Method for predicting temperature and curing degree of large-size blade through coupling of finite element analysis and mechanical learning
Through the coupling method of finite element analysis and mechanical learning, a temperature and curing degree prediction model of large-sized blades of composite materials was established, which solved the problem of uneven temperature field during the forming of the hot press tank, and achieved accurate prediction of temperature and curing degree and optimize, improving molding accuracy and life.
Patent Information
- Application Number
- CN202411204411.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-29
- Publication Date
- 2024-11-15
AI Technical Summary
During the molding process of large-sized blades of composite materials, the temperature field is uneven due to thickness differences, resulting in internal residual stress and deformation, reducing the molding accuracy and service life. It is difficult for the prior art to effectively predict and optimize the temperature and curing degree.
Using the method of coupling finite element analysis and mechanical learning, the finite element model of large-size blades of composite materials is established through ABAQUS software, and a user subroutine is written to realize thermal-chemical model calculation, establish a heating curve for sub-region heating, and use LSTM neural network and Gray wolf optimization algorithm optimizes the temperature and curing degree prediction model to optimize temperature field uniformity.
It effectively reduces the temperature field unevenness of large-sized blades of composite materials, shortens the curing cycle, improves the molding accuracy and service life, and improves the accuracy and stability of temperature and curing degree prediction.
Smart Images

Figure CN118965910A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to simulation molding of a large-sized composite blade, and in particular to a method for predicting the temperature and curing degree of a large-sized blade by coupling finite element analysis with machine learning. Background Art
[0002] Thermosetting resin-based composite materials have a series of excellent properties such as strong designability, high specific strength and modulus, strong fatigue resistance, heat resistance, corrosion resistance and good overall forming performance, so they are widely used in aviation, aerospace, weapons, automobiles and other fields. However, when using the autoclave molding process for curing, there are still many problems to be solved. The uneven temperature field during the molding process of the part is a relatively serious problem. Due to the anisotropic thermal conductivity, the thermal conductivity of thermosetting composite materials along the fiber direction is often greater than the thermal conductivity perpendicular to the fiber direction, which makes it difficult to achieve uniform distribution of the temperature field during the curing process of the part. The larger the size of the part and the thicker the cross-sectional thickness, the more serious the unevenness of the temperature field. In addition, the larger the size of the part, the longer the curing time is usually, which increases the production cycle. Due to the large size of the composite large-sized blade and the large difference in thickness of the blade body, extension root and tenon root, during the molding process, the heat exchange conditions between the blade surface and the air in the tank are good, and the temperature rises quickly. On the contrary, due to the thickness difference inside the blade, the blade body area is faster, and the extension root area and tenon root area heat up slowly, which causes a large temperature difference inside the blade and inconsistent curing time. In the later stage of curing, the thermosetting resin undergoes a violent chemical reaction during the curing process, and a large amount of heat accumulates in the root area and the tenon area, exacerbating the unevenness of the temperature field. The uneven temperature field will cause residual stress inside, and the large-sized blades of composite materials will produce large warping deformation because the thickness of the blade area is much smaller than the tenon area. The blade body is much larger than the root and tenon, which will also increase the deformation, and ultimately reduce the molding accuracy and service life of the large-sized blades of composite materials. Therefore, in order to reduce the unevenness of the temperature field of large-sized blades of composite materials during the autoclave curing process, it is necessary to predict the temperature and curing degree of the workpiece in advance during the curing process. When the predicted temperature and curing degree gradient is too large, the input parameters of the autoclave should be adjusted in time to ensure product quality.
[0003] At present, most of the research objects are composite materials with equal thickness, and the problem of temperature field optimization of large-sized composite materials with variable thickness is not considered. In addition, most of them establish a neural network proxy model of temperature, and few of them link the degree of curing with temperature to establish a proxy model. Due to the large thickness difference of large-sized composite blades, a regional optimization method is proposed. A heat source is added to the root area of large-sized composite blades. The temperature curve is adjusted by considering the influence of the degree of curing field on the temperature field. Considering that large-sized composite blades are large in size and have large thickness differences in the blade body, extension root and root, the numerical simulation is more complicated and the curing cycle is longer. LSTM (Long Short-Term Memory) neural network is used to establish proxy models of temperature and degree of curing respectively and combine the two models. The Grey Wolf Optimization Algorithm (GWO) is used to optimize the series LSTM network to achieve the coupling of finite element thermochemical process model and artificial neural network. Summary of the invention
[0004] Due to the large size of composite large-sized blades and the characteristics that the thickness is gradually reduced from the root to the blade body, and the thickness is thick in the middle and thin on both sides, the numerical simulation using finite element software is more complicated and takes a lot of time. Curing based on the traditional method of optimizing the temperature curve will produce a large uneven temperature field in the root area. In order to solve this problem, the present invention proposes a method of coupling finite element analysis and machine learning to predict the temperature and curing degree of large-sized blades by establishing a regional optimization simulation analysis model and building a series LSTM neural network agent model.
