensor measurement points of a test piece provided in one embodiment of the present application.

[0011] Figure 5 A schematic diagram of quantifying crack size during fatigue crack propagation based on a physical decoding variational neural network provided in one embodiment of the present application.

[0012] Figure 6 A schematic diagram of a skin structural feature quantification process during fatigue crack propagation based on a physical decoding variational neural network provided in one embodiment of the present application.

[0013] Figure 7 A schematic diagram of a structural feature of a long stringer during fatigue crack propagation quantified by a physical decoding variational neural network provided in one embodiment of the present application.

[0014] Figure 8 A schematic diagram showing the comparison of the results of two methods provided in one embodiment of the present application for diagnosing and predicting structural damage.

[0015] Fig. 9 Two methods for material parameters provided in an embodiment of the present application and Schematic diagram for comparing the update results.

[0016] Fig.10 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0017] 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.

[0018] 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.

[0019] This application provides a fleet digital twin damage diagnosis and prediction method considering structural characteristics, such as Figure 1 and Figure 2 The fleet digital twin damage diagnosis and prediction method considering structural characteristics includes steps 101 to 104.

[0020] Step 101: Collect monitoring signals of individual structures in the fleet.

[0021] Step 102: Input the monitoring signal into the physical feature extraction model to obtain the physical features of the individual structure; the physical features include damage status and structural features; the physical feature extraction model is obtained by training the physical decoding variational neural network using a training set; the physical decoding variational neural network includes a feature extraction network and a signal reconstruction network connected in sequence; the feature extraction network is used to input the monitoring signal and output damage status parameters and structural features; the signal reconstruction network is used to input damage status parameters and structural features and output the monitoring signal; the signal reconstruction network is a digital twin model constructed using a reduced-order method.

[0022] Step 103: establishing a similarity measure between different individual structures according to the physical characteristics of different individual structures; and using a link function to determine a joint probability distribution of damage states between different individual structures according to the similarity measure.

[0023] Step 104: Taking the individual structure with monitoring signal output at the current moment as the target individual, in the particle filtering-based structural damage state prediction process, updating the damage state of the target individual according to the damage state of the target individual, and collaboratively updating the damage state of other individuals according to the damage state of the target individual and the joint probability distribution; the other individuals are individual structures without monitoring signal output at the current moment.

[0024] The overall idea of ​​this application is: first, integrate the physics-based reduced-order signal reconstruction model and the data-driven physical feature extraction model to construct a physical decoding variational neural network. Secondly, based on the sensor monitoring signal, the physical decoding variational neural network is used to quantify the structural damage of different individuals in the fleet and accurately extract their structural characteristics. Then, based on the individual structural feature information, the connection function is used to establish the similarity measure between different individuals, and the joint probability distribution of the damage state is modeled. Finally, in the structural damage state prediction process based on particle filtering, the monitoring results of the existing structure are used to update the damage state of itself, as well as the coordinated update of the damage state of other structures. The monitoring signals of the target individuals are effectively utilized, the uncertainty of the damage state update is reduced, and the accuracy of damage prediction is improved.

[0025] In an exemplary embodiment, each sample in the training set includes damage state parameters, structural features and monitoring signals.

[0026] In the process of training the physical decoding variational neural network using the training set, the feature extraction network in the trained physical decoding variational neural network is used as the physical feature extraction model. The physical feature extraction model realizes damage state diagnosis.

[0027] The construction of the training set specifically includes: determining the structural features that affect the response of the intact structure and the damaged structure in the reduced-order signal reconstruction model based on the domain knowledge of structural integrity, and sampling the structural features. Each sample point represents a set of influencing parameters for the full-order simulation model. Then, the sampled parameters are input into the full-order simulation model. 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, structural characteristics and sensor signals obtained by simulation. The sensor signal is the monitoring signal, the full-order simulation model is the finite element model of the individual structure, and the reduced-order signal reconstruction model is the trained signal reconstruction network.

[0028] In order to reduce the impact of noise in the sensor monitoring signal on the robustness of the model, this application adds noise to the simulation signal to form a hypothetical noise signal: (1) in, and They represent the monitoring signal with noise and the monitoring signal obtained by simulation, respectively. The added noise , represents a normal distribution, is the standard deviation determined by the noise level of the structural health monitoring sensor. Perform model training. Replace the simulation data set , get the training set.

