A physical information neural network-based intelligent performance prediction method and system for automated multi-material fiber placement on complex curved surfaces of aircraft

Abstract

This invention relates to a physical information neural network-based intelligent performance prediction method and system for automated multi-material fiber placement on complex curved surfaces in aircraft. The invention includes: collecting fiber placement process parameters; performing standardized preprocessing and physical constraint-enhanced sampling to obtain an enhanced training dataset; constructing a high-order nonlinear sparse regression candidate library based on the enhanced training dataset; jointly identifying piecewise smooth physical differential equation systems and their mode switching logic using sparse Bayesian regression and Hidden Markov Models to obtain an embeddable inverse mechanism model; constructing a Physical Information Embedded Generative Adversarial Network (PI-GAN); collaboratively optimizing the generator and discriminator through adversarial training to output surface defect indices and mechanical performance indices; and initiating incremental self-learning when introducing new process scenarios to achieve model self-evolution. This invention achieves end-to-end accurate prediction of multi-dimensional quality indices such as surface defects and mechanical performance.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing of composite materials and artificial intelligence, and in particular to a physical information neural network intelligent performance prediction method and system for automated fiber placement (AFP) of complex curved surfaces in aircraft. Background Technology

[0002] Automated fiber placement technology, as a core process in the manufacturing of high-performance composite material components, has been widely applied in high-end equipment fields such as aerospace. It achieves high-precision forming of complex curved surface components by precisely controlling multi-dimensional process parameters such as placement speed, temperature, pressure, tension, placement angle, and turning radius. However, there is a highly nonlinear, strongly coupled, and multimodal mapping relationship between process parameters and the final component quality (such as surface defects: bubbles, gaps, wrinkles, overlaps, torsion; mechanical properties: tensile strength, flexural strength, interlaminar shear strength, compressive strength), making it difficult to accurately model and predict process performance. Currently, fiber placement process performance modeling mainly relies on two types of methods: empirical / mechanistic modeling and purely data-driven modeling.

[0003] Empirical and mechanistic modeling methods: Although these methods have a certain degree of physical interpretability, they rely heavily on prior knowledge and simplification assumptions, making it difficult to characterize complex dynamic behaviors such as multi-field coupling of thermo-mechanical-fluid processes, material nonlinearity, and time-varying boundary conditions in actual processes. Moreover, their modeling process requires a large number of trial-and-error experiments, which are time-consuming, costly, and have limited generalization ability, making it particularly difficult to adapt to new materials and new configuration scenarios.

[0004] Pure data-driven modeling: In recent years, researchers have attempted to introduce machine learning methods such as Support Vector Regression (SVR), XGBoost, LightGBM, and Multilayer Perceptron (MLP) to establish the mapping relationship between process parameters and quality indicators. These methods can fit complex nonlinear relationships to a certain extent, but they have significant limitations: (1) Small sample sensitivity: High-quality fiber placement experiments are costly and time-consuming, and the actual number of effective samples that can be obtained is limited, which can easily lead to model overfitting and poor generalization ability. (2) Lack of physical consistency: The model relies only on statistical correlation and ignores the underlying physical laws such as heat conduction, viscoelastic deformation, and interface bonding. The output results may violate basic physical constraints (such as energy conservation and stress continuity), resulting in unreliable prediction results that are difficult to guide engineering practice. (3) Weak multi-output coupling modeling capability: Existing models are mostly designed for single quality indicators, making it difficult to synchronously and collaboratively predict multi-dimensional coupled outputs such as surface defects and mechanical properties, and they are not robust enough to parameter disturbances.

[0005] To address these issues, some studies have attempted to introduce Generative Adversarial Networks (GANs) for data augmentation or combine Physical Information Neural Networks (PINNs) with partial differential equation constraints. However, existing GAN methods mostly employ fully connected or CNN structures, making it difficult to effectively model long-range dependencies and dynamic interactions between process parameter sequences. Traditional PINNs, on the other hand, are mostly based on feedforward networks, which have limited expressive power and lack mechanisms for effectively judging and guiding the authenticity and diversity of generated samples, making them prone to mode collapse or physical-data mismatch.

[0006] Crucially, there is currently a lack of a unified modeling paradigm that combines mechanistic interpretability, data adaptability, and physical consistency: one that can both inversely identify the dominant process mechanism (such as the dynamic switching logic under different deployment modes) from limited experimental data and integrate this mechanistic knowledge to guide the construction of deep models; one that can both generate high-fidelity, physically consistent virtual samples through adversarial learning to alleviate the bottleneck of small samples and achieve end-to-end collaborative prediction of multiple quality indicators.

[0007] Therefore, there is an urgent need to propose a new method for predicting process performance, which can break through the dual dilemma of "black box modeling - physical deviation" and "mechanism simplification - insufficient expression", and provide highly reliable, interpretable and self-evolving algorithmic support for intelligent planning, online monitoring and closed-loop control of automated fiber placement process. Summary of the Invention

[0008] To address this, this invention provides a physical information neural network-based intelligent performance prediction method and system for automated multi-material fiber placement on complex curved surfaces in aircraft. It integrates inverse mechanism modeling with a Physical-Informed Generative Adversarial Network (PI-GAN) for predicting and optimizing the process performance of fiber placement machines. The aim is to overcome the dual shortcomings of existing technologies, such as insufficient expressive power of mechanism modeling and physical deviations in purely data-driven models, as well as weak generalization with small samples. This provides a high-precision, highly interpretable, and self-evolving method for predicting the process performance of fiber placement machines. This method achieves end-to-end accurate prediction of multi-dimensional quality indicators such as surface defects (bubbles, gaps, wrinkles, overlaps, torsion) and mechanical properties (tensile strength, flexural strength, interlaminar shear strength, compressive strength) through a four-stage closed-loop process: inverse mechanism modeling, physical information embedding, adversarial enhancement, and multi-objective collaborative prediction. It is applicable to the performance prediction challenges of thermosetting / thermoplastic composite components during the wire layup process, which are caused by strong nonlinear and multi-coupling relationships between multidimensional process parameters, surface quality indicators, and mechanical performance indicators. It can be widely used in intelligent process planning, online quality prediction, and closed-loop parameter control systems for composite components in high-end equipment fields such as aviation and aerospace.

[0009] To address the aforementioned technical problems, this invention provides a physical information neural network-based intelligent performance prediction method for automated multi-material fiber placement on complex curved surfaces of aircraft, comprising: The collected fiber placement process parameters include continuous and discrete variables. The continuous variables include placement speed, placement pressure, placement head temperature, fiber tension, radius of curvature of turning path, ribbon width, and substrate temperature. The discrete variables include material type, process type, placement angle, plate geometry, and substrate material. The fiber placement process parameters are standardized and preprocessed, and physical constraint enhancement sampling is performed to obtain an enhanced training dataset; Based on the enhanced training dataset, a high-order nonlinear sparse regression candidate library is constructed; wherein, the high-order nonlinear sparse regression candidate library includes a heat conduction dominant term characterizing the heat exchange between the filament bundle and the substrate, a mechanical deformation coupling term characterizing compaction, shearing and buckling instability, and a flow spreading dynamics term characterizing the lateral spreading and edge instability of the filament bundle. Based on the aforementioned high-order nonlinear sparse regression candidate library, sparse Bayesian regression and hidden Markov models are used to jointly identify piecewise smooth physical differential equation systems and their mode switching logic, thereby obtaining an embeddable inverse mechanism model. A physical information embedding generative adversarial neural network (PI-GAN) is constructed, wherein the generator of the PI-GAN adopts the Transformer-PINN architecture, and the discriminator adopts a lightweight multilayer perceptron; the inverse mechanism model is embedded as a physical constraint term into the total loss function of the generator; wherein the total loss function of the generator includes adversarial loss, data fidelity loss, and physical constraint loss, and the physical constraint loss is constructed based on the piecewise smooth physical differential equation system and its mode switching logic; The generator and the discriminator are co-optimized using adversarial training to obtain a trained model. Based on the trained model, surface defect indices and mechanical performance indices are output. The mean square error is used as the evaluation index for the surface defect indices and mechanical performance indices. The surface defect indices include the number of bubbles, gaps, wrinkles, overlaps, and torsional deformations. The mechanical performance indices include tensile strength, flexural strength, interlaminar shear strength, and compressive strength. When a new process scenario is introduced, incremental self-learning is initiated. The historical model knowledge of the trained model is retained through knowledge distillation, and a representative sample replay strategy is combined to prevent catastrophic forgetting, thereby achieving model self-evolution.

