A modeling method for composite material curing process

Abstract

A composite material curing process modeling method can realize the mapping of curing process to the state information of the curing process of composite parts. It is characterized in that, first, a set of curing process data is obtained, then the state information of the composite parts in the curing process under each curing process is obtained, and then a neural operator model is trained using the above sample data, and the mapping of the curing process to the state information of the curing process of composite parts is realized through four processes of dimensionality increase mapping, iterative kernel integration, dimensionality reduction mapping, and physical constraints. The present invention can be widely used in composite material curing process modeling, and the constructed model can be used for rapid prediction of curing state information, curing process optimization, curing process monitoring, etc.

Description

Technical Field

[0001] The present invention relates to the technical field of composite material curing and molding, in particular to a composite material curing process, specifically a composite material curing process modeling method which inputs a curing process and outputs curing process state information. Background Art

[0002] In recent years, composite materials have been widely used in aerospace, automobile, shipbuilding and other fields due to their light weight and high strength. Composite material curing molding refers to the process of using a certain curing process to cause a series of physical and chemical changes inside the composite material and finally forming it. It is a key link in the production and manufacturing process of composite materials.

[0003] The curing process directly determines the evolution of the temperature field and curing degree field of composite parts during the curing process, so it must be carefully designed. If the design is improper, incomplete curing, severe temperature overshoot, residual stress and excessive deformation will occur, which will greatly reduce the quality of composite molding. In order to circumvent the above problems, the traditional physical model method is based on the finite element method, finite difference method or commercial simulation software, etc., by setting parameters such as geometric shape, boundary conditions, initial conditions and mesh division, to simulate and obtain the temperature, curing degree and other state information of the composite parts during the curing process under a given curing process. The physical model method can obtain high-fidelity curing process state information, but it needs to be re-simulated for new curing processes. For large and complex composite parts with a long single simulation time, the physical model method cannot meet the actual needs in terms of simulation efficiency.

[0004] The emerging data-driven method can use the high-fidelity simulation data obtained by the curing process data and the physical model method to train an efficient and accurate proxy prediction model. However, whether it is a neural network or other machine learning methods, it can only realize the mapping learning from low-dimensional or finite-dimensional input to low-dimensional or finite-dimensional output. Therefore, most methods simplify the input curing process into several features, such as heating rate, insulation temperature, insulation time, etc., and simplify the output curing process state information into several key process state indicators, such as thermal hysteresis and exothermic peak of the curing process temperature state information. Although this type of method has high prediction efficiency, it is obviously simplified and cannot truly achieve the modeling goal of inputting the curing process and outputting the curing process state information. Therefore, the present invention proposes a composite material curing process modeling method, which can use sample data and physical prior knowledge to train a neural operator model that meets physical constraints, and realize efficient prediction from the curing process to the curing process state information of composite parts. Summary of the invention

[0005] The purpose of the present invention is to solve the problems of long modeling cycle and low simulation efficiency of existing composite material curing process models. A method for modeling a composite material curing process is provided to realize the mapping of curing process to state information of composite parts during curing process. The constructed model can be used for rapid prediction of curing state information, curing process optimization, curing process monitoring, etc.

[0006] The technical solution of the present invention is:

[0007] A composite material curing process modeling method, characterized in that it comprises the following steps:

[0008] First, obtain a set of curing process data N is the number of curing processes, each curing process x i (t) includes temperature process information T(t) and / or pressure process information P(t), where T(t) and P(t) represent the temperature and pressure at each time point in the curing process, respectively; then, obtain each curing process x i (t) The state information y of the composite part during the curing process i (t), y i (t) including one or more of the temperature, degree of cure, stress, strain, stiffness, compliance and deformation at each point or certain specified points of the composite part during the curing process; finally, using the sample data And physical prior knowledge is used to train a neural operator model that satisfies physical constraints to achieve the mapping of the curing process to the state information of the composite part curing process.