[0005] A method for predicting temperature and curing degree of large-size blades by coupling finite element analysis with machine learning, comprising:
[0006] A: Based on the finite element software ABAQUS and the finite element modeling method, a finite element model of a large-sized composite blade is established. The heat transfer model and the curing kinetics model are coupled through temperature, and the calculation of the thermal-chemical model is realized by writing a user subroutine, and finally a numerical simulation model of the temperature field curing of a large-sized composite blade is established.
[0007] B: Based on the relationship between temperature and degree of curing, the second heating curve is changed to a segmented type, and the method of regional heating is adopted to increase the heating temperature of the tenon root area in the second heating stage.
[0008] C: Establish a series LSTM neural network, collect the curing temperature and curing degree data of large-size composite blades, set a sliding window, establish temperature and curing degree data sets respectively, and divide them into training set and test set.
[0009] D: Use the Grey Wolf Algorithm (GWO) to optimize the LSTM model, optimize the initial learning rate, the number of hidden layer neurons, the batch size, and the number of training iterations to obtain the best network hyperparameter values of LSTM;
[0010] E: Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R 2 ) to verify the prediction accuracy, reliability and stability of the LSTM model by the optimization algorithm.
[0011] in:
[0012] Step A: Based on the finite element software abaqus and the finite element modeling method, a finite element model of a large-sized composite blade is established. The heat transfer model of the large-sized composite blade is coupled with the curing kinetics model through temperature, and the calculation of the thermal-chemical model is realized by writing a user subroutine, and finally a numerical simulation model of the temperature field curing of a large-sized composite blade is established.
[0013] The specific design steps include:
[0014] A1: Based on the ABAQUS finite element software and the finite element modeling method, a finite element model of a large-size composite blade is established.
[0015] A2: Define the curing kinetics equation by writing a user subroutine to achieve the calculation of the thermal-chemical model.
[0016] A3: Determine the convection heat transfer boundary conditions on the upper and side surfaces of large-sized composite blades and the heat conduction boundary conditions on the lower surface in contact with the mold by writing user subroutines.
[0017] After completing the design process of step A, proceed to the design of step B, where:
[0018] Step B: Based on the relationship between temperature and curing degree, the second heating curve is changed to a segmented type, and a regional heating method is adopted to increase the heating temperature of the tenon root area in the second heating stage.
[0019] The specific design steps include:
[0020] B1: Collect the temperature and curing degree data of the bottom surface in contact with the mold and the temperature and curing degree data of the center of the tenon root, and calculate the temperature difference and curing degree difference at the two points at different times.
[0021] B2: According to the distribution of the difference in curing degree, the first-order differential or second-order differential is used to design the insulation time and heating rate, and the second heating curve is changed to a segmented type to obtain the optimized curve at the tenon root.
[0022] B3: The optimized curve at the tenon root is used as the overall solidification curve for simulation to obtain the temperature difference and solidification degree difference between the blade body and the tenon root.
[0023] B4: Add a heat source at the tenon root, adjust the heat source size according to the distribution of the difference in curing degree between the blade body and the tenon root, and then optimize the uniformity of the temperature field during blade curing.
[0024] After completing the design process of step B, proceed to the design of step C, where:
[0025] Step C: Establish a series LSTM neural network, collect the curing temperature and curing degree of epoxy resin composite materials, set a sliding window, establish temperature and curing degree data sets respectively, and divide them into training set and test set.
[0026] The specific design steps include:
[0027] C1: Construct an LSTM neural network and introduce a "gate" mechanism (forget gate, input gate, output gate) to solve the gradient vanishing or gradient exploding problems that are prone to occur in the traditional RNN training process, so as to more effectively capture the long-term dependency between temperature and curing degree during the autoclave curing process.
[0028] Forget Gate t is to change the state unit c of the previous moment t-1 The relevant information is conditionally selected and saved to the current state unit c t The calculation formula is:
[0029] f t =σ(W f ×[h t-1 , x t ]+b f ) (1)
[0030] Where: W f and bf-the weight matrix and bias matrix of the forget gate; the σ(·) gate is the sigmoid activation function; the output of the previous unit is h t-1 The matrix composed of the current moment input.