[0010] In one embodiment of the present invention, the fiber placement process parameters are subjected to standardized preprocessing and physical constraint enhanced sampling to obtain an enhanced training dataset, including: The parameter set of the fiber placement process parameters as follows: ; in, Laying speed, mm / s; Laying pressure, N / mm; : Laying head temperature, °C; : Tow tension, N; : Radius of curvature, mm; Ribbon width, mm; ∈{carbon fiber / epoxy, glass fiber / PEEK,…}: Material type; ∈{Automatic Fiber Placement (AFP), Automatic Tape Placement (ATL)}: Process category; ∈{0° ,±45° ,90°}: Paving angle; ∈{flat plate, hyperbolic paraboloid}: plate type; Substrate material; Substrate temperature, °C; : A 12-dimensional real vector space; The continuous variables are normalized using Z-score or Min-Max: ; in, , These are the mean and standard deviation of the training set, respectively, to prevent differences in dimensions from dominating gradient updates; For the normalized first One variable value; For the index of the variable; Embedding encoding of the discrete variables: for categorical variables Construct learnable embeddings: ; ; ; in, An embedding vector is a real-valued vector that transforms a discrete category into a vector to preserve semantic information for neural network learning. For embedding weight matrix; It is a one-hot encoding function; For the embedded dimension; For the number of categories, Preserve semantic similarity; The physical constraint-enhanced sampling includes: For minority class samples Nearest neighbor convex hull linear interpolation generates candidate samples: ; ; in, For newly generated candidate samples; This is the currently selected original sample; Neighboring samples; These are the interpolation coefficients; A uniform distribution between 0 and 1; Tomek Links are used to remove boundary noise pairs, preserving physically consistent samples. A physical feasibility constraint is introduced for filtering; the constraint is as follows: ; and ; Where R is the radius of curvature of the turning path; Temperature of the laying head; Minimum allowable turning radius for the material; The glass transition temperature of the resin matrix; For safety margin.

[0011] In one embodiment of the present invention, a high-order nonlinear sparse regression candidate library is constructed based on the enhanced training dataset, including: High-order nonlinear sparse regression candidate library These can be represented as the following three types of basis functions: The dominant heat conduction term includes: Fourier thermal conductivity: ; in, For the Laplace operator; Temperature of the laying head, in °C; The substrate temperature is ℃; Convection heat transfer terms: ; in, The convective heat transfer coefficient; Phase transition latent heat approximation term: ; Where H is the Heaviside function, which characterizes the abrupt change from the glassy state to the elastic state; is the glass transition temperature; m is the polynomial power; The mechanical deformation coupling term includes: Pressure Force – Gap Relationship: ; in, Laying pressure, N / mm; Width of the ribbon, in mm; The radius of curvature of the bend is in mm; Influence of bending stiffness: ; in , representing curvature; For the laying angle; , representing the equivalent bending stiffness; Frictional shear stress: ; in, The coefficient of friction is temperature-dependent. The flow spreading dynamics term includes: Capillary-related terms: ; in, The apparent viscosity of the resin; Surface tension; The laying speed is in mm / s; for the Arrhenius type, ; The reference viscosity is the theoretical limiting viscosity value at infinitely high temperatures. The activation energy is the minimum energy required for molecules to flow. The molar gas constant; Related terms of Weissenberg number: ; in, The relaxation time characterizes the elastic effect; All candidates passed the dimension consistency check to ensure that each one is dimensionless or has a definite physical dimension. The final high-order nonlinear sparse regression candidate library is as follows: ; Its i-th line is: ; Where N is the number of samples; P≈45, including constant terms, first / second order single terms, and interaction terms; Let P be the dimension of the row vector, and P be the total number of all possible physical terms. Let be the state vector at the i-th position.

[0012] In one embodiment of the present invention, sparse Bayesian regression and hidden Markov models are jointly used to identify piecewise smooth physical differential equation systems and their mode switching logic, thereby obtaining an embeddable inverse mechanism model, including: Set quality indicators The time / space evolution satisfies the implicit differential equation: ; ; in, This represents 9 quality indicators; , is a sparse coefficient matrix; sparse coefficient matrix The dimensions are P rows corresponding to physical items in the candidate library, and 9 columns corresponding to 9 quality indicators; It is Gaussian noise; For noise variance; It is the identity matrix; This serves as a candidate library for high-order nonlinear sparse regression. Solving SBL using Sparse Bayesian learning: ; in, It is the probability density function; For the observation data matrix; These are hyperparameters that control the relevance of each weight. Among them, the prior is set as a hierarchical automatic correlation determination prior: ; in, These are the weighting coefficients corresponding to the k-th candidate physics item; It is a 9-dimensional identity matrix; For the corresponding Weight parameters; The shape parameter of the Gamma distribution; is the scale parameter of the Gamma distribution; Iterative updates are achieved by maximizing the expected value of EM or by variational inference. With Ξ, automatic model selection and uncertainty quantification are achieved; in, The objects updated iteratively by the algorithm are all value; When process parameters cross critical thresholds, the quality evolution law undergoes discontinuous abrupt changes, requiring the introduction of an event-driven mode switching mechanism. Define the set of switching surfaces Each switching surface is defined by physical critical conditions: ; Where m is the total number of switching surfaces / modes; Let be the state space, representing the mathematical space containing all possible combinations of process parameters; ; ;

[0013] ; ; in, This is the equivalent elastic modulus of the composite material; The thickness of the prepreg tape / filament bundle; Patch length or effective contact length; The basic glass transition temperature; These are the empirical fitting coefficients; For different devices, The values ​​differ: 0.18 for AFP models and 0.32 for ATL models. ; Furthermore, a hidden Markov model is used to analyze the unobserved modes. Identify states ∈{1,2,…,K}, where K is the total number of hidden states: Observation sequence: quality indicators; State transition: ; in, , Let be the hidden state variables at time t and time t-1; This is the state transition matrix, which represents the transition from the state at the previous time step. Change to the current state The probability of; Launch distribution: ; in, The observation vector is the actual measured quality index data. An index for a certain state; For the first The covariance matrix under each state; By training using the Baum-Welch algorithm or ELBO maximization, we obtain: Optimal number of modes ; Mechanism equations corresponding to each mode ; Mapping of switching probability matrix A to the critical surface ; The final output is a piecewise smooth physical system, i.e., the inverse mechanism model: ; in, For the first Residual / noise under these modes; For the first in the state space Sub-regions; Indicates the switching surface of the system; The system is in two states

[0014] In one embodiment of the present invention, the total loss function is: ; in, and For dynamic weighting coefficients, Number of training iterations; adversarial loss It is expressed as follows: ; in, For discriminator networks; Data fidelity loss It is expressed as follows: ; Among them, weight Dynamically updated based on historical failure analysis; The first of the quality metrics calculated for the generator network One quality indicator; The physical constraint loss is expressed as follows: ; in, ; ; in, For uncertainty weighting matrix, The posterior standard deviation of the output of sparse Bayesian learning.

