Method and system for dynamic compensation of a towpreg head tcp based on internal displacement sensors

Abstract

The application discloses a filament laying head TCP dynamic compensation method and system based on internal displacement sensors, and relates to the technical field of intelligent manufacturing. The method comprises the following steps: obtaining a digital model of a target mold and a filament laying scheme, extracting mold geometric features and determining an ideal trajectory of a tool center point of the filament laying head; establishing a data-driven correction thickness model to determine the correction thickness by taking the mold geometric features and the current number of laid layers as inputs, and superimposing the correction thickness on the ideal trajectory to generate a corrected trajectory; combining the deviations of the mold geometric features, the ideal trajectory and the corrected trajectory, and determining the dynamic compensation air pressure through a compensation air pressure decision model based on Gaussian process regression; controlling the bonding pressure according to the dynamic compensation air pressure, and collecting measured thickness data through an internal displacement sensor of a pressure roller to perform closed-loop updating on the correction thickness model. The application realizes accurate control of the interlayer thickness and online adaptive adjustment of the air pressure in the complex curved surface filament laying process, and improves the laying quality of the composite material component.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, specifically to a TCP dynamic compensation method and system for filament placement heads based on internal displacement sensors. Background Technology

[0002] Automated fiber placement technology is a core process in the manufacturing of aerospace composite components, widely used in the placement and molding of large curved surface structures such as aircraft fuselages and wing skins. The trajectory accuracy of the tool center point (TCP) of the fiber placement head directly affects the placement position and compaction effect of the fiber bundle on the mold surface. As aerospace structural components develop towards larger and more integrated designs, the geometric complexity of the mold surface continues to increase, and features such as hyperbolic surfaces and variable curvature transition zones pose challenges to the quality of fiber placement.

[0003] Existing technologies have the following main shortcomings: First, traditional methods use constant air pressure to drive the pressure roller, applying uniform pressure to all areas of the die surface. This fails to consider the influence of geometric factors such as local curvature and normal angle, leading to fiber accumulation in concave areas and thinner thickness in convex areas, resulting in significant interlayer thickness deviations. Second, existing sensors are mostly installed outside the pressure roller, indirectly inferring layup quality through force signals. Spatially offset between the measurement point and the pressing point prevents in-situ real-time measurement. Third, current process optimization lacks a quantitative correction model that couples multiple physical field factors such as curvature, layup number, and temperature field, resulting in long adaptation cycles and high costs across dies and material systems. Therefore, there is an urgent need for a dynamic compensation method for the center point of the filament layup tool that can achieve in-situ layer thickness measurement and integrate a multi-physics field coupling model. Summary of the Invention

[0004] This invention addresses the technical problems in existing technologies, such as poor thickness consistency between layers on complex curved surfaces due to constant pressure control, the inability of external sensors to achieve in-situ real-time measurement, and the lack of a multi-physics field coupling quantitative correction model leading to long adaptation cycles across working conditions. It provides a TCP dynamic compensation method and system for wire placement heads based on an internal displacement sensor.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a TCP dynamic compensation method for a wire placement head based on an internal displacement sensor, comprising: Obtain the digital model and filament placement scheme of the target mold, extract the geometric features of the mold, and determine the ideal trajectory of the center point of the filament placement head tool accordingly; A data-driven correction thickness model is established. Using the mold geometry and the current layup number as input, the correction thickness corresponding to each position on the mold surface is determined and superimposed on the ideal trajectory to generate the correction trajectory. Combining the mold geometry and the deviation between the ideal trajectory and the corrected trajectory, the dynamic compensation air pressure of the filament placement head is determined using a compensation air pressure decision model based on Gaussian process regression. The bonding pressure of the filament placement head at the current number of lay-up layers is controlled by the dynamic compensation air pressure, and the measured thickness data is collected by the displacement sensor arranged inside the pressure roller of the filament placement head to perform closed-loop update of the corrected thickness model.

