Digital twinning method and system for predicting complex structure state based on machine learning, and medium

Abstract

The invention discloses a digital twinning method and system for complex structure state prediction based on machine learning, and a medium. The method comprises the following steps: step 1, data acquisition; step 2, model training; step 3, state acquisition; step 4, state prediction; 5, verifying the model; and step 6, repeating the step 3 and the step 4 to realize real-time prediction of the structure response. If the geometry, material, service state and working condition parameters of the structure are changed, that is, when the parameters exceed the prediction range of the neural network in the step 2, the steps 1-5 need to be completed again to construct the digital twin system, and then the step 3 and the step 4 are used for state real-time prediction. According to the method, the data set is generated through finite element numerical simulation in the early stage, the structure state real-time prediction model is generated through the neural network, and the response state of the structure can be predicted in real time by giving real-time input data.

Description

Technical Field

[0001] The present invention relates to a real-time prediction method for the state of a large and complex structure, and particularly to a digital twin method, system and medium for predicting the state of a large and complex structure based on machine learning. Background Art

[0002] In the state prediction of large and complex structures, it is difficult to obtain the real-time service state of the structure due to a large number of degrees of freedom. The service state includes, but is not limited to, state variables such as the stress, deformation, temperature, and irradiation dose of the structure. Therefore, for large and complex structures, the digital twin model based entirely on finite element simulation in the past has low efficiency and is difficult to model and obtain practical applications. Summary of the Invention

[0003] In order to solve the problem that the finite element method has insufficient computational efficiency in the mechanical analysis of large-scale complex structures and cannot perform real-time prediction, the present invention provides a digital twin method, system and medium for predicting the state of a complex structure based on machine learning. The present invention generates a data set by using finite element numerical simulation in the early stage, and uses a neural network to generate a real-time prediction model for the structure state. By giving real-time input data, the response state of the structure can be predicted in real time.

[0004] The object of the present invention is achieved by the following technical solutions:

[0005] A digital twin method for predicting the state of a complex structure based on machine learning, comprising the following steps:

[0006] Step 1, data acquisition:

[0007] Calculate the response result data Ri of the structure under different state parameters to obtain an input database {Gi, Di, Li, Bi, Mi, Si; Ri}, where G represents geometric parameters, D represents motion parameters, L represents load parameters, B represents boundary condition parameters, M represents material parameters, S represents state parameters, i = 1, 2,..., N, N is the number of groups of state variables, it is recommended that N be more than 100. For the present invention, there can be many state parameters S for different specific application scenarios. The data before the semicolon is input data, and the data after the semicolon is output data;

[0008] Step 2, model training:

[0009] Use the data obtained in Step 1 as an input database for training to obtain the response result Ri in different input parameter spaces. The specific steps are as follows:

[0010] Step 21, create and initialize a neural network;

[0011] Step 22, calculate the output data according to the input database, compare the error and update the network parameters;

[0012] Step 23: After all input data are input into the network, record the number of learning times and check whether the neural network error at this time meets the requirement of being less than X 1 %, if not, continue learning or increase the data volume until the requirement is met, then save the network parameters at this time, and the network training is completed;

[0013] Step 3: State acquisition:

[0014] Obtain the service state {Gt, Dt, Lt, Bt, Mt, St} of the structure under the basic geometric parameters, motion parameters, load parameters, boundary conditions, material parameters, and state parameters according to the digital twin sensor, where t represents the current time, that is, the time to be predicted;

[0015] Step 4: State prediction:

[0016] According to the service state obtained in Step 3, that is, {Gt, Dt, Lt, Bt, Mt, St}, obtain the real-time state Rt through the neural network model trained in Step 2, so as to achieve accurate prediction of the structural state;

[0017] Step 5: Model verification:

[0018] Give random input parameters {Gvi, Dvi, Lvi, Bvi, Mvi, Svi}, input them into the neural network model trained in Step 2, the obtained output result is Rvi, calculate the output result R'vi of the structure under the random input parameters {Gvi, Dvi, Lvi, Bvi, Mvi, Svi} through the finite element model, and compare whether the error between the two results of Rvi and R'vi is less than X 2 %, if it is less than X 2 %, it is considered that the prediction system meets the accuracy. Otherwise, increase the input data, adjust the state parameter Sv and improve the neural network in Step 2 to retrain until the requirement is met, or increase the value of X 2 % to an acceptable larger value, and this value needs to meet the user requirements;

[0019] Step 6: Repeat Step 3 and Step 4 to achieve real-time prediction of the structural response; if the geometry, material, service state, and working condition parameters of the structure change, that is, beyond the prediction range of the neural network in Step 2, then it is necessary to re-complete Steps 1 - 5 to build the digital twin system, and then use Steps 3 and 4 for real-time state prediction.