[0015] In one embodiment of the present invention, the discriminator employs WGAN-GP loss to improve training stability and sample diversity, and its expression is as follows: ; in: , for Wasserstein distance estimation; Let $\mathematical_expectation$ represent the average value over the sampling process. Scoring based on real data; To generate scores for the data; This is a gradient penalty;

[0016] ; , where is the penalty coefficient; The actual value; This is a predicted value; For discriminator networks, it refers to the process of interpolating samples. The output score obtained by inputting into the discriminator; The gradient operator is used to calculate the output of the discriminator. Relative to its input The rate of change; It is an L2 norm.

[0017] In one embodiment of the present invention, the evaluation is performed using mean square error as an evaluation index, including: For the nine-dimensional quality indicators, the following evaluation indicators are constructed: ; ; ; in, For predicted values, The actual value is the label; Representing the There are 9 quality indicators in total. Representing the One sample; The mean squared error should be as small as possible, but it is sensitive to large errors. The smaller the mean absolute error, the better, as it reflects the actual error distance; The coefficient of determination is as close to 1 as possible, indicating a high degree of model fit.

[0018] In one embodiment of the present invention, the historical model knowledge of the trained model is preserved through knowledge distillation, including: Define attention-guided distillation loss: ; in, It is softmax; Temperature coefficient; This is the generator output for the old model, i.e., the teacher model; The generator output is for the new model, i.e., the student model; This is the self-attention matrix of the last layer of the encoder; .

[0019] In one embodiment of the present invention, the representative sample replay employs a feature space nearest neighbor replay strategy, including: A high-dimensional feature library of old task data is built offline. When training a new task, k nearest neighbor old samples are dynamically retrieved for each new sample and mixed in a preset ratio for fine-tuning; where k=5 and the mixing ratio is 1:9.

[0020] This invention also provides a physical information neural network intelligent performance prediction system for automated multi-material fiber placement on complex curved surfaces of aircraft, comprising: The data acquisition module is used to acquire fiber placement process parameters; wherein, the fiber placement process parameters include: continuous variables and discrete variables, the continuous variables include placement speed, placement pressure, placement head temperature, fiber tension, turning path curvature radius, ribbon width and substrate temperature; the discrete variables include: material type, process type, placement angle, plate geometry and substrate material; The data preprocessing and enhancement module is used to perform standardized preprocessing and physical constraint enhancement sampling on the wire laying process parameters to obtain an enhanced training dataset. The candidate library construction module is used to construct a high-order nonlinear sparse regression candidate library based on the enhanced training dataset; wherein, the high-order nonlinear sparse regression candidate library includes a heat conduction dominant term characterizing the heat exchange between the filament bundle and the substrate, a mechanical deformation coupling term characterizing compaction, shearing and buckling instability, and a flow spreading dynamics term characterizing the lateral spreading and edge instability of the filament bundle. The inverse mechanism modeling module is used to identify piecewise smooth physical differential equation systems and their mode switching logic based on the high-order nonlinear sparse regression candidate library, using sparse Bayesian regression and hidden Markov models, to obtain an embeddable inverse mechanism model. A network construction module is used to construct a Physical Information Embedded Generative Adversarial Neural Network (PI-GAN). The generator of the PI-GAN adopts a Transformer-PINN architecture, and the discriminator adopts a lightweight multilayer perceptron. The inverse mechanism model is embedded as a physical constraint term in the generator's total loss function. The generator's total loss function includes adversarial loss, data fidelity loss, and physical constraint loss, wherein the physical constraint loss is constructed based on the piecewise smooth physical differential equation system and its mode switching logic. An optimization module is used to collaboratively optimize the generator and the discriminator in an adversarial training manner to obtain a trained model. Based on the trained model, surface defect indices and mechanical performance indices are output. The surface defect indices and mechanical performance indices are evaluated using mean square error as the evaluation index. The surface defect indices include the number of bubbles, gaps, wrinkles, overlaps, and torsional deformations. The mechanical performance indices include tensile strength, flexural strength, interlaminar shear strength, and compressive strength. The incremental self-learning module is used to initiate incremental self-learning when a new process scenario is introduced. It retains the historical model knowledge of the trained model through knowledge distillation and combines a representative sample replay strategy to prevent catastrophic forgetting, thereby achieving model self-evolution.

[0021] The technical solution of the present invention has the following advantages compared with the prior art: The present invention discloses a physical information neural network intelligent performance prediction method and system for automated multi-material fiber placement on complex curved surfaces of aircraft. This method and system can be widely applied to process digital twin systems, online quality prediction modules, and closed-loop parameter control systems in automated fiber placement production lines for aerospace composite materials. It provides a new generation of intelligent prediction paradigms for high-end equipment manufacturing, characterized by "mechanism guidance, data-driven approach, and continuous evolution," and offers the following beneficial effects: 1. Unified Interpretability of Mechanisms and Flexibility of Data: A clear physical framework is constructed through reverse mechanism modeling, and PI-GAN is introduced to efficiently characterize the details of process data. While maintaining physical consistency, the expressive power of the model is improved, avoiding the black box risk of pure data-driven models, and supporting process engineers to understand, verify and correct the model results.

[0022] 2. Strong generalization ability under small sample conditions: Generative adversarial networks are used to generate high-fidelity virtual samples that conform to physical laws. Combined with the SMOTE data augmentation strategy, the problem of data scarcity caused by high experimental costs and difficulty in data acquisition is effectively alleviated, and the prediction stability and generalization performance of the model in small sample scenarios are significantly improved.

[0023] 3. Multi-quality index collaborative prediction capability: Through end-to-end modeling, it simultaneously outputs 9-dimensional coupled quality indices, breaking through the limitations of traditional single quality index modeling, realizing the collaborative prediction of surface defects and mechanical properties, and providing a unified data foundation for comprehensive quality assessment and multi-objective process parameter optimization.

[0024] 4. Dynamic operating condition adaptive capability: By explicitly modeling the subsystem switching logic and combining it with an incremental learning mechanism, the model can be adaptively updated to non-stationary operating conditions such as material switching and curvature abrupt changes, ensuring the continuity and reliability of predictions in the process of complex process evolution.

[0025] 5. Engineering-friendly design: It adopts a standardized input / output interface design, combined with a lightweight discriminator structure and model compression technology, including knowledge distillation and pruning, to reduce computing resource requirements while ensuring prediction accuracy, and meet the requirements of real-time reasoning and stable operation in industrial sites. Attached Figure Description

[0026] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0027] Figure 1 This is a flowchart of the physical information neural network intelligent performance prediction method for automatic filament placement of multiple materials on complex curved surfaces of aircraft, based on the present invention.

[0028] Figure 2 This is the segmented smooth physical system and mode switching logic diagram of the present invention.

[0029] Figure 3 This is a verification diagram of the collaborative prediction of quality indicators in this invention. Detailed Implementation

[0030] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0031] In this invention, when directions (up, down, left, right, front, and back) are described, it is only for the convenience of describing the technical solution of this invention, and does not indicate or imply that the technical features referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0032] In this invention, "several" means one or more, "multiple" means two or more, "greater than," "less than," "exceeding," etc., are understood to exclude the stated number; "above," "below," "within," etc., are understood to include the stated number. In the description of this invention, the terms "first" and "second" are used only to distinguish technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0033] In this invention, unless otherwise explicitly defined, the terms "setting," "installing," and "connecting" should be interpreted broadly. For example, they can refer to a direct connection or an indirect connection through an intermediate medium; a fixed connection, a detachable connection, or an integrally formed connection; a mechanical connection, an electrical connection, or a connection capable of mutual communication; or the internal connection of two components or the interaction between two components. Those skilled in the art can reasonably determine the specific meaning of the above terms in this invention based on the specific content of the technical solution.

[0034] Example 1 Reference Figure 1 As shown, this embodiment provides a physical information neural network intelligent performance prediction method for automated multi-material fiber placement on complex curved surfaces of aircraft, including... Phase 1: Data Standardization and Enhancement.

[0035] 1.1 Input the set of process parameters.

