Adaptive Prediction Method for Curing Deformation of Composite Components Based on Point-Level Sensitivity
By adopting an adaptive prediction method based on point-level sensitivity, the problems of accuracy and efficiency in the prediction of curing deformation of composite components are solved. It achieves high-precision and rapid prediction of carbon/quartz hybrid layup configurations, adapts to layup configurations with different material ratios, and improves the robustness and prediction efficiency of the model.
Patent Information
- Application Number
- CN202611383933.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-08
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to simultaneously balance the accuracy, configuration adaptability, and computational efficiency of curing deformation prediction for composite components. This is especially true for carbon/quartz hybrid layup configurations, where existing methods suffer from issues such as strong subjectivity in manual zoning, poor adaptability to hybrid configuration feature schemes, insufficient prediction accuracy for high-risk areas, and low simulation computation efficiency.
An adaptive prediction method based on point-level sensitivity is adopted. Through the construction and preprocessing of multi-source datasets, the point-level sensitivity index is calculated. Combining the differences in mechanical properties between carbon fiber and quartz fiber, an adaptive feature selection mechanism is constructed, and continuous sample weights are introduced to achieve point-by-point refined training, replacing the traditional manual partitioning method and improving the robustness and generalization ability of the model.
It significantly improves the prediction accuracy of stress concentration areas such as edges and corners, enhances the fitting ability of high deformation risk areas, improves prediction efficiency, supports rapid iterative optimization of multiple schemes for process parameters, and reduces trial production costs and R&D cycle.
Smart Images

Figure CN123120150A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of precision manufacturing and intelligent quality prediction technology of composite materials, and particularly relates to a method for predicting the curing deformation of fiber-reinforced composite material components, which is especially suitable for adaptive prediction of curing deformation of carbon / quartz hybrid layup composite material components. Background Technology
[0002] Fiber-reinforced composite materials, with their superior properties such as high specific strength, high specific modulus, strong designability, corrosion resistance, and fatigue resistance, have become core materials for achieving structural lightweighting and performance upgrades in aerospace, high-end equipment, new energy, and rail transportation. Curing is the core process in the fabrication of fiber-reinforced composite components. During the molding process, the resin undergoes a cross-linking and curing reaction, accompanied by the coupling effects of multiple physical fields such as temperature, stress, and chemical shrinkage. After demolding, the component will experience significant curing deformation. This deformation directly affects the component's dimensional accuracy, assembly pass rate, and structural mechanical properties, and in severe cases, can lead to component failure. Therefore, accurately predicting the curing deformation distribution of components under different layup configurations and process parameters is a crucial prerequisite for optimizing layup design, controlling process parameters, reducing trial production costs, and shortening product development cycles.
[0003] Currently, composite material curing deformation prediction technology is mainly divided into two categories: mechanism-based finite element simulation method and data-driven prediction method.
[0004] Finite element method (FEM) simulation, based on curing kinetics and thermoviscoelastic constitutive theory, constructs a multi-physics coupled numerical model of the curing process. It can reflect the evolution of curing deformation at the mechanistic level and is currently a widely used prediction method in engineering. However, this method has significant shortcomings in practical applications: Firstly, simulation calculations rely on fine mesh generation and accurate material constitutive parameters. Simulations of a single-configuration complete curing process often take hours to days, resulting in extremely low computational efficiency and making it difficult to support rapid iterative optimization of multiple process parameters. Secondly, simulation results are highly dependent on the engineering experience of the modelers. Differences in mesh generation methods, boundary condition settings, and material parameter selection can lead to significant deviations in simulation results, resulting in poor consistency and repeatability, which is insufficient to meet the needs of rapid research and development.
[0005] With the development of artificial intelligence and big data technologies, data-driven methods for predicting solidification deformation have gradually become a research hotspot. These methods train machine learning models based on historical measured or simulated data, directly establishing a nonlinear mapping relationship between input parameters and deformation. The prediction speed is far higher than traditional finite element simulation, providing a new solution for rapid process iteration. However, existing data-driven methods still have three major shortcomings in engineering applications: First, most existing methods employ manual partitioning, dividing the component surface into several regions such as the central area, edge area, and corner area, and constructing a prediction model based on region-level features. The partitioning rules rely entirely on manual experience, are highly subjective, cannot adapt to components of different sizes and ply configurations, and have poor generalization adaptability; moreover, the region-level granularity is relatively coarse, making it impossible to achieve fine-grained prediction point by point, and the prediction error is generally large in areas with stress concentration and high deformation risk, such as edges and corners, which is difficult to meet the requirements of high-precision engineering applications.