[0029] In an exemplary embodiment, the structural features and their distribution ranges are shown in Table 1.

[0030] Table 1 Structural characteristics and their distribution range

[0031] When the individual structure is a wall panel specimen, the wall panel specimen The impact parameter set consists of the following 4 parameters: .

[0032] The generation process of the helicopter rear fuselage panel simulation data set mainly includes Latin hypercube sampling, parameter model establishment in ABAQUS, automatic simulation and data analysis. Finally, a simulation data set containing 2000 samples is obtained: ABAQUS is a powerful finite element software for engineering simulation.

[0033] in, , and They respectively represent the damage state parameters, structural characteristics and monitoring signals obtained by finite element simulation of the wall panel.

[0034] In an exemplary embodiment, a training set is used to train a physical decoding variational neural network, specifically including: using the training set to train the signal reconstruction network to obtain a pre-trained signal reconstruction network, so as to establish a mapping between damage status and structural characteristics and sensor monitoring signals, and using the signal reconstruction network as a replacement for the full-order simulation model. The pre-trained signal reconstruction network replaces the signal reconstruction network in the physical decoding variational neural network to obtain a pre-trained physical decoding variational neural network; and the training set is used to train the pre-trained physical decoding variational neural network to obtain a trained physical decoding variational neural network.

[0035] According to formula (1), the physical decoding variational neural network is trained based on the assumed noise signal, and the simulation data set is divided into a training subset and a validation subset in a ratio of 0.8 / 0.2. Train to establish a mapping between damage states and structural features and sensor monitoring signals as a full-order simulation model The network has an input size of 5, representing the damage state and structural characteristics, and an output size of 20, corresponding to the number of fiber Bragg Grating (FBG) sensors used to collect monitoring signals. The pre-trained reduced-order signal reconstruction model consists of two hidden layers with 7 and 23 neurons, respectively.

[0036] Physical feature extraction model There are also two hidden layers, but with 21 and 11 neurons respectively. Compared with the pre-trained reduced-order signal reconstruction model, the input and output sizes of the physical feature extraction are reversed.

[0037] Physical decoding variational neural network such as Figure 3 As shown, Figure 3 The left part in the middle represents the data-driven physical feature extraction model , which consists of parameterized, whose structure depends on the number of monitored signals and the number of latent variables The middle part represents the physical characteristics. ,in is the damage state parameter, are structural characteristics. These latent variables are expressed by means and standard deviation Probabilistic representation is performed, and the latent variables include damage state and structural characteristics. The right part of the physical decoding variational neural network is composed of a physics-based reduced-order signal reconstruction model It takes the sampling of latent variables as input and the reconstructed sensor signal as output.

[0038] 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 the training set and the loss function defined in formula (1), the physical decoding variational neural network is trained. In this stage, the simulation data set is also divided into training subsets and validation subsets in 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 2.

[0039] Table 2 Parameter settings of the physical decoding variational neural network for panel damage quantification

[0040] Finally, a trained physical feature extraction model is obtained for identifying structural damage states and structural features. .

[0041] In an exemplary embodiment, using the training set to train the pre-trained physical decoding variational neural network specifically includes: model training based on variational inference, the variational inference method within the Bayesian framework is used to train the physical decoding variational neural network, where the physical decoding variational neural network is a pre-trained physical decoding variational neural network, and the loss function is expressed as: in, represents the loss value, represents the expected log-likelihood of the monitoring signal under the variational posterior, Indicates is the variational posterior of the parameter With prior The KL (Kullback-Leibler) divergence of To monitor the signal, For physical characteristics, It represents the neural network parameters in the physical feature extraction model, which include the weights and thresholds of neurons; is the posterior distribution, that is, Under the conditions that occur, Probability of occurrence; is the likelihood function, that is, Under the conditions that occur, Probability of occurrence; is the prior distribution, that is, in the absence of any new information, Probability of occurrence.

[0042] Physics-decoding variational neural networks simultaneously learn the parameters and posterior distribution of the inferred model in a probabilistic framework .

[0043] This application will extract the physical feature model And the reduced-order signal reconstruction model obtained by pre-training Integrate it into the physical decoding variational neural network, and use the training set and the loss function in formula (1) to train the model to obtain a trained physical feature extraction model for identifying structural damage states and structural features. .