[0036] Collect the following parameters for the fiber placement process: placement speed, placement pressure, placement temperature, fiber tension, turning radius, ribbon width, material type, process category, placement angle, substrate type, substrate material, and substrate temperature. The original set of parameters for the fiber placement process is as follows: ; in, Laying speed (mm / s); Laying pressure (N / mm); : Laying head temperature (°C); : Tow tension (N); : Radius of curvature (mm); : Ribbon width (mm); ∈{carbon fiber / epoxy, glass fiber / PEEK,…} : Material type (discrete); ∈{Automatic Fiber Placement (AFP), Automatic Tape Placement (ATL)}: Process category (discrete); ∈{0° ,±45° ,90°} : Paving angle (discrete); ∈{flat plate, hyperbolic paraboloid} : Plate type (discrete); Substrate material (discrete); Substrate temperature (°C); : A 12-dimensional real vector space.

[0037] The physical meaning and dimensions of each parameter are shown in Table 1.

[0038] Table 1: Physical Meaning and Dimensions of Parameters

[0039] 1.2 Standardized preprocessing.

[0040] For continuous parameters, Z-score or Min-Max normalization is applied: ; in, , These are the mean and standard deviation of the training set, respectively, to prevent differences in dimensions from dominating gradient updates; For the normalized first One variable value; For the index of the variable; Embedding encoding of the discrete variables: for categorical variables Construct learnable embeddings: ; ; ; in, An embedding vector is a real-valued vector that transforms a discrete category into a vector to preserve semantic information for neural network learning. For embedding weight matrix; It is a one-hot encoding function; For the embedded dimension; For the number of categories, Preserve semantic similarity.

[0041] 1.3 Physically Constrained Enhanced Sampling (GSMOTE+).

[0042] To address the small sample size problem, a hybrid strategy combining SMOTE / ADASYN oversampling and Tomek Links undersampling is employed to generate virtual samples in the parameter space that conform to the process characteristics, thereby expanding the training set size and improving class imbalance. An augmented sample generator is defined. Based on the SMOTE-ENN hybrid strategy: (1) Generate candidate samples by linear interpolation within the K-nearest neighbor convex hull of the minority class (such as samples with high defect rates): ; ; in, For newly generated candidate samples; This is the currently selected original sample; Neighboring samples; These are the interpolation coefficients; A uniform distribution between 0 and 1; Tomek Links are used to remove boundary noise pairs, preserving physically consistent samples.

[0043] (2) Use Tomek Links to remove boundary noise pairs and retain physically consistent samples; (3) Introduce physical feasibility constraints: Only retain virtual samples that meet the following conditions, the constraints being: ; and ; Where R is the radius of curvature of the turning path; Temperature of the laying head; Minimum allowable turning radius for the material; The glass transition temperature of the resin matrix; For safety margin.

[0044] Phase 2: Reverse mechanism modeling and dynamic switching logic identification.

[0045] This phase aims to automatically extract interpretable, differentiable, and embeddable dominant physical laws from high-dimensional, nonlinear, and non-stationary process-quality data, and construct a piecewise smooth, mode-switching differential equation system as the hard constraint skeleton for subsequent Physical Information Neural Networks (PINN). Its core idea is: "Data-driven structure discovery, physical priors guarantee causality," avoiding the deviation risks of pure black-box modeling, while overcoming the modeling failure problem caused by strong multi-field coupling and complex boundaries in traditional forward mechanism modeling.

[0046] 2.1 Construction of candidate function libraries.

[0047] A physics-guided model is constructed to address the thermo-mechanical-fluid-structure interaction and defect evolution mechanisms during automated fiber placement (AFP / ATL) processes, balancing mathematical expressiveness with engineering interpretability. A high-order nonlinear sparse regression candidate library is also built. It is spanned by the following three types of basis functions: (1) The dominant thermal conductivity term (characterizing heat exchange between the filament bundle and the substrate) includes: Fourier thermal conductivity: ; in, For the Laplace operator; Temperature of the laying head, in °C; The substrate temperature is ℃; Convection heat transfer terms: ; in, The convective heat transfer coefficient; Phase transition latent heat approximation term: ; Where H is the Heaviside function, which characterizes the abrupt change from the glassy state to the elastic state; is the glass transition temperature; m is the polynomial power.

[0048] (2) Mechanical deformation coupling terms (characterizing compaction, shear, and buckling instability), including: Pressure Force – Gap Relationship: ; in, Laying pressure, N / mm; Width of the ribbon, in mm; The radius of curvature of the bend is in mm; Influence of bending stiffness: ; in , representing curvature; For the laying angle; , representing the equivalent bending stiffness; Frictional shear stress: ; in, The coefficient of friction is temperature-dependent. (3) Flow spreading dynamics (characterizing lateral spreading and edge instability of the filament bundle), including: Capillary-related terms: ; in, The apparent viscosity of the resin; Surface tension; The laying speed is in mm / s; for the Arrhenius type, ; The reference viscosity is the theoretical limiting viscosity value at infinitely high temperatures. The activation energy is the minimum energy required for molecules to flow. The molar gas constant; Related terms of Weissenberg number: ; in, The relaxation time characterizes the elastic effect.

[0049] All candidate items have passed dimensional consistency verification to ensure that each item is dimensionless or has a clear physical dimension (such as stress / energy / strain rate) to avoid spurious correlations.

[0050] The final high-order nonlinear sparse regression candidate library is as follows: ; Its i-th line is: ; Where N is the number of samples; P≈45, including constant terms, first / second order single terms, and interaction terms; Let P be the dimension of the row vector, and P be the total number of all possible physical terms. Let be the state vector at the i-th position.

[0051] 2.2 Sparse Bayesian Regression Identification.

[0052] Variational Sparse Bayesian Learning (VSBL) is employed to automatically identify high-dimensional nonlinear dynamic equations. A quality index is set... The time / space evolution satisfies the implicit differential equation: ; ; in, This represents 9 quality indicators; , is a sparse coefficient matrix; sparse coefficient matrix The dimensions are P rows corresponding to physical items in the candidate library, and 9 columns corresponding to 9 quality indicators; It is Gaussian noise; For noise variance; It is the identity matrix; It serves as a candidate library for high-order nonlinear sparse regression.

[0053] Solving using Sparse Bayesian Learning (SBL): ; in, It is the probability density function; For the observation data matrix; These are hyperparameters that control the relevance of each weight. The prior is set as the Hierarchical Automatic Relevance Determination (ARD) prior: ; in, These are the weighting coefficients corresponding to the k-th candidate physics item; It is a 9-dimensional identity matrix; For the corresponding Weight parameters; The shape parameter of the Gamma distribution; is the scale parameter of the Gamma distribution; Iterative updates are achieved by maximizing the expected value of EM or by variational inference. With Ξ, automatic model selection and uncertainty quantification are achieved; in, The objects updated iteratively by the algorithm are all value.

[0054] 2.3 Subsystem switching logic modeling.

[0055] Reference Figure 2 As shown, when the process parameters cross the critical threshold, the quality evolution law undergoes a discontinuous jump, requiring the introduction of an event-driven mode switching mechanism.

[0056] When process parameters cross critical thresholds, the quality evolution law undergoes discontinuous abrupt changes, requiring the introduction of an event-driven mode switching mechanism. Define the set of switching surfaces Each switching surface is defined by physical critical conditions: ; Where m is the total number of switching surfaces / modes; Let be the state space, representing the mathematical space containing all possible combinations of process parameters; ; ;

[0057] ; ; in, This is the equivalent elastic modulus of the composite material; The thickness of the prepreg tape / filament bundle; Patch length or effective contact length; The basic glass transition temperature; These are the empirical fitting coefficients; For different devices, The values ​​differ: 0.18 for AFP models and 0.32 for ATL models. ; Furthermore, a Hidden Markov Model (HMM) is used to analyze the unobserved modes. Identify ∈{1,2,…,K}: Observation sequence: quality indicators; State transition: ; in, , Let be the hidden state variables at time t and time t-1; This is the state transition matrix, which represents the transition from the state at the previous time step. Change to the current state The probability of; Launch distribution: ; in, The observation vector is the actual measured quality index data. An index for a certain state; For the first The covariance matrix under each state; By training using the Baum-Welch algorithm or ELBO maximization, we obtain: Optimal number of modes ; Mechanism equations corresponding to each mode ; Mapping of switching probability matrix A to the critical surface ; The final output is a piecewise smooth physical system, i.e., the inverse mechanism model: ; in, For the first Residual / noise under these modes; For the first in the state space Sub-regions; Indicates the switching surface of the system; The system is in two states

[0058] This model can directly embed the physical loss term of PINN as a hard / soft switching constraint (through a smooth maximization function). (Approximately achieves differentiable switching).