[0006] Secondly, for hybrid layup configurations of carbon fiber and quartz fiber, existing methods generally employ a uniform feature input scheme, failing to consider the fundamental differences in the mechanical properties of the two types of fiber materials. Carbon fiber composites exhibit strong anisotropy, with curing deformation primarily dominated by layup parameters. Spatial coordinate features have low marginal contributions, and their introduction can actually create redundant information that interferes with the model's learning of the dominant laws. Quartz fiber composites, on the other hand, are approximately isotropic, with curing deformation significantly influenced by spatial boundary constraints. The lack of coordinate features makes it impossible to accurately capture the spatial distribution of deformation. The uniform feature scheme cannot simultaneously account for the deformation characteristics of both materials, resulting in insufficient robustness and generalization ability of the model under hybrid layup configurations.
[0007] Third, most existing prediction models are trained using equal sample weights and do not specifically enhance the fitting of sensitive points with high deformation risk. The models are not good at fitting the extreme deformation regions and cannot meet the accuracy control requirements of high-risk areas in engineering applications.
[0008] In summary, existing technologies struggle to simultaneously achieve prediction accuracy, configuration adaptability, and computational efficiency. Developing a point-level refined method for predicting the curing deformation of composite materials that adapts to hybrid material configurations and balances high accuracy and high efficiency is a pressing technical problem that needs to be solved in this field. Summary of the Invention
[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide an adaptive prediction method for curing deformation of composite material components based on point-level sensitivity, which solves the problems of strong subjectivity of manual partitioning, poor adaptability of mixed configuration feature schemes, insufficient prediction accuracy of high-risk areas, and low simulation calculation efficiency in the prior art.
[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Step S1: Construction and Preprocessing of Multi-Source Deformation Dataset Curing deformation data of composite components with different layup configurations were collected to construct a fusion dataset of "measured data + simulation data". Standardization preprocessing was performed on all data to obtain a standard sample set containing the deformation values of all mesh coordinate points on the component surface. Specifically, the following steps were included: Step S11: Multi-source data acquisition Simulation data is used to establish a curing process model through multiphysics coupling simulation software. The input layup parameters (material type, layup angle, and layup sequence for each layer) and process parameters (heating rate, cooling rate, holding time, and curing pressure) are used to output the curing deformation of each coordinate point of a 41×41 grid on the surface of the component in batches. The measured data is collected by a three-dimensional coordinate measuring instrument to collect the actual coordinates of the grid points on the surface of the component after curing. The deformation is extracted by comparing with the theoretical model and the coordinate points are automatically matched by image recognition technology.
[0011] The dataset covers 16 typical layup configurations, including different carbon fiber / quartz ratios, angle combinations, and layup sequences. The process parameters range from 1 to 5℃ / min, 1 to 5℃ / min, 60 to 100 min, and 500 to 900 kPa. The dataset uses individual grid coordinate points as independent training samples, and after cleaning, it contains approximately 27,000 valid coordinate point samples.
[0012] Step S12: Data Alignment and Cleaning First, coordinate alignment is performed: the grid coordinates of all samples are uniformly mapped to the standard coordinate system of [0,400]×[0,400] to eliminate coordinate offset caused by differences in component size and clamping position; Next, outlier removal is performed: the interquartile range (IQR) method is used to trim the deformation of all grid points within a single configuration. The upper quartile Q3 and lower quartile Q1 of the deformation are calculated, and the interquartile range IQR = Q3 - Q1. Extreme outliers that exceed the range of [Q1 - 1.5 × IQR, Q3 + 1.5 × IQR] are removed to avoid outliers interfering with model training. Finally, missing value completion is performed: bilinear interpolation is used to complete the deformation for a small number of missing coordinate points, ensuring that each configuration contains complete 41×41 grid data.
[0013] Step S13: Numerical Standardization and Feature System Construction All process and layup type input features are processed using min-max normalization to map the values to the [0,1] interval, eliminating the influence of different dimensions on model training; the deformation label retains the original physical quantity value for subsequent accuracy evaluation and engineering applications.