[0044] This application inputs the monitoring signal of the individual structure into the trained physical feature extraction model , the damage state and structural characteristics of individual structures are identified online, and the related uncertainties are quantified. The specific steps are as follows.

[0045] 1) Generate test data set. In this application, a set of fatigue test data of helicopter rear fuselage panel specimens is used for verification. Panel specimens are No. 1, No. 2 and No. 3. The panel specimen simulates a piece of skin connected to the truss by rivets. Each panel specimen is equipped with 20 FBGs, such as Figure 4 As shown, FBGs are arranged along the rivet line to collect the static strain under peak load.

[0046] The three panel specimens were measured 9, 12 and 10 times respectively, and the strain measurements and corresponding crack lengths of 20 FBG sensors were recorded. In order to reduce the influence of load changes on strain measurement during fatigue crack growth, a damage index was defined based on the monitoring signal. Assuming that the rivets are arranged along the rivet direction, the strain measurement values ​​of 20 FBG sensors and the corresponding crack lengths were recorded. A strain sensor is used to collect the static strain at peak load , then, Sensor measurements The injury index Defined as: .

[0047] Then, the obtained test data set is defined as: .

[0048] in, Corresponding to the tests of three wall panel specimens, It is The crack length of each wall panel specimen is It is The damage index collection of the wall panel specimens.

[0049] 2) Use a physical feature extraction model based on a physical decoding variational neural network The damage state of the panel specimens is identified and the structural characteristics are measured.

[0050] Figure 5 The crack length identification results implemented based on FBG monitoring signals during fatigue crack growth tests are shown. Figure 5 Parts (a), (b) and (c) are the crack length identification results of wall panels No. 1 to No. 3 respectively. The crack lengths measured for the three wall panel specimens are very close to the predicted results and are basically within the 95% confidence interval, indicating that the physical decoding variational neural network used can effectively identify structural damage and give reliable uncertainty measurement results.

[0051] Furthermore, Figure 6 and Figure 7 The results of the structural characteristic calibration of three wall panel specimens are shown. Figure 6 Parts (a), (b) and (c) show the changes in skin thickness during the crack extension process of panels 1 to 3. Figure 6 Parts (d), (e) and (f) show the changes in the Young's modulus of the skin during the crack length extension process of panels No. 1 to No. 3, respectively. Figure 7 Parts (a), (b) and (c) show the changes in the thickness of the long stringer during the crack extension process of panels 1 to 3. Figure 7 Parts (d), (e) and (f) show the changes in the Young's modulus of the long stringer during the crack length extension of panels 1 to 3. The nominal dimensions and material properties of the skin and stiffeners of these three panel specimens are exactly the same. However, based on the features extracted from the monitoring signals, it is obvious that the thickness of the skin and stiffeners and the Young's modulus differ from the nominal values. These geometric and material properties are subsequently used in fleet structural similarity measurement and damage collaborative diagnosis.

[0052] In an exemplary embodiment, based on individual structural feature information, a link function is used to establish a similarity measure between different individuals, and a joint probability distribution of damage states is modeled, specifically including: 1) Use KL divergence to measure the difference between the distributions of individual structural features. The greater the difference between the distributions of two individual structural features, the greater the KL divergence. The definition of KL divergence is as follows: in, and The target distribution is and matching distribution The probability density function of . When the two distributions match exactly, , , is a random variable.

[0053] Use KL divergence to measure the similarity of the distribution of structural feature parameters and the similarity of the predicted distribution of crack lengths .

[0054] 2) Convert the KL divergence measurement result into a similarity measurement result. The basic idea is that the more similar the distributions of two individuals are, the higher the correlation coefficient will be. Conversely, when the damage states of two individuals are very different, the correlation coefficient should be reduced accordingly. Based on this concept, the conversion formula is given as follows: in, and These are all hyperparameters that need to be adjusted; represents the Sigmoid function, It is expressed as: (5) Sigmoid function is used to process .because , so it is shifted to the left and normalized to , thus converting it to the range [0,1]. Due to the characteristics of the Sigmoid function, when When becomes very large (indicating that the two individuals’ structural damage state distributions are very different), the term associated with it in formula (4) will tend to zero.