[0031] Input Gate i i Update the information in the input unit state and store it in the cell state. The calculation formula is:
[0032] i t =σ(W t ·[ht-1 , x t ]+b1) (2)
[0033] m t =tanh(w c ·[h t-1 , x t ]+b o ) (3)
[0034] c t =f t ⊙c t-1 +i t ⊙m t (4)
[0035] Where: ⊙ - Hadamard product; tanh(·) - tanh activation function; W i and b i -The weight matrix and bias term of the input gate; W c and b c -The weight matrix and bias term of the updated value; m i - the current input unit status; c t-1 -The state value at the last moment; c t -The current state value.
[0036] o t The output gate is mainly used to determine the output information of the unit state at the current moment. The formula is as follows:
[0037] o t =σ(W 。 ·[h t-1 , x t ]+b o ) (5)
[0038] h t =o t ⊙tanh(c t ) (6)
[0039] Where W o and b o -The weight and bias of the output gate; h t - Output the current time.
[0040] C2: Connect two LSTM networks in series, define two custom neural networks LSTM1 and LSTM2, each with an LSTM network. Then, define a new neural network ConcatNet, which connects these two neural networks in series. Finally, instantiate LSTM1, LSTM2, and ConcatNet, and use ConcatNet for training or prediction.
[0041] C3: Extract the temperature and curing degree of a point at the center of the tenon root, extension root and blade area as the original data, set the sliding window to achieve single-step or multi-step prediction. Divide the training set and test set in a ratio of 7:3.
[0042] After completing the design process of step C, proceed to the design of step D, where:
[0043] After completing the design process of step C, proceed to the design of step D, where:
[0044] The Grey Wolf Algorithm (GWO) was used to optimize the LSTM model, optimize the initial learning rate, the number of neurons in the hidden layer, the batch size, and the number of training iterations, and obtain the best network hyperparameter values of LSTM.
[0045] The core idea is to use the position information of α, β and δ wolves to guide other wolves to move towards the optimal solution, thereby achieving a balance between global search and local search. The core idea is to use the position information of α, β and δ wolves to guide other wolves to move towards the optimal solution, thereby achieving a balance between global search and local search. The mathematical model of the gray wolf's behavior of surrounding prey is defined as follows:
[0046] D=|C·Xp(t)-X(t)| (7)
[0047] X(t+1)=Xp(t)-A·D (8)
[0048] A=2a·r1-a (9)
[0049] C=2·R2 (10)
[0050] Xp(t) - current prey position vector;
[0051] X(t)-current position vector of the gray wolf;
[0052] X(t+1)-the position vector of the gray wolf after the next iteration;
[0053] D - distance between the wolf and the prey;
[0054] a-convergence factor, which decreases linearly from 2 to 0 as the number of iterations increases;
[0055] r1, R2 - random numbers in the interval [0, 1].
[0056] When the gray wolf identifies the location of the prey, α guides β and δ to attack the prey. The first three optimal solutions are selected, and the remaining gray wolves update their positions based on α, β, and δ.
[0057]
[0058]
[0059]
[0060] Where: D α -α the distance between α and other gray wolves;
[0061] D β - distance between β and other gray wolves;
[0062] D δ -δ distance to other gray wolves;
[0063] X α (t)-αThe position vector at the current number of iterations;
[0064] X β (t)-βThe position vector at the current iteration number;
[0065] X δ (t)-δThe position vector at the current number of iterations;
[0066] X1-the vector of a gray wolf in the wolf pack moving towards α;
[0067] X2-the vector of a gray wolf in the wolf pack moving toward β;
[0068] X3-The vector of a gray wolf in the wolf pack moving towards δ.
[0069] After completing the design process of step D, proceed to the design of step E, where:
[0070] The root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R 2 ) to verify the prediction accuracy, reliability and stability of the LSTM model by the optimization algorithm.
[0071] Among them, the smaller the RMSE and MAE values are, the more accurate the model prediction results are. 2 The closer it is to 1, the greater the goodness of fit and the better the model prediction effect. 2 The specific expression is:
[0072]
[0073]
[0074]
[0075] Where: N-number of prediction samples;
[0076] yi′-temperature prediction value
[0077] yi-true value of temperature;
[0078] y - mean of the true values.