[0059] Phase 3: Constructing a Physical Information Embedded Generative Adversarial Neural Network (PI-GAN).

[0060] This stage uses the piecewise smooth physical system identified in stage 2. To strengthen prior knowledge, a Physics-Informed Generative Adversarial Network (PI-GAN) is constructed to achieve collaborative optimization of data-driven modeling and mechanistic constraints. The overall architecture consists of a Transformer-PINN generator and a lightweight discriminator, forming an end-to-end differentiable, interpretable, and embeddable adversarial learning framework.

[0061] 3.1 Detailed explanation of the Transformer-PINN generator structure.

[0062] generator With standardized process parameters As input, output a 9-dimensional quality index prediction. To adapt to the sequence modeling paradigm, it is treated as a "pseudo-temporal" input of length L=12 (each dimension is treated as an independent time step), and sequence information is injected through learnable positional encoding.

[0063] (1) Input embedding and position encoding Discrete variable embedding: for Using the embedding vector from stage 1 ; Continuous variables can be used directly after standardization. Encode into a unified feature vector ,in =512; Incorporate Learned Positional Encoding: ; in, This is the initial feature input vector; It is a size of The matrix ( It is the sequence length. (It is the feature dimension). This represents the vector corresponding to the i-th position in the input sequence; These are the mean and standard deviation of a normal distribution.

[0064] (2) Encoder Layer.

[0065] The number of encoder stacking layers is Layer, each layer executes sequentially: Multi-head self-attention mechanism: ; ; in, These are the query, key, and value matrices, which are the core of the self-attention mechanism. The model determines how much attention weight to give to the value by calculating the matching degree (dot product) between the query and the key.

[0066] ; Among them, the number of attention heads Attention mechanism linear projection weight matrix Multi-head attention weight matrix .

[0067] Residual connectivity and layer normalization: ; Feedforward Neural Network (FFN): ; Wherein, the input weight matrix Intermediate layer dimensions of feedforward networks (FFN) GELU is the Gaussian error linear unit activation function, which improves the nonlinear fitting ability. This represents the bias term.

[0068] Second residual connection: ; (3) Decoder Layer.

[0069] Decoders are also stacked Each layer contains three sub-layers: Masked Multi-Head Self-Attention: Prevents future information leakage (although this task is static prediction, the standard structure is still retained to be compatible with temporal extensions). Encoder-decoder cross-attention: ; Among them, decoder As the querying party, I went to examine the process features extracted by the encoder. This allows the decoder to dynamically focus on the combination of process parameters most relevant to the current prediction in the encoder output.

[0070] FFN + Residual + LayerNorm (same as encoder).

[0071] (4) Output layer decoder last layer output The final prediction is obtained through linear projection: ; in, This is the output vector of the last layer of the decoder; , These are the weights and biases for the output layer.

[0072] The output order must strictly correspond: ; 3.2 Discriminator Design (Lightweight Discriminator).

[0073] Discriminator This is a 3-layer MLP, with a quality index vector as input, and the structure is as follows: ; ; ; in, Hidden layer states; The discrimination probability is the final result (between 0 and 1) output by the discriminator, representing the probability that the input sample is "true".

[0074] Gradient penalty term: Introducing Wasserstein-GP constraint to improve training stability: ; in

[0075] ; This is the penalty coefficient; Let $\mathematical_expectation$ represent the average value over the sampling process. For interpolation distribution, it refers to the above. The constructed sampling space; For discriminator networks, it refers to the network that interpolates samples. The output score obtained by inputting into the discriminator; The gradient operator is used to calculate the output of the discriminator. Relative to its input The rate of change; It is an L2 norm.

[0076] Phase 4: Collaborative optimization of adversarial training and multi-target loss.

[0077] This stage constructs an adversarial training framework for alternating generators and discriminators, and designs a multi-objective weighted loss function to achieve synergistic optimization of data fitting accuracy, physical consistency, and distribution realism. The core idea is: "Drive distribution alignment with adversarial training, ensure causal rationality with physical constraints, and achieve dynamic balance with adaptive scheduling," avoiding the mode collapse and physical anomalousness problems in traditional GAN ​​training.

[0078] 4.1 Design of the generator's total loss function.

[0079] generator The goal is to generate high-quality, physically consistent quality indicator predictions. Its total loss is a weighted sum of three terms: , The definitions of each sub-item are as follows: (1) Adversarial Loss.

[0080] Using non-saturated GAN loss (NS-GAN) to improve the stability of the generator gradient signal: ; in, For the discriminator network.

[0081] (2) Data Fidelity Loss.

[0082] Component-weighted mean square error is used to assign different importance to different quality indicators (based on prior process sensitivity): ; Among them, weight Dynamically updated based on historical failure analysis; The first of the quality metrics calculated for the generator network One quality indicator.

[0083] (3) Physical constraint loss The two types of physical knowledge identified in the second stage of integration: ; in, ; ; in For uncertainty weighting matrix, The posterior standard deviation of the output of Sparse Bayesian Learning (SBL) in Stage 2.

[0084] 4.2 Discriminator Loss Function.

[0085] Discriminator D employs the standard Wasserstein GAN with Gradient Penalty (WGAN-GP) loss to improve training stability and sample diversity. ; in: , for Wasserstein distance estimation; For discriminator networks; Scoring based on real data; To generate scores for the data; This is a gradient penalty;

[0086] ; , where is the penalty coefficient; The actual value; This is a predicted value; For discriminator networks, it refers to the process of interpolating samples. The output score obtained by inputting into the discriminator; The gradient operator is used to calculate the output of the discriminator. Relative to its input The rate of change; It is an L2 norm.

[0087] Phase 5: Collaborative optimization of adversarial training and multi-target loss.

[0088] This phase constructs a closed-loop enhancement mechanism of "evaluation-driven optimization and optimization-supported evolution" to achieve a leap from "static high precision" to "dynamic self-evolution." The core objectives are: to quantitatively verify the model's generalization ability on multi-dimensional quality indicators; to automatically optimize the network structure and hyperparameters based on evaluation feedback; and to support online incremental learning under new materials, new configurations, and new processes, avoiding the waste of computing power and knowledge forgetting caused by retraining.

[0089] 5.1 Multi-dimensional model evaluation system.

[0090] For the nine-dimensional quality indicators, the following evaluation indicators are constructed: ; ; ; in, For predicted values, The actual value is the label; Representing the There are 9 quality indicators in total. Representing the One sample; The mean squared error should be as small as possible, but it is sensitive to large errors. The smaller the mean absolute error, the better, as it reflects the actual error distance; The coefficient of determination is as close to 1 as possible, indicating a high degree of model fit.

[0091] 5.2 Incremental self-learning system.

[0092] When introducing new technological scenarios, traditional retraining is costly and leads to the forgetting of old knowledge. This invention designs a three-stage incremental learning framework: (1) Knowledge Distillation.

[0093] set up For the old model (teacher). For the new model (student), define the attention-guided distillation loss: , in, It is softmax; Temperature coefficient; This is the generator output for the old model, i.e., the teacher model; The generator output is for the new model, i.e., the student model; This is the self-attention matrix of the last layer of the encoder; .