[0014] The input features of this invention are divided into three main categories: global basic features, sensitivity features, and spatial boundary derived features, wherein: Global basic characteristics: total number of layup layers, carbon fiber percentage, 0° layup percentage, 90° layup percentage, ±45° layup percentage, layup symmetry factor; process characteristics include heating rate, cooling rate, holding time, and curing pressure; Sensitivity characteristics: Point-level sensitivity index corresponding to the coordinate point; Spatial boundary derived features: i.e., spatial features, normalization x Coordinates, normalized y-coordinate, normalized distance from the coordinate point to the four boundaries of the component, normalized distance from the nearest corner point, and binary identifier of the boundary point (marked as 1 when the normalized distance from the coordinate point to any boundary of the component is less than 0.1, otherwise marked as 0).
[0015] The carbon fiber-dominated configuration uses only global basic features and sensitivity features as input, while the quartz-dominated / hybrid configuration uses all three types of features as input.
[0016] Step S2: Data-driven point-level sensitivity index calculation Based on all training samples, a globally unified point-level sensitivity index is calculated and used for training weight allocation of all layer group models.
[0017] This invention does not use the traditional manual partitioning method, but directly calculates the deformation sensitivity of each coordinate point independently, fundamentally eliminating the subjectivity and lack of adaptability of manual partitioning.
[0018] Calculate the mean, standard deviation, and maximum value of the deformation sequence (i.e., the deformation sequence after outlier removal, referring to the sequence of deformation values of the same coordinate point in all training samples) for each coordinate point after outlier removal in step S12. After normalizing the three indicators to their global maximum values, sum them according to preset weights to obtain the point-level sensitivity index. The weight coefficients for the three indicators are ω1, ω2, and ω3, respectively. Specifically, the formula for calculating the point-level sensitivity index S(x,y) is: in: This is the mean absolute value of the deformation at this coordinate point after IQR anomaly removal, reflecting the average severity of the deformation at this point; The standard deviation of the corresponding deformation sequence reflects the fluctuation stability of the deformation at that point under different process conditions; The maximum value of the corresponding deformation sequence reflects the extreme deformation risk at that point; the denominator term D max , , These are the global maximum values of the statistics corresponding to all coordinate points on the surface of the component, used for normalization to unify the three indicators to the [0,1] dimension. , , The weighting coefficients of the three indicators satisfy the following conditions: + + =1, and , , ≥0.
[0019] Weight , , Determining the value is the core issue at the feature construction level -- if equal weight allocation (i.e. = = If the value is 1 / 3, then the implicit assumption that the mean, standard deviation, and maximum value contribute equally to deformation risk lacks theoretical basis. In reality, the dominance of these three attributes differs significantly under different ply configurations: in carbon fiber-dominated configurations, ply parameters determine the main deformation law, and extreme value risk is often concentrated under specific ply angle combinations; in quartz-dominated configurations, boundary constraints dominate the deformation distribution, and fluctuation stability (standard deviation) may be more critical. Therefore, this invention employs a Bayesian adaptive optimization method at the feature level to automatically determine the optimal weight ratio in a data-driven manner. in To ensure goodness of fit for configurational cross-validation, optimization employs the T (Tree-structured Parzen Estimator) sampler from the Optuna framework, with the search space being: , [0,1], =1- - This optimization process solves the problem of "what indicators to use and how to weight them" at the feature construction level, making the definition of the sensitivity index itself data-adaptive.
[0020] Step S3: Adaptive Feature Selection Based on Carbon Fiber Ratio This invention addresses the differences in mechanical properties between carbon fiber and quartz fiber by constructing an adaptive feature selection mechanism driven by the proportion of carbon fiber. Carbon fiber reinforced composites exhibit extremely strong anisotropy, with significant differences in mechanical properties along the fiber direction (longitudinal) and perpendicular to the fiber direction (transverse): the longitudinal tensile strength reaches 876±36 MPa, while the transverse tensile strength is only 17.9±3.3 MPa, a difference of nearly 50 times; the longitudinal coefficient of thermal expansion is approximately -0.1×10⁻⁶. -6 / ℃, and the horizontal range is as high as +35×10. -6 / ℃. Therefore, the curing deformation of carbon fiber-dominated components is mainly determined by the laying parameters (number of layers, angle, sequence), and the marginal contribution of spatial coordinate features is low. Introducing them may introduce redundant information and interfere with model learning.
[0021] Quartz fiber composites exhibit near-isotropic characteristics, with an anisotropy ratio not exceeding 1.4:1. They also possess relatively low strength (tensile strength 1–3 GPa) and typically employ a balanced braided structure. Therefore, the deformation behavior of quartz-dominated components is closer to that of isotropic thin plates, primarily controlled by spatial boundary conditions—the deformation at free edges and corners is typically 2–3 times that of the central region due to weaker constraints. Normalized coordinate features can provide effective spatial prior information, significantly improving the model's ability to capture the spatial distribution of deformation.