[0079] Compared with the existing autoclave curing for variable thickness composite materials, the beneficial effects of the present invention based on regional curing and predicting temperature and curing degree through a series LSTM neural network are as follows:
[0080] The present invention provides a method for coupling finite element analysis with machine learning to predict changes in temperature and degree of cure of a temperature field. The method takes into account the special structure of a large-sized composite blade during curing in an autoclave, because the thickness of the tenon root is much greater than that of the blade body, and the thickness becomes thinner on both sides during the process of excessively decreasing toward the blade body. The traditional method of optimizing the temperature curve cannot meet the requirements of temperature field uniformity. In addition, the traditional method of optimizing the temperature curve requires multiple simulations and takes a lot of time.
[0081] The abaqus finite element analysis software is used to calculate the thermal-chemical model of large-sized blades of composite materials by writing user subroutines, and finally a numerical simulation model of the temperature field curing of large-sized blades of composite materials is established. Considering that residual stress will be generated after the curing degree exceeds the gel point, the second heating curve is optimized based on the first-order differential and second-order differential of the curing degree difference, and the heating curve is changed to a segmented type. The first section is a slow heating, and the heating rate is determined by finite element analysis, so that Δα increases slowly. In order to take into account the uniformity of the curing degree and the total curing time, it ends when the first-order differential or second-order differential of the curing degree difference is equal to zero, which can reduce Δα and no longer increase sharply in the subsequent curing process. Considering that the thickness of the tenon stem area is large, the curing speed is slow in the early stage of the second heating, and the internal heat accumulation is large, a temperature-related heat source is added to the first stage of the second heating of this area to reduce the temperature gradient and the curing degree gradient. After reaching the specified temperature of the process, it is kept warm until the curing is completed.
[0082] After the abaqus finite element analysis is completed, the proxy model is established through the LSTM series neural network, and the Bayesian optimization algorithm and the gray wolf algorithm are used for optimization. 2 As a reference indicator, the optimal series LSTM neural network is found to ensure the accuracy and generalization ability of the proxy model.
[0083] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present invention in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0084] Figure 1 It is a schematic diagram of the finite element mesh of a large-size blade of composite materials;
[0085] Figure 2 It is the temperature curve of three points A, B and C before optimization;
[0086] Figure 3 It is a curve chart obtained by the difference in temperature and the difference in curing degree at point C and the bottom surface in contact with the mold;
[0087] Figure 4 It is a segmented temperature rise curve of the temperature field of a large-size blade of composite materials;
[0088] Figure 5 Comparison of the curing degree difference and temperature difference obtained by simulating the curing process of a large-sized composite blade by optimizing the heating temperature curve before and after;
[0089] Figure 6 It is the overall optimization curve and the tenon root area optimization curve diagram;
[0090] Figure 7 It is a diagram of the structure of a serial LSTM neural network;
[0091] Figure 8 This is a flow chart of the temperature field simulation prediction model for large-size composite blades based on GWO-LSTM. DETAILED DESCRIPTION
[0092] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific implementation disclosed below.
[0093] Explanation of the embodiment:
[0094] First, perform step A, wherein:
[0095] A: Based on the finite element software abaaqus and the finite element modeling method, a finite element model of a large-sized composite blade is established. The heat transfer model of the large-sized composite blade is coupled with the curing kinetics model through temperature, and the calculation of the thermal-chemical model is realized by writing a user subroutine, and finally a numerical simulation model of the temperature field curing of a large-sized composite blade is established.
[0096] The specific design steps include:
[0097] A1: Based on the ABAQUS finite element software and the finite element modeling method, a finite element model of a large-size composite blade is established.
[0098] like Figure 1 As shown in the figure, the large-size composite blade consists of a tenon root, an extension root and a blade body. The tenon root is a dovetail tenon root. The dovetail tenon root is narrow and the end is wide, forming a big head shape. The thickest part is more than twice as large as the blade body. AS4 / 3501-6 is used as the material for thermosetting sequential coupling, and the unit type is DC3D8.
[0099] A2: Define the curing kinetics equation by writing a user subroutine to realize the calculation of the thermal-chemical model.
[0100] Define field variables such as initial curing degree and curing rate in the USDFLD (Redefine field variables at a material point) subroutine; define the heat generated by the resin curing reaction inside the material in the HETVAL (Provide internal heat generation in heat transfer analysis) subroutine, and calculate the curing degree and curing rate based on the resin curing kinetic equation. Define the USDFLD subroutine and HETVAL subroutine in the abaqus material property location
[0101] The autocatalytic model of AS4 / 3501-6 is:
[0102]
[0103]
[0104] Where: K i , A i , ΔE i (i=1, 2, 3) are the rate parameter, frequency factor and activation energy of the autocatalytic reaction respectively.