[0009] The method for acquiring the state information of the solidification process may be: solving by finite element method; solving by finite difference method; solving by commercial simulation software or monitoring the solidification experiment process.

[0010] The neural operator model that satisfies physical constraints is a neural network that approximates operator mapping, and the network includes four parts: dimensionality increase mapping, iterative kernel integration, dimensionality reduction mapping and physical constraints.

[0011] The dimension-raising mapping is to use one or more fully connected layers or convolutional layers to transform x i (t) Mapping into a higher-dimensional implicit representation can be seen as a local operator that can improve the representation ability of the network;

[0012] The iterative kernel integral is a kernel convolution that uses multiple iterative solvers to approximate an unknown function, which can be regarded as a nonlinear operator. The iterative kernel integral in the content of the present invention can be implemented by using one or more Fourier layers. Each Fourier layer includes five parts: fast Fourier transform, low-pass filtering, linear transformation, inverse fast Fourier transform, and nonlinear activation processing. Among them, the fast Fourier transform is used to quickly obtain the frequency domain feature representation of the time domain function. Low-pass filtering is used to retain low-frequency features and remove high-frequency features. Linear transformation is used to process the low-frequency features retained by low-pass filtering, which can learn highly nonlinear features in the original space and ensure that the number of network parameters is independent of the discrete resolution of the time domain function. Inverse fast Fourier transform is used to quickly convert the low-frequency features after linear transformation back to the time domain function representation. Nonlinear activation processing is used to increase the nonlinear expression ability of the network.

[0013] The dimension reduction mapping is to use one or more fully connected layers or convolutional layers to reduce the dimension of the implicit representation after the iterative kernel integration and output the prediction result, which can be regarded as a local operator;

[0014] Physical constraints refer to constraints on network outputs, which can be reflected in the network loss function. The physical constraints can include the relationship between the curing process and the curing process state information, such as the similar trends between the temperature curing process and the temperature state information, which can introduce a residual model; it can include the physical properties of the curing state information itself, such as the curing degree state information is continuous and monotonically increasing in time, and the temperature state information is continuous in time; it can include the physical equations satisfied by the curing state information, such as the geometric equilibrium equations and constitutive equations satisfied by the stress and strain state information.

[0015] The beneficial effects of the present invention are:

[0016] The composite material curing process modeling method proposed in the present invention can use sample data and physical prior knowledge to train a neural operator model that satisfies physical constraints, thereby realizing efficient prediction directly from the curing process to the curing process state information, laying a foundation for curing process optimization, curing process monitoring, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is an implementation flow chart of the present invention.

[0018] Figure 2 It is a double insulation platform curve diagram of the temperature curing process in a specific embodiment of the present invention.

[0019] Figure 3 It is a "part-mold" system in a specific embodiment of the present invention.

[0020] Figure 4 Schematic diagram of the neural operator model that satisfies physical constraints described in the present invention.

[0021] Figure 5 Schematic diagram of a single Fourier layer in the iterative kernel integration described in the present invention.

[0022] Figure 6 A schematic diagram of a neural operator model satisfying physical constraints for predicting temperature state information in a specific embodiment of the present invention.

[0023] Figure 7 This is a diagram of the prediction effect of a neural operator model that meets physical constraints in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0024] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0025] like Figure 1-7 shown.