[0022] Based on the above mechanism, the carbon fiber ratio R c The ratio of the total number of carbon fiber layup layers to the total number of layup layers in the component is calculated using the following formula: A carbon fiber percentage threshold of 0.45 was set. This threshold was determined through comparative experiments with multiple sets of different cutoff values, and is considered the optimal cutoff value that balances the prediction accuracy of both configurations. If R c If R > 0.45, it is determined to be the dominant configuration of carbon fiber. A modeling scheme without coordinate features is adopted to avoid redundant spatial information interfering with the learning of the dominant law of laying parameters; if R c If the value is ≤0.45, it is determined to be a quartz-dominated or hybrid configuration. Normalized coordinates and spatial boundary derived features are introduced to supplement the prior information of spatial boundary constraints. The specific composition of the spatial boundary derived features is shown in the spatial feature list in step S13.
[0023] Step S4: Training the prediction model with continuous sample weights Based on a point-level sensitivity index, continuous loss weights are assigned to each coordinate point sample, enabling the model to automatically increase the fitting weights for highly sensitive points during training. The formula for calculating the sample loss weight w corresponding to the coordinate point (x, y) is as follows: in, This is the weight scaling factor, which controls the degree to which the sensitivity adjusts the sample weights. The value of directly affects the gradient update strength at highly sensitive points during model training: On the smaller scale, insufficient weight differentiation makes it difficult to effectively strengthen high-risk areas; If the value is too large, it may distort the original data distribution, leading to overfitting. Optimal This is closely related to the sensitivity distribution characteristics of a specific dataset. Therefore, this invention employs a Bayesian adaptive optimization method at the weight level, where {w1} is determined at the feature level. w2 w3 Based on this, the optimal search is performed using the same objective function. : The search interval [0.1, 2.0] covers the complete control range from "almost no reinforcement of high-sensitivity points" (α→0) to "the weight of high-sensitivity points is 3 times that of low-sensitivity points" (when α=2.0, the weight of high-sensitivity points can reach up to 1.0+2.0=3.0), with 20 trials.
[0024] Furthermore, this invention employs a Bayesian adaptive optimization method at the model level. Based on the aforementioned optimization results, the hyperparameters of the prediction model are optimized with leave-one-out cross-validation goodness of fit as the objective, thereby improving the generalization performance of the model.
[0025] Step S5: Rapid prediction of multi-point deformation of the configuration to be predicted Input the process parameters of the layup configuration to be predicted, calculate the carbon fiber ratio and match the corresponding feature scheme with the training model, and output the deformation prediction values of all grid coordinates on the surface of the component point by point, including the following steps: Step S51: Input parameter preprocessing Input the layup parameters (material type, layup angle, layup sequence) and process parameters (heating rate, cooling rate, holding time, curing pressure) of the configuration to be predicted, and the system will automatically calculate the total number of layers, carbon fiber ratio, angle statistics and other derived parameters.
[0026] Step S52: Automatic feature and model matching Based on the calculated carbon fiber percentage R c The adaptive feature selection rules are executed to generate the input feature set of the corresponding dimension; at the same time, the corresponding type of the trained prediction model of the corresponding layer group is matched according to the total number of layers and the carbon fiber proportion type.
[0027] Step S53: Point-by-point deformation prediction output The generated feature set is input into the matching prediction model, and the deformation prediction values of all coordinate points of the grid are calculated in batches, outputting the complete surface deformation distribution results; it can quickly complete the prediction of a single configuration.
[0028] Compared with the prior art, the present invention has the following beneficial effects: By employing a refined training mechanism with point-level sensitivity weights, the deformation differences at different locations of the component can be accurately captured, significantly improving the prediction accuracy of stress concentration areas such as edges and corners, and addressing the problem of large prediction errors in stress concentration areas.
[0029] We propose a data-driven point-level sensitivity index that completely replaces the traditional manual partitioning method. By combining continuous sample weights, we achieve refined training and control from the region level to the point level. The data automatically identifies high deformation risk points without the need for manual partitioning rules, thus completely eliminating the subjectivity and poor configuration adaptability of manual partitioning and enhancing the model's ability to fit high deformation risk areas.