[0105] A3: Determine the convective heat transfer boundary conditions and heat conduction boundary conditions by writing user subroutines.
[0106] According to the structural characteristics of the composite large-size blade finite element model constructed by A1, the boundary conditions of the composite large-size blade are set: 1) The upper surface and side surface of the blade conduct convective heat exchange with the medium in the tank, which is equivalent to the convective heat transfer coefficient between the blade and the air; 2) The lower surface temperature of the blade is equivalent to the upper surface temperature of the molding mold plate. When the finite element model is established, the overall initial temperature is set to room temperature (about 25℃) in the load module. The blade curing process curve and ambient temperature are defined using the FILM (Define surface-based nonuniform film coefficient) subroutine and the DISP (Specify prescribed boundary conditions) subroutine.
[0107] After completing the design process of step A, proceed to the design of step B, where:
[0108] B: According to the relationship between temperature and curing degree, the second heating curve is changed to a segmented one. The heating method of the area is adopted to increase the temperature of the tenon root area in the second heating stage.
[0109] The specific design steps include:
[0110] B1: Collect the temperature and curing degree data of the bottom surface in contact with the mold and the temperature and curing degree data of the center of the tenon root, and calculate the temperature difference and curing degree difference at the two points at different times.
[0111] Select the data at points A, B, and C. Point A represents the temperature and curing degree changes at the center of the blade body, point B represents the temperature and curing degree changes at the center of the root, and point C represents the temperature and curing degree changes at the center of the tenon root. Figure 2 As shown in Figure 2, before optimization, the temperature change at point C is the largest. Figure 3 As shown, based on the post-processing of abaqus simulation results by python, the difference between the curing degree at point C and the curing degree of the bottom surface in contact with the mold is collected, and the first-order differential of the maximum curing degree is calculated.
[0112] B2: According to the distribution of the curing degree difference, the first-order differential or second-order differential is used to design the holding time and heating rate, and the second heating curve is changed to a segmented type to obtain the optimized curve at this point. The optimized curve at this point is used as the overall curing curve for simulation to obtain the temperature difference and curing degree difference of the large-sized blade of the composite material.
[0113] like Figure 4 As shown in the figure, the second heating curve adopts a segmented method, with the first-order differential of the curing rate As the segmentation point, the heating rate of the first section is reduced, the curing speed of the tenon root area is reduced, excessive internal heat accumulation is avoided, and the heating rate of the second section is increased. The second insulation platform can be reached faster and the overall curing cycle can be reduced.
[0114] like Figure 5 The figure shows the curing degree difference and temperature difference comparison of the curing process of large-size composite blades simulated by the heating temperature curves before and after optimization. Taking into account the material characteristics of AS43501-6 and the temperature change when α is greater than 0.57, it is concluded that the overall temperature field uniformity is improved by adopting the segmented heating method, but there is still a large temperature gradient and curing degree gradient at the root of the tenon.
[0115] B3: Add a heat source at the root of the tenon, adjust the size of the heat source according to the distribution of the difference in the degree of solidification between the blade body and the root of the tenon, and then optimize the uniformity of the temperature field of the blade solidification. In the abaqus load module, create a surface heat flux boundary condition and define the DFLUX (Define nonuniform distributed flux in heat transfer analysis) subroutine.
[0116] like Figure 6 The overall optimization curve and the root area optimization curve are shown. The DFLUX subroutine is used to define the internal heat source added to the bottom of the root area of the composite large-size blade. In the abaqus finite element software, the added internal heat source is transferred in the form of surface heat flux. Therefore, by adjusting the size of the surface heat flux, the temperature gradient at the root of the tenon can be reduced. The temperature gradient optimization comparison diagram at the root of the tenon is shown in Figure 7 shown.
[0117] After completing the design process of step B, proceed to the design of step C, where:
[0118] Step C: Establish a series LSTM neural network, collect the curing temperature and curing degree of epoxy resin composite materials, set a sliding window, establish temperature and curing degree data sets respectively, and divide them into training set and test set.
[0119] The specific design steps include:
[0120] C1: Construct an LSTM neural network and introduce a "gate" mechanism (forget gate, input gate, output gate) to solve the gradient vanishing or gradient exploding problems that are prone to occur in the traditional RNN training process, so as to more effectively capture the long-term dependency between temperature and curing degree during the autoclave curing process.
[0121] Based on the MATLAB neural network toolbox, open the deep network designer. Since the LSTM neural network is used for prediction, the sequence to sequence module is selected, the input parameter is 4, the output parameter is 1, and the LSTM layer is temporarily set to 3.