[0020] A digital twin device for implementing the above digital twin method for predicting the state of a complex structure based on machine learning, including a data acquisition module, a machine learning module, a state acquisition module, a state prediction module, and a system verification module, where:

[0021] The data acquisition module is used to model through input parameters and conduct a large number of finite element calculations to provide a large amount of input data for building a neural network model;

[0022] The machine learning module is used to build a neural network algorithm, train and verify the neural network to obtain a reliable neural network;

[0023] The state acquisition module is used to connect to digital twin sensors to obtain the service state of the structure in real time, and record and update the time;

[0024] The state prediction module is used to execute the trained neural network to achieve accurate prediction of the structure state;

[0025] The system verification module is used to verify the neural network.

[0026] A computer device system includes a memory, a processor, and a computer program stored on the memory. The processor executes the computer program to implement the steps of the above method.

[0027] A computer-readable storage medium stores a computer program / instructions. When the computer program / instructions are executed by a processor, the steps of the above method are implemented.

[0028] A computer program product includes a computer program / instructions. When the computer program / instructions are executed by a processor, the steps of the above method are implemented.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] 1. The present invention can quickly predict the real-time state of large complex structures and mechanical systems, including but not limited to real-time prediction of aircraft states, nuclear reactor cores and numerical reactors, real-time monitoring of automobile body states, etc. The real-time state includes physical parameters such as stress, displacement, temperature, and damage amount of the structure.

[0031] 2. The present invention can overcome the problem that the existing model has a large amount of calculation and cannot be predicted in real time, that is, by putting the finite element simulation that consumes computing resources in the early stage of structure service, accumulating data through finite element simulation to provide training data for the neural network, and subsequently using a reliable neural network for digital twin and state prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flowchart for the construction and use of the prediction system;

[0033] Figure 2 It is a contour map of structural deformation under a certain typical example;

[0034] Figure 3It is a neural network prediction model;

[0035] Figure 4 It is the training process of the neural network;

[0036] Figure 5 It is the architecture design diagram of the state rapid prediction software;

[0037] Figure 6 They are data twin hardware and computer devices. Specific implementation manners

[0038] The technical solutions of the present invention will be further described below in conjunction with the accompanying drawings, but are not limited thereto. Any modification or equivalent replacement of the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention shall be covered within the protection scope of the present invention.

[0039] Specific implementation manner 1: This implementation manner provides a digital twin method for predicting the state of a large complex structure based on machine learning. The method identifies the state parameters of the complex structure, including geometric parameters, material parameters, load parameters, boundary parameters, and state parameters, etc., conducts a large number of preliminary finite element simulations under different parameter space combinations, and uses the finite element simulation results under each working condition as input data to train the neural network model. Finally, a prediction model of the structure under this series of service working conditions is formed, and the real-time response results of the structure in the service state can be obtained by giving input parameters, that is, a digital twin model of the structure is constructed. As Figure 1 shown, the specific steps are as follows:

[0040] Step 1, data acquisition: Calculate the response result data Ri of the structure under different state parameters. The relevant state parameters include: geometric parameter G, motion parameter D, load parameter L, boundary condition parameter B, material parameter M, and state parameter S. In this step, the finite element method is used to calculate the response results under different parameter combinations, that is, Ri. This step can obtain the input database {Gi, Di, Li, Bi, Mi, Si; Ri}, where i represents the number of data groups, i = 1, 2,..., N, representing the possible working condition combinations under different parameters. Before the semicolon is the input data, and after the semicolon is the output data. For different application scenarios, the parameters included in the state parameter S are different. For example, in the automotive field, the state parameter S includes fuel quantity and road surface condition; in the nuclear power field, the state parameter S includes irradiation dose and temperature, etc. For example, when predicting the pipeline deformation, the required parameter space is {material parameter, geometric parameter, state parameter; displacement}, and the displacement nephogram is as Figure 2 shown. The state parameters of the pipeline deformation working condition include temperature parameter and irradiation parameter).

[0041] Step 2, Model Training: Use the data obtained in Step 1 as the input database {Gi, Di, Li, Bi, Mi, Si; Ri} for training to obtain the response result Ri of {Gi, Di, Li, Bi, Mi, Si} under different input parameter spaces. According to the calculation accuracy requirements, the deviation of this model is less than a specific error X 1 %, temporarily take 0.01%, and its range can be taken as 10 -10 ~10 -1 .