[0030] Based on the differences in mechanical properties between carbon fiber and quartz, an adaptive feature selection mechanism driven by the proportion of carbon fiber is constructed. Redundant spatial features are eliminated for carbon fiber-dominated configurations, while spatial prior information is supplemented for quartz-dominated configurations. This adapts to layup configurations with different material ratios, significantly improving the robustness and generalization ability of the model under hybrid layup configurations.
[0031] The prediction efficiency is greatly improved, achieving an order-of-magnitude improvement compared to traditional finite element simulation. It can support rapid iterative optimization of multiple schemes for process parameters, effectively reducing trial production costs and R&D cycle. Attached Figure Description
[0032] Figure 1 This is a schematic diagram of the overall process of the adaptive prediction method for curing deformation of composite material components based on point-level sensitivity according to the present invention. Figure 2 This is a schematic diagram of the point-level spatial risk prior and dual utilization mechanism and the three-step Bayesian optimization of the present invention; Figure 3 This is a schematic diagram of the adaptive feature configuration strategy driven by the carbon fiber ratio of the present invention. Detailed Implementation
[0033] The following is in conjunction with the appendix Figure 1 To be continued Figure 3 The present invention will be further illustrated by specific embodiments. It should be noted that these embodiments are implemented based on the technical solution of the present invention, and detailed implementation methods and specific operation processes are given, but the protection scope of the present invention is not limited to the following embodiments.
[0034] This embodiment applies to the scenario of predicting multi-point deformation during curing of carbon fiber / quartz hybrid layup composite components. The component to be predicted has a square plane shape, discretized using a 41×41 uniform grid, containing a total of 1681 coordinate points. The overall process of the adaptive prediction method for curing deformation of composite components based on point-level sensitivity described in this invention is as follows: Figure 1 As shown, the specific implementation process is as follows: Step S1: Construction and Preprocessing of Multi-Source Deformation Dataset Curing deformation data of composite components with different layup configurations were collected to construct a "real measurement + simulation" fusion dataset. Standardization preprocessing was performed on all data to obtain a standard sample set containing the deformation of all grid coordinate points on the component surface, with each individual grid coordinate point serving as an independent training sample.
[0035] S11 Multi-Source Data Acquisition Simulation data is used to establish a thermodynamic coupling model of the curing process using multiphysics coupling simulation software. The input layup parameters (material type, layup angle, and layup sequence for each layer) and process parameters (heating rate, cooling rate, holding time, and curing pressure) are used to output the curing deformation of each coordinate point of a 41×41 grid on the surface of the component in batches. The measured data is collected by a three-dimensional coordinate measuring instrument to collect the actual coordinates of the grid points on the surface of the component after curing. The deformation is extracted by comparing with the theoretical model and the coordinate points are automatically matched by image recognition technology.
[0036] The dataset in this embodiment covers 16 typical layup configurations, including pure carbon fiber, pure quartz fiber, and mixed layups with different proportions. After cleaning, it contains approximately 27,000 valid coordinate point samples.
[0037] S12 Data Alignment and Cleaning First, coordinate alignment is performed: the grid coordinates of all samples are uniformly mapped to the standard coordinate system of [0,400]×[0,400] to eliminate coordinate offsets caused by differences in component size and clamping position. Second, outlier removal is performed: the interquartile range (IQR) method is used to trim the deformation of all grid points in a single configuration, and the upper quartile Q3 and lower quartile Q1 of the deformation are calculated. The interquartile range IQR = Q3 - Q1 is used to remove extreme outliers that exceed the range of [Q1 - 1.5 × IQR, Q3 + 1.5 × IQR] to avoid outliers interfering with model training. Finally, missing value completion is performed: bilinear interpolation is used to complete the deformation of a small number of missing coordinate points to ensure that each configuration contains complete 41×41 grid data.
[0038] S13 Numerical Standardization Processing All process and layup type input features are processed using min-max normalization to map the values to the [0,1] interval, eliminating the influence of different units on model training; the deformation label retains the original physical quantity value (unit: mm) for subsequent accuracy evaluation and engineering applications.