[0122] C2: Connect two LSTM networks in series and define two custom neural networks LSTM1 and LSTM2, each with an LSTM network. Then, define a new neural network ConcatNet, which connects the two neural networks in series. Finally, instantiate LSTM1, LSTM2, and ConcatNet, and use ConcatNet for training or prediction. The structure of the tandem LSTM network is as follows: Figure 7 shown.
[0123] C3: Extract the temperature and curing degree of a point at the center of the tenon root, extension root and blade area as the original data, set the sliding window to achieve single-step or multi-step prediction. Divide the training set and test set in a ratio of 7:3.
[0124] It is assumed that the composite material is completely uncured and its temperature is equal to 25°C at the beginning of the process. The heating rate and holding temperature of the second stage are defined as 0.5-3.5°C / min and 170-220°C, respectively. The curing simulation of the large-scale blade of the composite material was simulated using abaqus, and three different positions were selected ( Figure 1 A, B and C in the figure), observe the three control points at each position (located at the root of the tenon, the root of the extension and the middle of the blade), and get 24 sets of data. Due to the differences in units and dimensions of various data, this paper uses normalization to process the data, and divides the data groups into 7:3 after normalization. The data normalization formula is:
[0125]
[0126] Where: y′-normalized conversion value;
[0127] y-true value;
[0128] y max - maximum value of samples;
[0129] y min - Sample minimum value.
[0130] The autoclave temperature and the temperature and time of adding the heat source are selected as inputs, and the temperatures at points A, B, and C are obtained as outputs after passing through the first LSTM1 network. Then, the temperatures at points A, B, and C are imported into the second LSTM2 network, and the curing degrees at points A, B, and C are obtained as outputs.
[0131] After completing the design process of step C, proceed to the design of step D, where:
[0132] GWO was used to optimize the LSTM model, optimize the initial learning rate, the number of neurons in the hidden layer, the batch size, and the number of training iterations, and obtain the optimal network hyperparameter values of LSTM.
[0133] The specific design steps include:
[0134] The optimization goal of the LSTM tandem neural network is achieved by simulating the predation behavior of gray wolf groups and the predation mechanism based on wolf group collaboration. The core idea is to use the location information of α gray wolf, β gray wolf and δ gray wolf to guide other gray wolf individuals to move towards the optimal solution, thereby achieving a balance between global search and local search.
[0135] This paper uses the Gray Wolf Optimization Algorithm to obtain the optimal network hyperparameter values of LSTM, including the initial learning rate, the number of neurons in the hidden layer, the batch size, and the number of training iterations, namely [lr, L1, Batch, k]. Increasing the number of hidden layers can improve the nonlinear fitting ability of the model, but it also makes the model more complex, the prediction time increases, and even causes overfitting problems. Therefore, this paper controls the range of the number of hidden layers to 1 layer. The constraints of the optimization parameters are set as follows:
[0136] Table 1 Constraints of optimization parameters
[0137]
[0138] The simulation prediction model process of temperature field of large-size composite blades based on GWO-LSTM proposed in this paper is as follows: Figure 8 According to the flowchart, the specific steps of constructing the temperature field simulation prediction model of large-size composite blades based on GWO-LSTM are as follows.
[0139] Step 1: After outliers are removed, missing values are supplemented, and data is normalized, the original time series sample data is divided into a training set and a test set in a ratio of 7:3.
[0140] Step 2: Determine the initial data of GWO, including the number of gray wolf populations, initial coordinates, and number of iterations. Convert the number of hidden layer units, learning rate, and number of iterations of the LSTM network into the location coordinates of the wolf pack, and select a training sample set to train the LSTM model.
[0141] Step 3: Calculate the individual fitness values of the wolf pack. The fitness function of the gray wolf is set to Where y is the measured value, y iis the model prediction value, T is the length of the time series, and the smaller the fitness function, the better. Update the individual positions of the wolf pack according to the fitness value. When the search reaches the maximum number of iterations or the global optimal position meets the minimum limit, the optimal solution of the three model parameters of the LSTM network, namely the number of hidden layer units, learning rate and number of iterations, is obtained.
[0142] Step 4: Select a test sample set and test the LSTM network with the above optimized parameters to obtain the optimal LSTM network model. Based on the optimal LSTM network model, predict the temperature and curing degree of large-size blades of composite materials.
[0143] After completing the design process of step D, proceed to the design of step E, where:
[0144] The root mean square error (RMSE), mean absolute error (MAE) and determination coefficient R 2 , verify the optimization algorithm’s prediction accuracy, reliability and stability on the LSTM model.