[0094] (2) Representative Replay Employing a Feature-space Nearest Neighbor Replay (FNNR) strategy: Offline construction of old data feature library; During training for a new task, for each new sample, retrieve its k=5 nearest neighbor old samples; Training is performed by mixing new and old samples at a ratio of 10%.

[0095] Example 2 This embodiment takes the automatic yarn laying process of thermosetting carbon fiber prepreg on convex curved surface components as the background, and aims to achieve high-precision collaborative prediction of 9 quality indicators (5 surface defects: number of bubbles, number of gaps, number of wrinkles, number of overlaps, number of twists; 4 mechanical properties: tensile strength, flexural strength, interlaminar shear strength, compressive strength).

[0096] Step 1: Experimental Data Preprocessing

[0097] 1.1 Experimental data acquisition.

[0098] Covering the typical value range of 12-dimensional process parameters, 300 sets of orthogonal experiments were designed, with each set of experiments corresponding to one component sample, and 9 quality indicators were collected: Surface defects: number of bubble defects, number of gap defects, number of wrinkle defects, number of overlapping defects, number of torsion defects; Mechanical properties: tensile strength (MPa), flexural strength (MPa), interlaminar shear strength (MPa), compressive strength (MPa).

[0099] 1.2 Data Standardization and Enhancement.

[0100] The six continuous variables were normalized using Min-Max to map them to the [-1, 1] interval to improve training stability. The six discrete variables were then embedded using an 8-dimensional learnable embedding (embedding_dim=8), and the embedding vectors were trained end-to-end with the network after initialization. The processed continuous features and discrete embedded features were concatenated to form an input feature vector with a dimension of 54 (6+6×8).

[0101] To address the problem of scarce original samples, a GSMOTE + physical constraint strategy is adopted for sample augmentation and screening. First, the SMOTE algorithm is used to interpolate in the 5-nearest neighbor space of the minority class to generate candidate samples. Then, the Tomek Links method is used to remove boundary noise pairs and retain valid candidate samples. Subsequently, a physical feasibility filtering mechanism is introduced to remove unreasonable samples that satisfy the radius R < R min = 200 mm or the temperature T h < T9 - 20 °C (where T9 = 135 °C is the glass transition temperature of the CF / EP material). Finally, the augmented samples are stratified into a valid training set, a validation set, and a test set in a ratio of 7:2:1.

[0102] Step 2: Inverse mechanism modeling.

[0103] Construct a candidate function library, which contains a total of P = 45 candidate basis functions, covering three types of mechanisms: heat, force, and flow. Taking the number of gaps as an example, its evolution equation is assumed to be: ; Use variational sparse Bayesian learning (VSBL) to solve the sparse coefficient : ; Iteratively update through the EM algorithm, retain the terms of the posterior mean , and finally identify the dominant equation for : ; This equation shows that the gap surges at a small curvature radius (R < 500 mm), which is in line with engineering experience.

[0104] Further use the hidden Markov model (HMM, K = 3) to identify mode switching: Mode 1 (steady laying): R ≥ 1000 mm, v ≤ 250 mm / s; Mode 2 (medium-speed turning): 500 ≤ R < 1000 mm; Mode 3 (sharp turn): R < 500 mm; The switching surface is defined as: ; Each mode corresponds to an independent differential equation system , and differentiable switching is achieved through the smooth maximum function: ; Where Distance to the modal center =10, used to control the steepness of the switch; This is the vector of currently input process parameters; For the first The center point of the physical mode of a typical operating condition identified by the system.

[0105] Step 3: PI-GAN model construction and training.

[0106] Building a Transformer-PINN generator : Input: A 54-dimensional feature concatenation vector, converted into a pseudo-sequence. ; Encoder: Layer, number of heads FFN intermediate dimension ; Decoder: Layer (decoding depth can be reduced due to static prediction); Output: ; This is the output of the hidden state of the last layer of the decoder.

[0107] Discriminator Utilizing a lightweight MLP: The activation function is LeakyReLU (negative slope). ).

[0108] The adversarial training employs a two-timescale update rule (TTUR): Fixed discriminator learning rate generator ; Each round of updates Second-rate ,Every Rotation Update Second-rate ; Batch size Total training rounds Stop early .

[0109] The generator's total loss is designed as follows: ; in: ; in, This is generated fake data; The score given by the discriminator; The expected value is the average of the results calculated for all samples in the current batch.

[0110] ; in, The most standard supervised learning loss is used; the predicted value is calculated. and the true value The mean square error between them; : This represents the loss weight, indicating that among these 9 indicators, some indicators may be more important than others, so the error penalty is heavier.

[0111] ; in, The Jacobian matrix is ​​the derivative of the output with respect to the input. . These are the right-hand side terms of the piecewise smooth physical equations previously extracted through sparse Bayesian learning.

[0112] in, The Jacobia matrix is ​​calculated using automatic differentiation. This is the right-hand item of the switching system output for stage 2.

[0113] Initial weights In verifying the loss Later dynamically adjusted to .

[0114] Step 4: Model evaluation and incremental self-learning.

[0115] When a new scenario is introduced (such as CF / PEEK materials), incremental self-learning is initiated: Knowledge Distillation: The Old Model As a teacher, the new model For the student, the loss is: ; in, It is a matrix representing the relationships of concern between layers within the Transformer. The attention matrix represents the old model. This represents the attention matrix of the new model currently being trained.

[0116] in, , This indicates the total number of attention heads.

[0117] Representative replay: from the old feature library In the process, retrieval is performed on each new sample. Neighboring, Mixed Ratio ;in, No. The encoder outputs a feature matrix for each sample.

[0118] After 100 rounds of fine-tuning, the new scenario was tested. Furthermore, performance degradation in older scenarios This verifies the absence of catastrophic forgetting.

[0119] Implementation effect verification.

[0120] Reference Figure 3 As shown, after optimizing process parameters and implementing closed-loop control using the filament placement machine process performance prediction method described in this embodiment, the actual filament placement effect is significantly improved. From the perspective of performance prediction accuracy, the Physical Information Embedded Generative Adversarial Network (PI-GAN) constructed in this invention achieves high-precision, strong consistency, and low-variance collaborative prediction of nine key quality indicators (five surface defects + four mechanical properties): Surface defect prediction is significantly better than traditional XGBoost and MLP models; the confidence interval width for mechanical property prediction is much narrower than that of empirical models, reflecting the model's strong robustness to process disturbances and fully meeting the reliability requirements of aerospace-grade components.

[0121] Example 3 Based on the same inventive concept, this embodiment provides a physical information neural network intelligent performance prediction system for automatic multi-material filament placement on complex curved surfaces of aircraft. The principle of solving the problem is similar to that of the physical information neural network intelligent performance prediction method for automatic multi-material filament placement on complex curved surfaces of aircraft, and the repeated parts will not be described again.