[0039] The input features in this embodiment are divided into three categories, as follows: I. Global Basic Features (Always Existing, 46 Dimensions in Total) 1. Laying Parameter Characteristics (36 Dimensions): Number of carbon fiber layers, total number of carbon fiber layers, average number of carbon fiber layers, maximum number of carbon fiber layers, minimum number of carbon fiber layers; Number of quartz layers, total number of quartz layers, average number of quartz layers, maximum number of quartz layers, minimum number of quartz layers; Total number of laid layers, total number of all layers, carbon fiber percentage, quartz percentage; Number of 1-layer carbon fiber, number of 2-layer carbon fiber, number of 3-layer carbon fiber, number of ≥5-layer carbon fiber; Number of 1-layer quartz, number of ≥10-layer quartz; Total number of the first 3 layers, total number of the last 3 layers, number of intermediate layers; Number of carbon fiber layers in the upper half, number of carbon fiber layers in the lower half; Number of quartz layers in the upper half, number of quartz layers in the lower half; Mean of layer number variation, standard deviation of layer number variation; First layer material, number of first layer layers; Last layer material, number of last layer layers; Number of material switching times, percentage of material switching; Theoretical deformation coefficient.
[0040] 2. Angular characteristics (5 dimensions): mean angle; standard deviation of angle; mean of angle variation; standard deviation of angle variation; percentage of ply at a specific angle.
[0041] 3. Trend characteristics (5 dimensions): fluctuation frequency; deformation range; spatial smoothness; periodicity; symmetry.
[0042] II. Sensitivity characteristics (1-dimensional, always present): point-level sensitivity index.
[0043] III. Spatial Boundary Derived Features (Add as needed): See the table below for details. Step S2: Data-driven point-level sensitivity index calculation Based on all training samples, a globally unified point-level sensitivity index is calculated for weight allocation in all layer models. This invention does not employ traditional manual partitioning; instead, it directly calculates the deformation sensitivity of each coordinate point independently, fundamentally eliminating the subjectivity and insufficient adaptability of manual partitioning. The mean, standard deviation, and maximum value of the deformation sequence (i.e., the sequence of deformation values for the same coordinate point in all training samples) after outlier removal in step S12 are calculated for each coordinate point. After global maximum normalization of the three indicators, they are weighted and summed according to preset weights to obtain the point-level sensitivity index. The point-level spatial risk prior and dual-utilization mechanism, as well as the three-step Bayesian optimization, are attached. Figure 2 As shown, the specific calculation formula is as follows: in: This is the mean absolute value of the deformation at this coordinate point after IQR anomaly removal, reflecting the average severity of the deformation at this point; The standard deviation of the corresponding deformation sequence reflects the fluctuation stability of the deformation at that point under different process conditions; The maximum value of the corresponding deformation sequence reflects the extreme deformation risk at that point; the denominator term D max , , These are the global maximum values of the statistics corresponding to all coordinate points on the surface of the component, used for normalization to unify the three indicators to the [0,1] dimension. , , The weighting coefficients of the three indicators satisfy the following conditions: + + =1, and , , ≥0.
[0044] Weight , , Determining the value is the core issue at the feature construction level -- if equal weight allocation (i.e. = = If the value is 1 / 3, then the implicit assumption that the mean, standard deviation, and maximum value contribute equally to deformation risk lacks theoretical basis. In reality, the dominance of these three attributes differs significantly under different ply configurations: in carbon fiber-dominated configurations, ply parameters determine the main deformation law, and extreme value risk is often concentrated under specific ply angle combinations; in quartz-dominated configurations, boundary constraints dominate the deformation distribution, and fluctuation stability (standard deviation) may be more critical. Therefore, this invention employs a Bayesian adaptive optimization method at the feature level to automatically determine the optimal weight ratio in a data-driven manner. in To reserve one configuration for cross-validation goodness of fit, optimization employs the T (Tree-structured Parzen Estimator) sampler from the Optuna framework, with a search space of: , [0,1], =1- - This optimization process addresses the questions of "what indicators to use and how to weight them" at the feature construction level, making the definition of the sensitivity index itself data-adaptive.
[0045] Step S3: Adaptive Feature Selection Based on Carbon Fiber Ratio This invention addresses the differences in mechanical properties between carbon fiber and quartz fiber by constructing an adaptive feature selection mechanism driven by the carbon fiber percentage. The adaptive feature configuration strategy driven by the carbon fiber percentage is as follows: Figure 3 As shown.
[0046] S31 carbon fiber percentage calculation For any layup configuration, calculate its carbon fiber percentage R. c The calculation formula is: S32 Threshold Determination and Feature Matching The optimal decision threshold was determined to be 0.45 through leave-one-out cross-validation. This threshold is the optimal boundary value that balances the prediction accuracy of both types of configurations. The following adaptive rule was then executed: When R c When the value is greater than 0.45, it is determined to be a carbon fiber dominant configuration. Since carbon fibers exhibit strong anisotropy, curing deformation is mainly determined by layup parameters. Therefore, only layup characteristics and process characteristics are used as inputs, without introducing spatial characteristics, to avoid redundant spatial information interfering with the learning of the dominant law of layup parameters; for example, an 8-layer pure carbon fiber configuration R... c =1.0, 12-layer carbon / quartz 7:3 hybrid configuration (R c (≈0.58) all adopt this feature scheme.