[0026] A method for modeling a composite material curing process is used to realize the mapping of the curing process to the curing process state information of the composite part. The model built can be used for rapid prediction of curing process state information, curing process optimization, curing process state monitoring, etc. It specifically includes the following steps (such as Figure 1 shown):

[0027] Step 1: Obtain a set of curing process data

[0028] Each curing process x i (t) includes temperature process information T(t) and / or pressure process information P(t), T(t) and P(t) respectively represent the temperature and pressure at each time point in the curing process, and N is the number of curing processes. In this embodiment, the curing temperature information is considered, and the curing temperature information is calculated as follows: Figure 2 Take the temperature curve of the double insulation platform shown in the figure as an example. Figure 2 Each parameter of the middle curve is randomly sampled within a given range to generate 50 sets of temperature process curves, that is, N = 50. The parameter sampling range is as follows:

[0029] parameter <![CDATA[t1]]> <![CDATA[t2]]> <![CDATA[t3]]> <![CDATA[t4]]> <![CDATA[t5]]> <![CDATA[T0]]> <![CDATA[T1]]> <![CDATA[T2]]> scope (20,80) <![CDATA[(t1+20,110)]]> <![CDATA[(t2+20,150)]]> <![CDATA[(t3+20,190)]]> 222 293 (50,120) (100,240)

[0030] For each generated temperature process curve, discrete sampling is performed at any sampling interval in the time domain to obtain x i (t), in this example, the total curing process time is 222min, the sampling interval is 1min, x i (t) is a one-dimensional vector of length 223.

[0031] Step 2: Get each curing process x i (t) The state information y of the composite part during the curing process i (t);

[0032] The method for obtaining the state information of the curing process can be: solving by finite element method; solving by finite difference method; solving by commercial software or monitoring the curing experiment process. In this embodiment, the commercial multi-physics simulation software COMSOL Multiphysics is used to obtain the state information of the composite part during the curing process, and the curing process of a one-dimensional flat composite part is considered, such as Figure 2 This is a schematic diagram of the “part-mold” system. The overall thickness of the system is 50 mm, and the thickness of the Invar steel mold is L t =20mm, thickness of composite part L c =30mm, such as Figure 3 As shown. The matrix of the composite material is 8552 epoxy resin, the reinforcement phase is AS4 carbon fiber, and the resin volume fraction is 42.6%. The initial condition of the system is T0 = 293K, and the boundary conditions act on the upper and lower surfaces of the "part-mold". In this embodiment, the third type of boundary conditions, namely the convection heat transfer boundary conditions, are considered, which are expressed as:

[0033]

[0034]

[0035] Where h c =120W / (m 2 K) and h t =70W / (m 2 ·K) are the convective heat transfer coefficients of the upper and lower surfaces, k c and k t represent the thermal conductivity of the composite material and the Invar steel mold, respectively, where T a For a given temperature process information.

[0036] The curing process status information y i (t) may include one or more of the temperature, curing degree, stress, strain, stiffness, flexibility and deformation at each point or certain specified points of the composite part during the curing process. This embodiment simulates and obtains 50 sets of temperature state information at x = 35 mm under the curing process, and performs discrete sampling in the time domain with the same sampling interval as the curing process, that is, the sampling interval is 1 min, y i (t) is a one-dimensional vector of length 223.

[0037] Step 3: Using sample data Train a neural operator model to map the curing process to the state information of the composite part curing process;

[0038] The neural operator model is a neural network that approximates the operator mapping. Figure 4It includes four parts: dimensionality increase mapping, iterative kernel integration, dimensionality reduction mapping and physical constraints.

[0039] The dimension-raising mapping is to use one or more fully connected layers or convolutional layers to transform the one-dimensional (single-channel) x i (t) Dimensional mapping is performed to multi-dimensional (multi-channel) data. In this embodiment, a fully connected layer is used for dimensional mapping, which is expressed as:

[0040]

[0041] Iterative kernel integration uses a single or multiple Fourier layers to process the data after dimensional mapping. Each Fourier layer is as follows Figure 5 The figure shows five parts including fast Fourier transform, low-pass filtering, linear transformation, inverse fast Fourier transform and nonlinear activation processing. In this embodiment, 4 Fourier layers are used to complete iterative kernel integration, and each low-pass filtering retains the first 16 low-frequency features, that is, k=16.