[0055] 3) Using a link function to determine the joint probability distribution of damage states between different individual structures according to the similarity measure, specifically including: using a Frank link function to approximate the relationship between damage states of individual structures in the fleet, and determining the joint probability distribution of damage states between different individual structures according to the similarity measure.

[0056] Frank's joint probability distribution is expressed as: in, express and The joint probability distribution of is the cumulative distribution function of the damage state of individual structure 1, is the cumulative distribution function of the damage state of individual structure 2, is the correlation parameter in Frank’s copula function, given by formula (4).

[0057] In addition, in order to ensure that the correlation between individual internal crack size and structural characteristic parameters and extension parameters is not lost before and after the crack size distribution is updated by the link function, it is necessary to use a Gaussian link function that is easy to construct and good at fitting multidimensional joint distributions to measure and preserve this correlation.

[0058] Particle filtering, as a Bayesian-based method, is effective in modeling crack propagation processes that contain both epistemic uncertainty and random uncertainty. However, related particle filtering-based digital twin methods mainly focus on the diagnosis and prediction of single-machine structural damage, focusing on updating the damage state of the corresponding structure using the inspection data of a specific individual, and rarely pay attention to the fleet level. In fact, although the damage states of multiple individuals in a fleet are different, they are also correlated. To this end, a simple approach is to consider the damage states of all structures in the fleet together and build a huge diagnosis and prediction model. However, as the number of individuals in the fleet increases, this approach will face the problem of dimensionality disaster, and the computational complexity and storage requirements will increase rapidly. Therefore, for the digital twin of fleet maintenance, effective means are needed to optimize the balance between the differences and similarities of fleet damage, so as to efficiently use the monitoring data in the fleet for collaborative diagnosis, thereby achieving cost reduction and efficiency improvement of fleet maintenance. However, there are still two major challenges to achieving this goal.

[0059] First, how to accurately quantify the similarities between different individuals in the fleet is a key challenge. These similarities are mainly reflected in the geometric characteristics and material properties of the structure, which need to be accurately identified. Changes in these properties often lead to differences in structural monitoring signals. Related health monitoring methods usually directly establish the relationship between monitoring signals and damage, but often ignore the impact of these structural differences. The physical decoding variational neural network integrates the reduced-order model containing damage states and individual structural characteristics into the decoder of the physical decoding variational neural network, thereby embedding physical knowledge related to structural damage in the damage quantification process based on monitoring signals. Through this method, the damage state and individual structural characteristics of the structure can be identified simultaneously and probabilistically in the analysis based on monitoring signals.

[0060] Second, after obtaining the individual features of the structures and measuring their similarities, how to achieve collaborative diagnosis and prediction of structures within the fleet based on the similarity of these features is another important challenge. When two structures have similarities, their damage states must obey a correlated joint distribution. Therefore, by observing one structure, this joint distribution can be updated, thereby synchronously updating the damage states of the two structures. The difficulty in constructing this joint distribution lies in how to effectively describe the dependency relationship between the damage distributions of the two structures. The link function decomposes the joint distribution of multiple random variables into marginal distributions and the dependency structure between them, thereby separating the marginal distribution from the dependency structure. This decomposition method provides a useful reference for establishing a joint distribution of damage states based on the similarity measure of physical decoding variational neural networks.

[0061] In an exemplary embodiment, in the particle filter-based structural damage state prediction process, the monitoring results of the existing structure are used to update the damage state of the structure itself, and to coordinate the update of the damage state of other structures. For structural damage diagnosis and prediction problems, the state of the system is expressed by the crack length. In the crack growth process, the growth rate It can be described as: (7) in, is the stress intensity factor range, are some uncertain parameters related to the crack growth process, N represents the number of load cycles, Increase for load cycles.

[0062] Put the parameters and crack length Appending them together, we can get the augmented state vector , then the state transition equation of the system can be described as: (8) The observation model of the particle filter is expressed as: (9) in, Indicates k The state vector at step , Indicates k The damage state parameters at the time of step, Indicates k The observed value of the damage state at step , Indicates k The measurement noise at step time follows a zero-mean Gaussian distribution, Indicates k The uncertainty parameter at step time, Indicates k Damage state parameters at -1 step, Indicates k Uncertainty parameter at -1 step, represents the damage state extended noise, Follows Gaussian distribution The crack extension noise satisfies , represents the variance set in the Gaussian distribution, represents parameter evolution noise, is the stress intensity factor range, is the damage state parameter of the state, is the uncertainty parameter.