[0047] When R c When the value is ≤0.45, it is determined to be a quartz-dominated or balanced hybrid configuration. Since quartz material is approximately isotropic, solidification deformation is mainly controlled by spatial boundary constraints. Therefore, all three types of features are introduced to supplement prior spatial information and improve the model's ability to capture the spatial distribution of deformation; for example, a 16-layer pure quartz configuration (R... c =0), 20-layer carbon / quartz 3:7 hybrid configuration (R c =0.3) all adopt this feature scheme.
[0048] Step S4 introduces the prediction model training with continuous sample weights. Based on the point-level sensitivity index, continuous loss weights are assigned to each coordinate point sample. The LightGBM prediction model is trained in layers and types according to "number of layers + material type". The model training is completed by combining regularization strategy and parameter optimization.
[0049] S41 Continuous Sample Weight Allocation Based on a point-level sensitivity index, continuous loss weights are assigned to each coordinate point sample. These weights change continuously with sensitivity, replacing the traditional method of fixed weights for discrete partitioning. This achieves fine-grained point-level control, enabling the model to automatically increase the fitting weights for highly sensitive points during training. The specific calculation formula is consistent with the invention content.
[0050] Weight scaling factor The optimization employs a weighted layer Bayesian adaptive optimization method. Based on the optimal weights in the feature layer, it searches for the optimal value in the interval [0.1, 2.0] with leave-one-out configuration cross-validation as the objective. The value balances the enhancement effect in high-risk areas with the model's generalization ability.
[0051] S42 model parameter optimization A two-stage hyperparameter optimization strategy is adopted: the first stage determines the optimal range of core parameters through grid search, and the second stage performs fine-tuning through Bayesian optimization, i.e., model-level Bayesian adaptive optimization. Based on the optimal parameters determined in the first two steps, the optimal hyperparameter Θ of LightGBM is searched with leave-one-out configuration cross-validation as the objective. in, Θ = {learning_rate, max_depth, num_leaves, n_estimators, subsample, colsample_bytree, reg_alpha, reg_lambda}, where Ω is the preset search range for each hyperparameter, and the number of trials is 50. The objective function for all three optimization steps is based on a generalized R-squared with leave-one-out configuration cross-validation. 2 This ensures that the optimization direction is aimed at improving the model's generalization ability.
[0052] In this embodiment, the core parameters of all models are uniformly set as follows: learning rate 0.05, maximum depth 8, number of leaf nodes 63, number of decision trees 800, L1 regularization parameter reg_alpha=0.1, and L2 regularization parameter reg_lambda=0.2. Regularization is used to control model complexity and avoid overfitting.
[0053] Step S5: Rapid prediction of multi-point deformation of the configuration to be predicted Input the process parameters of the layup configuration to be predicted, calculate the carbon fiber ratio and match the corresponding feature scheme and training model, and output the deformation prediction value of all grid coordinates on the surface of the component point by point.
[0054] S51 Input Parameter Preprocessing Input the layup parameters (material type, layup angle, layup sequence) and process parameters (heating rate, cooling rate, holding time, curing pressure) of the configuration to be predicted, and the system will automatically calculate the total number of layers, carbon fiber ratio, layup ratio at each angle, layup symmetry coefficient, and other derived parameters.
[0055] S52 Feature and Model Automatic Matching Based on the calculated carbon fiber percentage R c The adaptive feature selection rules are executed to generate the input feature set of the corresponding dimension; at the same time, the corresponding type of the trained prediction model of the corresponding layer group is matched according to the total number of layers and the carbon fiber proportion type.
[0056] S53 Point-by-Point Deformation Prediction Output The generated feature set is input into the matching prediction model, and the deformation prediction values of all coordinate points of the 41×41 grid are calculated in batches, outputting the complete surface deformation distribution results. In this embodiment, the complete prediction of a single configuration takes less than 0.1 seconds.
[0057] The method of this invention is evaluated using a leave-one-out cross-validation strategy. Compared with the traditional finite element simulation method, the method of this invention has significantly improved prediction efficiency, can support rapid iterative optimization of multiple schemes of process parameters, effectively reduce the number of trial productions for process optimization, and significantly reduce R&D costs and cycle.