[0042] The neural network model to be used in this step is shown in Figure 3 . For the sake of generality, the present invention adopts a fully connected neural network, a convolutional neural network, a sparse neural network or a random neural network, etc. with multiple input layers, multiple output layers and multiple hidden layers, where: the input layer, the hidden layer, the output layer and the connection method of the network can be selected as needed; network information such as activation functions and loss functions can be selected as needed; the initialization method includes but is not limited to random initialization, etc. The model of a general neural network is as shown in Figure 3 .

[0043] The training process of the neural network is shown in Figure 4 : After preparing the input data set, first initialize the neural network, including but not limited to random initialization; calculate the output data according to the input data set, compare the errors and update the network parameters, and the solvers used include but are not limited to methods such as SGD or ADAM, etc.; after all input data are input into the network, record the number of learning times and check whether the error of the neural network at this time meets the requirement of being less than X 1 %. If not, continue learning or increase the data volume until the requirement is met, and then save the network parameters at this time. The network training is completed.

[0044] Step 3, State Acquisition: Obtain the service states {Gt, Dt, Lt, Bt, Mt, St} of the basic geometric parameters, motion parameters, load parameters, boundary conditions, material parameters and state parameters of the structure according to the digital twin sensor, where t represents the current time.

[0045] Step 4, State Prediction: According to the service states {Gt, Dt, Lt, Bt, Mt, St} obtained in Step 3, obtain the real-time state Rt through the trained neural network model, so as to realize the accurate prediction of the structure state.

[0046] Step 5, Model Verification: By inputting random input parameters {Gvi, Dvi, Lvi, Bvi, Mvi, Svi} into the model trained in Step 2, the output result obtained is Rvi. Calculate the output result R’vi of the structure under the random input parameters {Gvi, Dvi, Lvi, Bvi, Mvi, Svi} through the finite element model, and compare whether the error between the two results of Rvi and R’vi is less than X 2 %, X 2 ranges from 10 -10 to 10 -1 , if it is less than X 2 %, it is considered that the prediction system meets the accuracy. Otherwise, it is necessary to consider increasing the input data, improving the neural network in Step 2 and retraining until the requirements are met, or appropriately increasing the value range of X 2 %, and at this time, it should be noted that the increased value is acceptable to users.

[0047] Step 6, After the system is constructed, repeating Step 3 and Step 4 can realize the real-time prediction of the structural response. If the parameters such as the geometry, material, service state, and working conditions of the structure change, that is, beyond the prediction range of the neural network in Step 2, it is necessary to re-complete Steps 1 - 5 to construct the digital twin system, and then use Steps 3 - 4 to conduct real-time state prediction. In Step 1, the finite element method is used to conduct mechanical simulation on the structure, and then a calculation network is generated through methods such as digital twin and machine learning. Finally, the real-time state of the structure can be simulated, and life prediction and failure analysis can be carried out. For example, if the pipeline material changes from zirconium alloy to 314 stainless steel, the neural network trained for zirconium alloy originally should be retrained, that is, finite element simulation is carried out using 316 stainless steel to obtain a large amount of data, a reliable neural network is constructed based on this data, and the deformation prediction and digital twin of the pipeline made of 316 stainless steel material are realized.

[0048] Without loss of generality, the following method can be applied to construct the digital twin model of complex structures and build the digital twin software and device equipment. The specific software architecture design of this digital twin model is as Figure 5 shown: According to specific usage scenarios, construct a digital twin software including but not limited to the following five modules:

[0049] 1. Data Acquisition Module: Model through input parameters and carry out a large number of finite element calculations to provide a large amount of input data for the neural network model. Although this process takes a long time and has a large amount of calculation, this step is a preliminary work, and the data obtained can provide input for the subsequent neural network.

[0050] 2. Machine learning module: Used to build neural network algorithms, including but not limited to training and validating neural networks using the Python language (machine learning libraries such as TensorFlow or PyTorch) to obtain a reliable neural network.

[0051] 3. State acquisition module: Used to connect to digital twin sensors to obtain the service state of the structure in real time, and record and update the time.

[0052] 4. State prediction module: Used to execute the trained neural network to achieve accurate prediction of the structure state.