[0122] This embodiment provides a physical information neural network intelligent performance prediction system for automated multi-material fiber placement on complex curved surfaces of aircraft, characterized by comprising: The data acquisition module is used to acquire fiber placement process parameters; wherein, the fiber placement process parameters include: continuous variables and discrete variables, the continuous variables include placement speed, placement pressure, placement head temperature, fiber tension, turning path curvature radius, ribbon width and substrate temperature; the discrete variables include: material type, process type, placement angle, plate geometry and substrate material; The data preprocessing and enhancement module is used to perform standardized preprocessing and physical constraint enhancement sampling on the wire laying process parameters to obtain an enhanced training dataset. The candidate library construction module is used to construct a high-order nonlinear sparse regression candidate library based on the enhanced training dataset; wherein, the high-order nonlinear sparse regression candidate library includes a heat conduction dominant term characterizing the heat exchange between the filament bundle and the substrate, a mechanical deformation coupling term characterizing compaction, shearing and buckling instability, and a flow spreading dynamics term characterizing the lateral spreading and edge instability of the filament bundle. The inverse mechanism modeling module is used to identify piecewise smooth physical differential equation systems and their mode switching logic based on the high-order nonlinear sparse regression candidate library, using sparse Bayesian regression and hidden Markov models, to obtain an embeddable inverse mechanism model. A network construction module is used to construct a Physical Information Embedded Generative Adversarial Neural Network (PI-GAN). The generator of the PI-GAN adopts a Transformer-PINN architecture, and the discriminator adopts a lightweight multilayer perceptron. The inverse mechanism model is embedded as a physical constraint term in the generator's total loss function. The generator's total loss function includes adversarial loss, data fidelity loss, and physical constraint loss, wherein the physical constraint loss is constructed based on the piecewise smooth physical differential equation system and its mode switching logic. An optimization module is used to collaboratively optimize the generator and the discriminator in an adversarial training manner to obtain a trained model. Based on the trained model, surface defect indices and mechanical performance indices are output. The surface defect indices and mechanical performance indices are evaluated using mean square error as the evaluation index. The surface defect indices include the number of bubbles, gaps, wrinkles, overlaps, and torsional deformations. The mechanical performance indices include tensile strength, flexural strength, interlaminar shear strength, and compressive strength. The incremental self-learning module is used to initiate incremental self-learning when a new process scenario is introduced. It retains the historical model knowledge of the trained model through knowledge distillation and combines a representative sample replay strategy to prevent catastrophic forgetting, thereby achieving model self-evolution.

[0123] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0124] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0125] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0126] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0127] Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A physical information neural network intelligent performance prediction method for automated multi-material fiber placement on complex curved surfaces of aircraft, characterized in that, include: The collected fiber placement process parameters include continuous and discrete variables. The continuous variables include placement speed, placement pressure, placement head temperature, fiber tension, radius of curvature of turning path, ribbon width, and substrate temperature. The discrete variables include material type, process type, placement angle, plate geometry, and substrate material. The fiber placement process parameters are standardized and preprocessed, and physical constraint enhancement sampling is performed to obtain an enhanced training dataset; Based on the enhanced training dataset, a high-order nonlinear sparse regression candidate library is constructed; wherein, the high-order nonlinear sparse regression candidate library includes a heat conduction dominant term characterizing the heat exchange between the filament bundle and the substrate, a mechanical deformation coupling term characterizing compaction, shearing and buckling instability, and a flow spreading dynamics term characterizing the lateral spreading and edge instability of the filament bundle. Based on the aforementioned high-order nonlinear sparse regression candidate library, sparse Bayesian regression and hidden Markov models are used to jointly identify piecewise smooth physical differential equation systems and their mode switching logic, thereby obtaining an embeddable inverse mechanism model. A physical information embedding generative adversarial neural network (PI-GAN) is constructed, wherein the generator of the PI-GAN adopts the Transformer-PINN architecture, and the discriminator adopts a lightweight multilayer perceptron; the inverse mechanism model is embedded as a physical constraint term into the total loss function of the generator; wherein the total loss function of the generator includes adversarial loss, data fidelity loss, and physical constraint loss, and the physical constraint loss is constructed based on the piecewise smooth physical differential equation system and its mode switching logic; The generator and the discriminator are co-optimized using adversarial training to obtain a trained model. Based on the trained model, surface defect indices and mechanical performance indices are output. The mean square error is used as the evaluation index for the surface defect indices and mechanical performance indices. The surface defect indices include the number of bubbles, gaps, wrinkles, overlaps, and torsional deformations. The mechanical performance indices include tensile strength, flexural strength, interlaminar shear strength, and compressive strength. When a new process scenario is introduced, incremental self-learning is initiated. The historical model knowledge of the trained model is retained through knowledge distillation, and a representative sample replay strategy is combined to prevent catastrophic forgetting, thereby achieving model self-evolution.

2. The physical information neural network intelligent performance prediction method for automated multi-material fiber placement on complex curved surfaces of aircraft, as described in claim 1, is characterized in that... The fiber placement process parameters are standardized and preprocessed, and physical constraint-enhanced sampling is performed to obtain an enhanced training dataset, including: The parameter set of the fiber placement process parameters as follows: ； in, Laying speed, mm / s; Laying pressure, N / mm; : Laying head temperature, °C; : Tow tension, N; : Radius of curvature, mm; Ribbon width, mm; ∈{carbon fiber / epoxy, glass fiber / PEEK,…}: Material type; ∈{Automatic Fiber Placement (AFP), Automatic Tape Placement (ATL)}: Process category; ∈{0° ,±45° ,90°}: Paving angle; ∈{flat plate, hyperbolic paraboloid}: plate type; Substrate material; Substrate temperature, °C; : A 12-dimensional real vector space; The continuous variables are normalized using Z-score or Min-Max: ； in, , These are the mean and standard deviation of the training set, respectively, to prevent differences in dimensions from dominating gradient updates; For the normalized first One variable value; For the index of the variable; Embedding encoding of the discrete variables: for categorical variables Construct learnable embeddings: ； ； ； in, An embedding vector is a real-valued vector that transforms a discrete category into a vector to preserve semantic information for neural network learning. For embedding weight matrix; It is a one-hot encoding function; For the embedded dimension; For the number of categories, Preserve semantic similarity; The physical constraint-enhanced sampling includes: For minority class samples Nearest neighbor convex hull linear interpolation generates candidate samples: ； ； in, For newly generated candidate samples; This is the currently selected original sample; Neighboring samples; These are the interpolation coefficients; A uniform distribution between 0 and 1; Tomek Links are used to remove boundary noise pairs, preserving physically consistent samples. A physical feasibility constraint is introduced for filtering; the constraint is as follows: ； and ; Where R is the radius of curvature of the turning path; Temperature of the laying head; Minimum allowable turning radius for the material; The glass transition temperature of the resin matrix; For safety margin.

3. The physical information neural network intelligent performance prediction method for automated multi-material fiber placement on complex curved surfaces of aircraft, as described in claim 1, is characterized in that... Based on the enhanced training dataset, a high-order nonlinear sparse regression candidate library is constructed, including: High-order nonlinear sparse regression candidate library These can be represented as the following three types of basis functions: The dominant heat conduction term includes: Fourier thermal conductivity: ； in, For the Laplace operator; Temperature of the laying head, in °C; The substrate temperature is ℃; Convection heat transfer terms: ； in, The convective heat transfer coefficient; Phase transition latent heat approximation term: ； Where H is the Heaviside function, which characterizes the abrupt change from the glassy state to the elastic state; is the glass transition temperature; m is the polynomial power; The mechanical deformation coupling term includes: Pressure Force – Gap Relationship: ； in, Laying pressure, N / mm; Width of the ribbon, in mm; The radius of curvature of the bend is in mm; Influence of bending stiffness: ； in , representing curvature; For the laying angle; , representing the equivalent bending stiffness; Frictional shear stress: ； in, The coefficient of friction is temperature-dependent. The flow spreading dynamics term includes: Capillary-related terms: ; in, This refers to the apparent viscosity of the resin. Surface tension; The laying speed is in mm / s; for the Arrhenius type, ; The reference viscosity is the theoretical limiting viscosity value at infinitely high temperatures. The activation energy is the minimum energy required for molecules to flow. The molar gas constant; Related terms of Weissenberg number: ； in, The relaxation time characterizes the elastic effect; All candidates passed the dimension consistency check to ensure that each one is dimensionless or has a definite physical dimension. The final high-order nonlinear sparse regression candidate library is as follows: ； Its i-th line is: ； Where N is the number of samples; P≈45, including constant terms, first / second order single terms, and interaction terms; Let P be the dimension of the row vector, and P be the total number of all possible physical terms. Let be the state vector at the i-th position.