[0145] Table 2 Validation values of the model
[0146]
[0147] In order to enhance the model's ability to learn from existing data, data standardization is essential to improve prediction accuracy and address the impact of different dimensions. In this study, data standardization involves adjusting the input data distribution to achieve a mean of 0 and a variance of 1. Table 2 lists the indices of the machine learning models applied to the training and test sets. It is worth noting that negative values in MAE and RMSE indicate superior performance as smaller values indicate higher accuracy. In contrast, positive values in the coefficient of determination (R2) are observed. Among the above machine learning models, the LSTM neural network model exhibited the lowest accuracy, which indicates that the assumption of a linear relationship between input and output is inaccurate. R of GWO-LSTM 2 The value is close to 1.0, reflecting its reliable prediction, and the GWO-LSTM neural network model appears to be closer to the true value, as shown by its lower MAE and RMSE.
Claims
1. A method for predicting the temperature and curing degree of large-size blades by coupling finite element analysis with machine learning. It is characterized in that include: A: Based on the finite element software abaqus and the finite element modeling method, a finite element model of a large-sized composite blade is established. The heat transfer model and the curing kinetics model are coupled through temperature, and the calculation of the thermal-chemical model is realized by writing a user subroutine, and finally a numerical simulation model of the temperature field curing of a large-sized composite blade is established; B: According to the relationship between temperature and curing degree, the second heating curve is changed to a segmented type, and the heating method of the area is adopted to increase the heating temperature of the tenon root area in the second heating stage; C: Establish a series LSTM (Long Short-Term Memory) neural network, collect the curing temperature and curing degree of epoxy resin composite materials, set a sliding window, establish temperature and curing degree data sets respectively, and divide them into training set and test set; D: Use the Grey Wolf Algorithm (GWO) to optimize the LSTM model, optimize the initial learning rate, the number of hidden layer neurons, the batch size, and the number of training iterations to obtain the best network hyperparameter values of LSTM; E: Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and Coefficient of Determination (R 2 ) to verify the prediction accuracy, reliability and stability of the LSTM model by the optimization algorithm.
2. According to claim 1, a method for predicting the temperature and curing degree of a large-sized blade by coupling finite element analysis and machine learning, It is characterized in that In step A, the specific steps include: A1: Based on the ABAQUS finite element software and the finite element modeling method, a finite element model of a large-size composite blade is established. The large-sized composite blade consists of a tenon, an extension root and a blade body. The tenon is a dovetail-type tenon with a narrow root and a wide end, forming a big head shape. The thickest part is more than twice as large as the blade body. A2: Define the curing kinetics equation by writing a user subroutine to realize the calculation of the thermal-chemical model. An abaqus user subroutine was written based on the AS4 3501-6 material properties of the large-size composite blade finite element model constructed by A1. Field variables such as initial curing degree and curing rate were defined in the USDFLD (Redefine field variables at a material point) subroutine. The heat generated by the resin curing reaction inside the material was defined in the HETVAL (Provide internal heat generation in heat transfer analysis) subroutine, and the curing degree and curing rate were calculated according to the resin curing kinetic equation. A3: Determine the convection heat transfer boundary conditions of the upper and side surfaces of the large-sized composite blade and the heat conduction boundary conditions of the lower surface in contact with the mold by writing a user subroutine. According to the structural characteristics of the composite large-size blade finite element model constructed by A1, the boundary conditions of the composite large-size blade are set: 1) The upper surface and side surface of the blade conduct convective heat exchange with the medium in the tank, which is equivalent to the convective heat transfer coefficient between the blade and the air; 2) The lower surface temperature of the blade is equivalent to the upper surface temperature of the molding mold plate. When the finite element model is established, the overall initial temperature is set to room temperature (about 25℃) in the load module. The blade curing process curve and ambient temperature are defined using the FILM (Define surface-based nonuniform film coefficient) subroutine and the DISP (Specify prescribed boundary conditions) subroutine.