[0063] In the Bayesian framework, The unknown state vector of the step The inference is based on the following observation sequence: 1) In the prediction step, according to the state variables of the previous time step and the state transition between two adjacent time steps to predict the state vector : in, Represents the observed quantity To Observable In known cases, Probability of occurrence; Represents the state vector In the known case Probability of occurrence; Represents the observed quantity To Observable In known cases, Probability of occurrence.

[0064] 2) In the update step, when the observation value When known, update the state variable The joint probability distribution of : in, Represents the observed quantity To Observable In known cases, The probability of occurrence, is the observation likelihood function, specifically Represents the state vector In known cases, Probability of occurrence; is a normalization constant, specifically Represents the observed quantity To Observable In known cases, Probability of occurrence.

[0065] Practical problems often correspond to nonlinear non-Gaussian systems, and formulas (10) and (11) are difficult to calculate analytically. In the particle filter method, Particles represent , and use sequential importance resampling to obtain its approximate solution: (12) in, is the number of particles, It's Dirac function, It is Particles, It is The importance weights of samples are given, and each particle represents a state vector. Each particle carries a set of parameters (including crack size, physical characteristics, crack extension model parameters, etc.) and uses a set of particles to represent the probability distribution of these parameters ( N The values ​​of a parameter carried by each particle constitute a distribution of the corresponding parameter).

[0066] In an exemplary embodiment, in the particle filtering-based structural damage state prediction process, the monitoring results of the existing structure are used to update the damage state of the structure itself, and the coordinated update of the damage state of other structures is realized, which specifically includes the following steps.

[0067] 1) Select the crack growth model. This example uses the Paris model to describe the crack growth process: .

[0068] in, is the half crack length, is the number of load cycles, is the stress intensity factor range, and is the unknown material parameter. Among them, The calculation is: .

[0069] in, is the geometric factor, is the far-field stress. Where, As a function of the structural geometry and the relative length of the crack, it can be expressed as , and there exists an empirical expression.

[0070] When the stringer stiffness At a certain time, Relative length to crack There is a definite relationship curve, where is the distance between the center lines of two adjacent long girders, and the stiffness of the long girders The definition is as follows: .

[0071] in, is the width of the stringer, and are the thickness of the stringer and the skin, and are the Young's modulus of the stringer and the skin, respectively. In this embodiment, and The value of is considered accurate, and , , and It is considered that it may deviate from the nominal value, and its true value needs to be calibrated through physical decoding variational neural network.

[0072] 2) Set the parameters of the particle filtering method and the correlation measurement method, as shown in Table 3.

[0073] Table 3 Parameter settings of particle filter

[0074] 3) Based on the damage quantification results, that is, the physical features extracted by the physical feature extraction model, the fleet structure damage diagnosis and prediction are carried out according to the aforementioned collaborative update method. The results compared with the traditional monomer-oriented particle filter method are as follows: Figure 8 As shown, Figure 8 Parts (a), (b) and (c) are the structural damage prediction results of wall panels No. 1 to 3 obtained by the benchmark method, and parts (e), (d) and (f) are the structural damage prediction results of wall panels No. 1 to 3 obtained by the method of the present application.

[0075] The specific process of collaborative updating in this application is as follows: 1. The damage state of the individual itself (the extraction result of the physical feature extraction model, equivalent to the observation value) is combined with the joint distribution of the damage states between different individuals (obtained based on the link function and the correlation measurement results) to transform it into the damage state of other structures (observation value). 2. Use the "observation value" of the damage state of other structures to update the damage state of other structures.

[0076] from Figure 8 It can be observed that during the entire crack propagation process, for each wall panel, when there is a monitoring signal from the sensor input, the physical decoding variational neural network will extract the individual damage state and structural characteristics from it, and perform an update step based on the damage state, reducing the uncertainty of the prediction result. Compared with the baseline method, the method proposed in this application further improves the accuracy of crack diagnosis and reduces the uncertainty of the crack propagation process through an update step based on the connection function. In particular, for the first and third wall specimens, there is a period of lack of monitoring data during the crack propagation process, but the collaborative update method can still better control the uncertainty during this period.