4. The physical information neural network intelligent performance prediction method for automated multi-material fiber placement on complex curved surfaces of aircraft, as described in claim 1, is characterized in that... A sparse Bayesian regression and hidden Markov model are jointly used to identify piecewise smooth physical differential equation systems and their mode switching logic, resulting in an embeddable inverse mechanism model, including: Set quality indicators The time / space evolution satisfies the implicit differential equation: ； ； in, This represents 9 quality indicators; , is a sparse coefficient matrix; sparse coefficient matrix The dimensions are P rows corresponding to physical items in the candidate library, and 9 columns corresponding to 9 quality indicators; It is Gaussian noise; For noise variance; It is the identity matrix; This serves as a candidate library for high-order nonlinear sparse regression. Solving SBL using Sparse Bayesian learning: ； in, It is the probability density function; For the observation data matrix; These are hyperparameters that control the relevance of each weight. Among them, the prior is set as a hierarchical automatic correlation determination prior: ； in, These are the weighting coefficients corresponding to the k-th candidate physics item; It is a 9-dimensional identity matrix; For the corresponding Weight parameters; The shape parameter of the Gamma distribution; is the scale parameter of the Gamma distribution; Iterative updates are achieved by maximizing the expected value of EM or by variational inference. With Ξ, automatic model selection and uncertainty quantification are achieved; in, The objects updated iteratively by the algorithm are all value; When process parameters cross critical thresholds, the quality evolution law undergoes discontinuous abrupt changes, requiring the introduction of an event-driven mode switching mechanism. Define the set of switching surfaces Each switching surface is defined by physical critical conditions: ； Where m is the total number of switching surfaces / modes; Let be the state space, representing the mathematical space containing all possible combinations of process parameters; ； ； ； ； ； in, This is the equivalent elastic modulus of the composite material; The thickness of the prepreg tape / filament bundle; Patch length or effective contact length; The basic glass transition temperature; These are the empirical fitting coefficients; For different devices, The values ​​differ: 0.18 for AFP models and 0.32 for ATL models. ; Furthermore, a hidden Markov model is used to analyze the unobserved modes. Identify states ∈{1,2,…,K}, where K is the total number of hidden states: Observation sequence: quality indicators; State transition: ； in, , Let be the hidden state variables at time t and time t-1; This is the state transition matrix, representing the transition from the state at the previous time step. Change to the current state The probability of; Launch distribution: ； in, The observation vector is the actual measured quality index data. An index for a certain state; For the first The covariance matrix under each state; By training using the Baum-Welch algorithm or ELBO maximization, we can obtain: Optimal number of modes ; Mechanism equations corresponding to each mode ; Mapping the switching probability matrix A to the critical surface ; The final output is a piecewise smooth physical system, i.e., the inverse mechanism model: ； in, For the first Residual / noise under these modes; For the first in the state space Sub-regions; Indicates the switching surface of the system; When the system is on the boundary between two states, the system is in a switching state.

5. The physical information neural network intelligent performance prediction method for automated multi-material fiber placement on complex curved surfaces of aircraft, as described in claim 1, is characterized in that... The total loss function is: ； in, and For dynamic weighting coefficients, Number of training iterations; adversarial loss It is expressed as follows: ； in, For discriminator networks; Data fidelity loss It is expressed as follows: ； Among them, weight Dynamically updated based on historical failure analysis; The first of the quality metrics calculated for the generator network One quality indicator; The physical constraint loss is expressed as follows: ； in, ; ； in, For uncertainty weighting matrix, The posterior standard deviation of the output of sparse Bayesian learning.

6. The physical information neural network intelligent performance prediction method for automated multi-material filament placement on complex curved surfaces of aircraft, as described in claim 1, is characterized in that... The discriminator employs WGAN-GP loss to improve training stability and sample diversity; its expression is as follows: ； in: , for Wasserstein distance estimation; Let $\mathematical expectation$ represent the average value obtained over the sampling process. Scoring based on real data; To generate scores for the data; This is a gradient penalty; ； ； , where is the penalty coefficient; The actual value; This is a predicted value; For discriminator networks, it refers to the process of interpolating samples. The output score obtained by inputting into the discriminator; The gradient operator is used to calculate the output of the discriminator. Relative to its input The rate of change; It is an L2 norm.

7. The physical information neural network intelligent performance prediction method for automated multi-material fiber placement on complex curved surfaces of aircraft, as described in claim 1, is characterized in that... The evaluation is conducted using mean squared error as the evaluation metric, including: For the nine-dimensional quality indicators, the following evaluation indicators are constructed: ； ； ； in, For predicted values, The actual value is the label; Representing the There are 9 quality indicators in total. Representing the One sample; The mean squared error should be as small as possible, but it is sensitive to large errors. The smaller the mean absolute error, the better, as it reflects the actual error distance; The coefficient of determination is as close to 1 as possible, indicating a high degree of model fit.

8. The physical information neural network intelligent performance prediction method for automated multi-material fiber placement on complex curved surfaces of aircraft, as described in claim 1, is characterized in that... Knowledge distillation preserves the historical model knowledge of the trained model, including: Define attention-guided distillation loss: ； in, It is softmax; Temperature coefficient; This is the generator output for the old model, i.e., the teacher model; The generator output is for the new model, i.e., the student model; This is the self-attention matrix of the last layer of the encoder; 。 9. The physical information neural network intelligent performance prediction method for automated multi-material fiber placement on complex curved surfaces of aircraft, as described in claim 1, is characterized in that... The representative sample replay employs a feature space nearest neighbor replay strategy, including: A high-dimensional feature library of old task data is built offline. When training a new task, k nearest neighbor old samples are dynamically retrieved for each new sample and mixed in a preset ratio for fine-tuning; where k=5 and the mixing ratio is 1:

9.

10. A physical information neural network intelligent performance prediction system for automated multi-material fiber placement on complex curved surfaces of aircraft, characterized in that, include: The data acquisition module is used to acquire fiber placement process parameters; wherein, the fiber placement process parameters include: continuous variables and discrete variables, the continuous variables include placement speed, placement pressure, placement head temperature, fiber tension, turning path curvature radius, ribbon width and substrate temperature; the discrete variables include: material type, process type, placement angle, plate geometry and substrate material; The data preprocessing and enhancement module is used to perform standardized preprocessing and physical constraint enhancement sampling on the wire laying process parameters to obtain an enhanced training dataset. The candidate library construction module is used to construct a high-order nonlinear sparse regression candidate library based on the enhanced training dataset; wherein, the high-order nonlinear sparse regression candidate library includes a heat conduction dominant term characterizing the heat exchange between the filament bundle and the substrate, a mechanical deformation coupling term characterizing compaction, shearing and buckling instability, and a flow spreading dynamics term characterizing the lateral spreading and edge instability of the filament bundle. The inverse mechanism modeling module is used to identify piecewise smooth physical differential equation systems and their mode switching logic based on the high-order nonlinear sparse regression candidate library, using sparse Bayesian regression and hidden Markov models, to obtain an embeddable inverse mechanism model. A network construction module is used to construct a Physical Information Embedded Generative Adversarial Neural Network (PI-GAN). The generator of the PI-GAN adopts a Transformer-PINN architecture, and the discriminator adopts a lightweight multilayer perceptron. The inverse mechanism model is embedded as a physical constraint term in the generator's total loss function. The generator's total loss function includes adversarial loss, data fidelity loss, and physical constraint loss, wherein the physical constraint loss is constructed based on the piecewise smooth physical differential equation system and its mode switching logic. An optimization module is used to collaboratively optimize the generator and the discriminator in an adversarial training manner to obtain a trained model. Based on the trained model, surface defect indices and mechanical performance indices are output. The surface defect indices and mechanical performance indices are evaluated using mean square error as the evaluation index. The surface defect indices include the number of bubbles, gaps, wrinkles, overlaps, and torsional deformations. The mechanical performance indices include tensile strength, flexural strength, interlaminar shear strength, and compressive strength. The incremental self-learning module is used to initiate incremental self-learning when a new process scenario is introduced. It retains the historical model knowledge of the trained model through knowledge distillation and combines a representative sample replay strategy to prevent catastrophic forgetting, thereby achieving model self-evolution.

Images