3. According to the method of coupling finite element analysis and machine learning to predict the temperature and curing degree of large-sized blades according to claim 1, It is characterized in that In step B, the specific steps include: B1: Collect the temperature and curing degree data of the bottom surface of the large-sized composite blade and the center of different thickness of the large-sized composite blade, and calculate the temperature difference and curing degree difference between the center of different thickness of the large-sized composite blade and the bottom surface of the large-sized composite blade. Collect data from three representative points at the center of different thicknesses of large-sized composite blades, and select the point with the largest temperature gradient as the optimization target. Based on Python, post-process the abaqus simulation results, collect the difference between the curing degree at this point and the curing degree of the bottom surface in contact with the mold, and calculate the first-order differential of the maximum curing degree. B2: According to the distribution of the curing degree difference, the first-order differential or second-order differential is used to design the holding time and heating rate, and the second heating curve is changed to a segmented type to obtain the optimized curve at this point. The optimized curve at this point is used as the overall curing curve for simulation to obtain the temperature difference and curing degree difference of the large-sized blade of the composite material. Considering the material characteristics of AS4 3501-6, the temperature curve when α is greater than 0.57 is optimized, that is, the second heating curve and the temperature of the insulation platform are optimized. The second heating curve adopts a segmented method, which reduces the heating rate of the first segment and increases the heating rate of the second segment. The segmented methods with the first-order differential equal to zero and the second-order differential equal to zero are simulated respectively. It is found that the segmented method with the first-order differential equal to zero produces a smaller temperature gradient and curing degree gradient. Therefore, the first-order differential of the maximum curing degree obtained at this point is selected. For segmentation point. B3: Add a heat source at the root of the tenon, adjust the size of the heat source according to the distribution of the difference in the degree of solidification between the blade body and the root of the tenon, and then optimize the uniformity of the temperature field of the blade solidification. In the abaqus load module, create a surface heat flux boundary condition and define the DFLUX (Define nonuniform distributed flux in heat transfer analysis) subroutine. The DFLUX user subroutine is defined using fortron language to find the internal heat source at the bottom of the tenon root area of a large-sized composite blade. The temperature gradient at the tenon root is reduced to within 10°C by adjusting the size of the internal heat source.
4. According to claim 1, a method for dynamically predicting temperature field temperature and curing degree changes by coupling finite element analysis and machine learning, It is characterized in that In step C, the specific steps include: C1: Construct two LSTM neural networks in series Based on MATLAB software, two custom neural networks LSTM1 and LSTM2 are defined, each with an LSTM network. Then, a new neural network ConcatNet is defined, which connects these two neural networks in series. Finally, LSTM1, LSTM2, and ConcatNet are instantiated, and ConcatNet is used for training or prediction. C2: Extract the temperature and curing degree of a point at the center of the root, extension root and blade area of the large-size blade of the tenon composite material as the original data, set the sliding window to achieve single-step or multi-step prediction. Divide the training set and test set in a ratio of 7:
3. 24 sets of data were collected and normalized, and then divided into data groups according to 7:3 after normalization.
5. The method for predicting the temperature and curing degree of large-sized blades by coupling finite element analysis and machine learning according to claim 1, It is characterized in that In step D, the specific steps include: The Grey Wolf Optimizer (GWO) algorithm is used to optimize the initial learning rate, number of hidden layer neurons, batch size and number of training iterations of the LSTM neural network. Step 1: After outliers are removed, missing values are supplemented, and data is normalized, the original time series sample data is divided into a training set and a test set in a ratio of 7:
3. Step 2: Determine the initial data of GWO, including the number of gray wolf populations, initial coordinates, and number of iterations. Convert the number of hidden layer units, learning rate, and number of iterations of the LSTM network into the location coordinates of the wolf pack, and select a training sample set to train the LSTM model. Step 3: Calculate the individual fitness values of the wolf pack. The fitness function of the gray wolf is set to Where y is the measured value, y i is the model prediction value, T is the length of the time series, and the smaller the fitness function, the better. Update the individual positions of the wolf pack according to the fitness value. When the search reaches the maximum number of iterations or the global optimal position meets the minimum limit, the optimal solution of the three model parameters of the LSTM network, namely the number of hidden layer units, learning rate and number of iterations, is obtained. Step 4: Select a test sample set and test the LSTM network with the above optimized parameters to obtain the optimal LSTM network model. Based on the optimal LSTM network model, predict the temperature and curing degree of large-size blades of composite materials.
6. According to claim 1, a method for dynamically predicting temperature field temperature and curing degree changes by coupling finite element analysis and machine learning, It is characterized in that In step E, the specific steps include: The root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination R 2 , verify the optimization algorithm’s prediction accuracy, reliability and stability on the LSTM model.
Citation Information
Cited By
Parameter optimization method and device for composite material forming process
CN119763744A
Parameter optimization method and device for composite material molding process
CN119763744B
Determination method of engine working blade directional solidification finite element simulation boundary condition
CN120354559A
Thermal-chemical simulation method for composite material curing based on physical information neural network
CN120409247A
A method for thermo-chemical simulation of composite curing based on physical information neural network
CN120409247B