[0077] The root mean square error (RMSE) and correlation index (Correlation Index, ) is used to test the model prediction error and its fitting effect on the observations. In addition, the average bandwidth (ABW) is used to measure the prediction accuracy of the model.

[0078] Table 4 lists the calculation results of the three indicators. Compared with the baseline method, the method proposed in this application has lower prediction error, better fitting effect and lower bandwidth. Overall, the prediction accuracy and goodness of fit are improved by 11%, while the uncertainty of the whole process is reduced by 8%. It should be noted that the crack propagation process of the first wall panel is significantly slower than that of the latter two wall panels. On the one hand, this increases the signal-to-noise ratio of the observed data and reduces the update effect based on observation; on the other hand, due to the difference in damage size with the other two individuals, the correlation of the damage state is also low, which makes the update result given by the proposed method slightly lower than that of the other two wall panels. However, overall, the performance of the method in this application in fleet diagnosis and prediction is still better than the baseline method.

[0079] Furthermore, the present application compares the updated results of the crack growth parameters. Fig. 9 In the figure, different colors correspond to Figure 5 The number of cycles corresponding to each test measurement value changes from blue to yellow, and the layers are superimposed one by one, and the number of cycles is also superimposed one by one. Fig. 9 Parts (a), (b) and (c) are the material parameters of panels 1 to 3 obtained using the benchmark method. and The updated results, Fig. 9 Parts (e), (d) and (f) are the material parameters of wall panels No. 1 to No. 3 obtained by the method of the present application. and The updated results from Fig. 9 As can be seen in the figure, the material parameters are and The update results are better than those of the traditional monomer-oriented particle filter method. This shows that the additional update step can improve the update efficiency of the digital twin model. Overall, the uncertainty of the material parameters is significantly reduced with the update of the crack size. This reduction is more obvious when the true damage state and structural characteristics deviate greatly from the initial distribution expectation, which shows that the proposed method effectively utilizes the additional information provided by the crack size update.

[0080] The present application discloses a fleet digital twin damage diagnosis and prediction method considering structural features, which is specifically a new fleet digital twin damage diagnosis and prediction method considering structural features, and provides a promising solution for the diagnosis and prediction of fleet structural damage. The method integrates a physics-based reduced-order signal reconstruction model and a data-driven physical feature extraction model to construct a physical decoding variational neural network. Based on sensor monitoring signals and using the network, the structural damage of different individuals in the fleet is quantified and their structural features are accurately extracted. Then, based on the individual structural feature information, a similarity measure between different individuals is established using a link function, and the joint probability distribution of the damage state is modeled. Finally, in the structural damage state prediction process based on particle filtering, the monitoring results of the existing structure are used to realize the update of its own damage state, as well as the coordinated update of the damage state of other structures. The present invention effectively utilizes the structural health monitoring signals of other structures to update the damage state of other structures in the fleet, and has good scalability.

[0081] Based on the same inventive concept, the embodiment of the present application also provides a fleet damage prediction device for implementing the above-mentioned fleet digital twin damage diagnosis and prediction method considering structural features. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above-mentioned method, so the specific limitations in one or more fleet damage prediction device embodiments provided below can refer to the limitations of the fleet digital twin damage diagnosis and prediction method considering structural features above, and will not be repeated here.

[0082] In an exemplary embodiment, the present application provides a fleet damage prediction device including: The monitoring signal acquisition module is used to collect monitoring signals of individual structures in the fleet.

[0083] The physical feature extraction module is used to input the monitoring signal into the physical feature extraction model to obtain the physical features of the individual structure; the physical features include damage status and structural features; the physical feature extraction model is obtained by training the physical decoding variational neural network using a training set.

[0084] The joint probability distribution determination module of the damage state is used to establish a similarity measure between different individual structures according to the physical characteristics of different individual structures; and adopt a link function to determine the joint probability distribution of the damage state between different individual structures according to the similarity measure.

[0085] The damage state prediction module is used to take the individual structure with monitoring signal output at the current moment as the target individual, update the damage state of the target individual according to the damage state of the target individual in the structure damage state prediction process based on particle filtering, and collaboratively update the damage state of other individuals according to the damage state of the target individual and the joint probability distribution; the other individuals are individual structures without monitoring signal output at the current moment.