[0042] Dimensionality reduction mapping is to use one or more fully connected layers or convolutional layers to reduce the dimension of the data after iterative kernel integration into a one-dimensional (single-channel) output. In this embodiment, a fully connected layer is used for dimensionality reduction mapping, which is expressed as:

[0043]

[0044] Adding physical constraints needs to consider the characteristics of the predicted state information of the curing process. When inputting the temperature curing process and predicting the temperature state information, because the temperature state information is continuous in time domain and the temperature state information is similar to the temperature process trend, the second-order derivative constraint can be minimized to ensure prediction continuity, and the residual constraint can be used to ensure prediction accuracy; when predicting the degree of curing, a monotonically increasing constraint in the time domain can be added to ensure that the prediction is reasonable; when predicting stress and strain, geometric equilibrium equations and constitutive equation constraints can be added to ensure prediction accuracy. The added physical constraints will be reflected in the calculation of the network loss function. In this embodiment, taking temperature state information as an example, the residual constraint is incorporated into the neural operator model, and the second-order derivative is minimized in the loss function to ensure prediction continuity, such as Figure 6 As shown, L, P and M represent the dimension increase mapping, dimension reduction mapping and physical constraint parts respectively.

[0045] In this embodiment, 50 groups of sample data Using the classic neural network training method, 200 iterations of training were performed to obtain a prediction model for the curing process state information of composite parts. The test results on the new curing process are as follows: Figure 7 shown.

[0046] The parts not involved in the present invention are the same as the prior art and are implemented by using the prior art.

Claims

1. A composite material curing process modeling method, characterized in that: It includes the following steps: First, obtain a set of curing process data N is the number of curing processes, each curing process x i (t) includes temperature process information T(t) and / or pressure process information P(t), where T(t) and P(t) respectively represent the temperature and pressure at each time point during the curing process, and N is the number of curing processes; Secondly, obtain each curing process x i (t) The state information y of the composite part during the curing process i (t), y i (t) one or more of the following quantities: temperature, degree of cure, stress, strain, stiffness, compliance and deformation at each or certain specified points of the composite part during the curing process; Third, using sample data A neural operator model that satisfies physical constraints is trained with physical prior knowledge to achieve the mapping of curing process to the state information of the curing process of composite parts; The neural operator model satisfying physical constraints is a neural network that approximates operator mapping, and the network includes four parts: dimensionality increase mapping, iterative kernel integration, dimensionality reduction mapping, and physical constraints; the physical constraints are constraints imposed on the network output, which are reflected in the network loss function; the physical constraints include the relationship between the curing process and the state information of the curing process, the physical properties of the state information itself, and the physical equations satisfied by the state information; The iterative kernel integration uses one or more Fourier layers to process the data after dimensional mapping in sequence; each Fourier layer includes five parts: fast Fourier transform, low-pass filtering, linear transformation, inverse fast Fourier transform, and nonlinear activation processing; The fast Fourier transform is used to quickly obtain the frequency domain feature representation of the time domain function; the low-pass filter is used to retain the low-frequency features and remove the high-frequency features; the linear transform is used to process the low-frequency features retained by the low-pass filter, which can learn the highly nonlinear features in the original space and ensure that the number of network parameters is independent of the discrete resolution of the time domain function; the inverse fast Fourier transform is used to quickly convert the low-frequency features after the linear transformation back to the time domain function representation; the nonlinear activation processing is used to increase the nonlinear expression ability of the network.

2. The method according to claim 1, characterized in that The method for acquiring the state information of the solidification process is: solving by using a finite element method, a finite difference method, or commercial simulation software, or obtaining the information by monitoring the solidification experiment process.

3. The method according to claim 1, characterized in that The dimension-raising mapping is to use one or more fully connected layers or convolutional layers to enhance x i (t) dimension.

4. The method according to claim 1, characterized in that: The dimensionality reduction mapping is to use one or more fully connected layers or convolutional layers to reduce the dimensionality of the data after iterative kernel integration and output the prediction result.

Images