[0053] 5. System verification module: The verification module is used to verify the neural network. It is planned to use random state parameters and finite element results for verification; note that in this software design, attention should be paid to the construction and initialization of the application scenario to be applied to a specific scenario.

[0054] Specific implementation method two: On the basis of specific implementation method one, this implementation method provides a computer device system, including a memory, a processor, and a computer program stored on the memory. The processor executes the computer program to implement the steps of the method described in specific implementation method one.

[0055] Specific implementation method three: On the basis of specific implementation method one, this implementation method provides a computer-readable storage medium, on which a computer program / instructions are stored. When the computer program / instructions are executed by a processor, the steps of the method described in specific implementation method one are implemented.

[0056] Specific implementation method four: On the basis of specific implementation method one, this implementation method provides a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processor, the steps of the method described in specific implementation method one are implemented.

Claims

1. A digital twin method for complex structural state prediction based on machine learning, characterized in that The method comprises the following steps: Step 1: Data acquisition: Calculate the response result data Ri of the structure under different state parameters and obtain the input database {Gi, Di, Li, Bi, Mi, Si; Ri}, where G represents geometric parameters, D represents motion parameters, L represents load parameters, B represents boundary condition parameters, M represents material parameters, S represents state parameters, i = 1, 2, ..., N, N is the number of groups of state quantities, the data before the semicolon is the input data, and the data after the semicolon is the output data; Step 2: Model training: The data obtained in step 1 is used as the input database for training to obtain the response results Ri under different input parameter spaces; Step 3: Status acquisition: According to the digital twin sensor, the basic geometric parameters, motion parameters, load parameters, boundary conditions, material parameters, and service status of the state parameters {Gt, Dt, Lt, Bt, Mt, St} of the structure are obtained, where t represents the current time; Step 4: Status prediction: According to the service status {Gt, Dt, Lt, Bt, Mt, St} obtained in step 3, the real-time status Rt is obtained through the neural network model trained in step 2, thereby realizing accurate prediction of the structural status; Step 5: Model verification: Given random input parameters {Gvi, Dvi, Lvi, Bvi, Mvi, Svi}, input them into the neural network model trained in step 2, and obtain the output result Rvi. Calculate the output result R'vi of the structure under the random input parameters {Gvi, Dvi, Lvi, Bvi, Mvi, Svi} through the finite element model, and compare whether the error between the two results Rvi and R'vi is less than X2%. If it is less than X2%, it is considered that the prediction system meets the accuracy. Otherwise, increase the input data, adjust the state parameter Sv, and improve the neural network in step 2 to retrain until the requirements are met, or increase the value of X2%, and the improved value needs to meet the user requirements. Step 6. Repeat steps 3 and 4 to achieve real-time prediction of structural response. If the geometry, material, service status, and operating parameters of the structure change, that is, if they exceed the prediction range of the neural network in step 2, it is necessary to re-complete steps 1 to 5 to build the digital twin system, and then use steps 3 and 4 to perform real-time status prediction.

2. The digital twin method for complex structure state prediction based on machine learning according to claim 1 is characterized in that The specific steps of step 2 are as follows: Step 21, create and initialize the neural network; Step 22: Calculate output data based on the input database, compare errors and update network parameters; Step 23, after all input data are input into the network, record the number of learning times and check whether the neural network error at this time meets the requirement of less than X1%. If not, continue learning or increase the amount of data until the requirement is met, save the network parameters at this time, and the network training is completed.

3. The digital twin method for complex structure state prediction based on machine learning according to claim 2 is characterized in that The value range of X1% is 10 -10 ~10 -1 .

4. The digital twin method for complex structure state prediction based on machine learning according to claim 1 is characterized in that The value range of X2% is 10 -10 ~10 -1 .

5. A digital twin device for implementing the digital twin method for predicting the state of a complex structure based on machine learning as described in any one of claims 1 to 4, characterized in that The device includes a data acquisition module, a machine learning module, a state acquisition module, a state prediction module and a system verification module, wherein: The data acquisition module is used to model the model by inputting parameters and carrying out a large number of finite element calculations, thereby providing a large amount of input data for building a neural network model; The machine learning module is used to build a neural network algorithm, train and verify the neural network, and obtain a reliable neural network; The state acquisition module is used to connect to the digital twin sensor to obtain the service status of the structure in real time, and record and update the time; The state prediction module is used to execute the trained neural network to achieve accurate prediction of the structural state; The system verification module is used to verify the neural network.

6. A computer device system, comprising a memory, a processor and a computer program stored in the memory, characterized in that The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

8. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images