[0086] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Fig.10 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, 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 fleet damage prediction 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 fleet digital twin damage diagnosis and prediction method considering structural characteristics is implemented.

[0087] Those skilled in the art will understand that Fig.10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. 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 processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0088] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0089] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0090] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0091] 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).

[0092] 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 data processing logic of a programmable logic device, etc., but is not limited thereto.

[0093] 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.

[0094] 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 fleet digital twin damage diagnosis and prediction method considering structural characteristics, characterized in that: The fleet digital twin damage diagnosis and prediction method considering structural characteristics includes: Collect monitoring signals of individual structures in the fleet; The monitoring signal is input into a physical feature extraction model to obtain the physical features of the individual structure; the physical features include damage status and structural features; the physical feature extraction model is obtained by training a physical decoding variational neural network using a training set; the physical decoding variational neural network includes a feature extraction network and a signal reconstruction network connected in sequence; the feature extraction network is used to input monitoring signals and output damage status parameters and structural features; the signal reconstruction network is used to input damage status parameters and structural features and output monitoring signals; the signal reconstruction network is a digital twin model constructed using a reduced order method; Establishing a similarity measure between different individual structures according to the physical characteristics of different individual structures; using a link function to determine the joint probability distribution of damage states between different individual structures according to the similarity measure; An individual structure with a monitoring signal output at the current moment is taken as a target individual. In the structural damage state prediction process based on particle filtering, the damage state of the target individual is updated according to the damage state of the target individual, and the damage state of other individuals is collaboratively updated according to the damage state of the target individual and the joint probability distribution; the other individuals are individual structures without a monitoring signal output at the current moment.

2. The fleet digital twin damage diagnosis and prediction method considering structural characteristics according to claim 1 is characterized in that: Each sample in the training set includes damage state parameters, structural features and monitoring signals; in the process of training the physical decoding variational neural network using the training set, the feature extraction network in the trained physical decoding variational neural network is used as the physical feature extraction model.

3. The fleet digital twin damage diagnosis and prediction method considering structural characteristics according to claim 2 is characterized in that: The physical decoding variational neural network is trained using the training set, including: Using the training set to train the signal reconstruction network to obtain a pre-trained signal reconstruction network; Replacing the signal reconstruction network in the physical decoding variational neural network with the pre-trained signal reconstruction network to obtain the pre-trained physical decoding variational neural network; The pre-trained physical decoding variational neural network is trained using the training set to obtain a trained physical decoding variational neural network.

4. The fleet digital twin damage diagnosis and prediction method considering structural characteristics according to claim 3 is characterized in that: When the training set is used to train the pre-trained physical decoding variational neural network, the loss function used is expressed as: in, represents the loss value, represents the expected log-likelihood of the monitoring signal under the variational posterior, represents the KL divergence with , To monitor the signal, For physical characteristics, is the posterior distribution, is the likelihood function, is the prior distribution.

5. The fleet digital twin damage diagnosis and prediction method considering structural characteristics according to claim 1 is characterized in that: The joint probability distribution of the damage states between different individual structures is determined by using the link function according to the similarity measure, specifically including: The Frank link function is used to determine the joint probability distribution of damage states between different individual structures based on the similarity measure.

6. The fleet digital twin damage diagnosis and prediction method considering structural characteristics according to claim 5 is characterized in that: The joint probability distribution is expressed as: in, express and The joint probability distribution of is the cumulative distribution function of the damage state of individual structure 1, is the cumulative distribution function of the damage state of individual structure 2, is the correlation parameter in Frank's copula function.

7. The fleet digital twin damage diagnosis and prediction method considering structural characteristics according to claim 1 is characterized in that: The state transfer equation of the particle filter is: The observation model of the particle filter is expressed as: in, Indicates k The state vector at step , Indicates k The damage state parameters at the time of step, Indicates k The observed value of the damage state at step , Indicates k The measurement noise at step time follows a zero-mean Gaussian distribution, Indicates k The uncertainty parameter at step time, Indicates k Damage state parameters at -1 step, Indicates k Uncertainty parameter at -1 step, represents the damage state extended noise, represents parameter evolution noise, is the stress intensity factor range, Increase for load cycles.

Images