[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An adaptive prediction method for curing deformation of composite material components based on point-level sensitivity, characterized in that, Includes the following steps: S1 Multi-Source Deformation Dataset Construction and Preprocessing: Collect curing deformation data of composite material components under different layup configurations, construct a "measured + simulated" fusion dataset, and obtain a standard sample set containing the deformation of all grid coordinate points on the component surface after standardization preprocessing, with each grid coordinate point as an independent training sample. S2 data-driven point-level sensitivity index calculation: The deformation sensitivity is calculated independently for each coordinate point. Based on the mean, standard deviation, and maximum value of the deformation sequence after outlier removal for each coordinate point, the point-level sensitivity index is obtained by normalizing the global maximum value and then weighting the results. S3 Adaptive Feature Selection Based on Carbon Fiber Ratio: For carbon fiber / quartz hybrid layup configurations, the carbon fiber ratio of the layup configuration is calculated. Based on the difference in mechanical properties between carbon fiber and quartz fiber, the combination of input features is adaptively selected: when the carbon fiber ratio is greater than a preset threshold, layup features, process features, and sensitivity features are used as inputs; when the carbon fiber ratio is not greater than the preset threshold, spatial boundary derived features are introduced as inputs based on layup features, process features, and sensitivity features. S4 introduces a prediction model training with continuous sample weights: Based on the point-level sensitivity index, a continuous loss weight is assigned to each coordinate point sample. The higher the sensitivity of the point, the greater the loss weight. The fixed deformation prediction model is trained using the weighted sample set. S5 Rapid Multi-Point Deformation Prediction for the To-Be-Predicted Configuration: Input the layup parameters and process parameters of the to-be-predicted ply configuration, calculate the carbon fiber ratio, match the corresponding feature scheme and the trained prediction model, and output the deformation prediction values of all grid coordinates on the component surface point by point.
2. The adaptive prediction method for curing deformation of composite material components based on point-level sensitivity according to claim 1, characterized in that, The standardization preprocessing in step S1 specifically includes: All sample grid coordinates were uniformly mapped to the [0,400]×[0,400] standard coordinate system to complete coordinate alignment; extreme outliers in the deformation of each configuration were removed using the interquartile range method; missing coordinate point deformations were filled using bilinear interpolation; min-max normalization was performed on the layup features and process features, and the original physical quantity values were retained for the deformation labels.
3. The adaptive prediction method for curing deformation of composite material components based on point-level sensitivity according to claim 1, characterized in that, The formula for calculating the point-level sensitivity index S(x,y) in step S2 is as follows: In the formula, This is the mean of the absolute values of the deformation after truncation at the coordinate point (x, y); This represents the standard deviation of the corresponding deformed sequence; D represents the maximum value of the corresponding deformed sequence. max , , These are the mean absolute value of deformation, the standard deviation of deformation, and the global maximum value of deformation at all coordinate points on the component surface, respectively. , , The weighting coefficients of the three indicators are such that their sum is 1 and none of them are less than 0.
4. The adaptive prediction method for curing deformation of composite material components based on point-level sensitivity according to claim 1, characterized in that, The specific rule for adaptive matching in step S3 is as follows: Calculate the carbon fiber percentage of the configuration: A threshold of 0.45 was set, and this threshold was determined through multiple sets of comparative experiments to be the optimal cutoff value that balances the prediction accuracy of both configurations; if R c If R > 0.45, it is determined to be a carbon fiber dominant configuration, and modeling is performed by inputting only layup characteristics, process characteristics, and sensitivity characteristics; if R c If the value is ≤0.45, it is determined to be a quartz-dominated or hybrid configuration. Modeling is performed by inputting layup characteristics, process characteristics, sensitivity characteristics, and spatial boundary derived characteristics.
5. The adaptive prediction method for curing deformation of composite material components based on point-level sensitivity according to claim 1, characterized in that, The formula for calculating the continuous loss weight w in step S4 is as follows: In the formula, is the weight adjustment coefficient; S(x,y) is the point-level sensitivity index of the corresponding coordinate point.
6. The adaptive prediction method for curing deformation of composite material components based on point-level sensitivity according to claim 1, characterized in that, The solidification deformation prediction model is the LightGBM model. During training, grid search combined with Bayesian optimization is used to adjust hyperparameters, and L1 and L2 regularization terms are set to control model complexity and avoid overfitting.