[0006] Secondly, the present invention provides a TCP dynamic compensation system for a wire placement head based on an internal displacement sensor, comprising: The trajectory acquisition module is used to acquire the digital model and filament placement scheme of the target mold, extract the geometric features of the mold, and determine the ideal trajectory of the center point of the filament placement head tool accordingly. The trajectory correction module is used to establish a data-driven correction thickness model. Taking the mold geometry features and the current layup number as input, it determines the correction thickness corresponding to each position on the mold surface and superimposes it onto the ideal trajectory to generate the correction trajectory. The air pressure decision module is used to determine the dynamic compensation air pressure of the filament placement head by combining the geometric features of the mold, the deviation between the ideal trajectory and the corrected trajectory, and through a compensation air pressure decision model based on Gaussian process regression. The closed-loop update module is used to control the bonding pressure of the filament placement head at the current number of layers according to the dynamic compensation air pressure, and to collect measured thickness data through displacement sensors arranged inside the pressure roller of the filament placement head, and to perform closed-loop update of the corrected thickness model.

[0007] The beneficial effects of this invention are: Compared to existing technologies, this invention first extracts geometric features such as curvature and normal angle from the digital model of the mold to generate an ideal trajectory for the filament placement head, providing a geometric benchmark for subsequent compensation. Secondly, a multi-physics coupled correction thickness model is constructed, incorporating the stacking effect caused by curvature, cumulative layer deviation, the influence of temperature and curing degree on resin flowability, and residual stress rebound into quantitative calculations to generate a correction trajectory. Thirdly, a Gaussian process regression model with embedded physical information is used to determine the dynamic compensation air pressure, employing Hertzian contact theory as a mean function prior. This allows for accurate prediction of the optimal air pressure with limited calibration data, and outputs uncertainty estimates to guide safe compensation. Finally, measured thickness data is collected in situ using a displacement sensor array embedded within the pressure roller. This data is then combined with online morphology reconstruction technology to continuously update the correction thickness model, and a transfer learning framework is used to achieve rapid adaptation across molds and materials. This invention improves the thickness consistency between layers on complex curved surfaces, suppresses bridging defects in concave regions, and shortens the process calibration cycle for new molds and materials. Attached Figure Description

[0008] Figure 1A flowchart illustrating the TCP dynamic compensation method for the filament placement head based on an internal displacement sensor provided by this invention; Figure 2 This is a schematic diagram of the TCP dynamic compensation system for the filament placement head based on an internal displacement sensor provided by the present invention.

[0009] In the attached diagram, the components represented by each number are as follows: The system includes a trajectory acquisition module 11, a trajectory correction module 12, a barometric pressure decision module 13, and a closed-loop update module 14. Detailed Implementation

[0010] Example 1, as Figure 1 As shown, this embodiment of the invention provides a TCP dynamic compensation method for wire placement heads based on internal displacement sensors, including: S10: Obtain the digital model and filament placement scheme of the target mold, extract the geometric features of the mold, and determine the ideal trajectory of the center point of the filament placement head tool accordingly.

[0011] In automated fiber placement, the fiber placement head needs to lay fiber bundles layer by layer on the die surface according to a preset path. The reference point for its motion control is called the tool center point. Specifically, the tool center point is the geometric center of the actual contact area between the pressure roller (compacting roller) and the die surface. The TCP is the programming reference point for CNC motion. The trajectory accuracy of the tool center point directly affects the actual placement position and compaction effect of the fiber bundle on the die surface. However, different dies have different surface shapes; some areas have large curvature, some have small curvature, some are concave, and some are convex. Different geometric features will significantly affect the compaction behavior of the fiber bundle. Therefore, this step first obtains the digital model of the target die and the fiber placement scheme, extracts geometric features such as curvature and normal angle, and determines the ideal trajectory without considering thickness deviation.

[0012] The digital model of the target mold is typically a CAD model generated by computer-aided design software and stored in STEP or IGES format. It contains precise geometric information about the mold surface, such as its shape, dimensions, and curvature distribution. The fiber placement scheme is a placement process document developed based on the design requirements of the composite component. It specifies the placement sequence, angle, path, and boundaries of each fiber bundle. The digital model and the fiber placement scheme together form the foundational data source for subsequent trajectory planning and compensation decisions. By extracting prior information, a geometric benchmark can be established for subsequent trajectory correction and pressure compensation, enabling compensation decisions to anticipate geometric changes in the surface and avoid response lag caused by relying solely on real-time measurements. Specifically, the purpose of this step is to introduce prior geometric knowledge of the mold into the compensation framework, achieving a shift from passive response to proactive prediction.

[0013] Specifically, the process involves acquiring the digital model of the target mold and the filament placement scheme, extracting the mold's geometric features, and determining the ideal trajectory of the filament placement head tool's center point, including: Based on the digital model of the target mold, extract the mold geometric features at each position on the mold surface. The mold geometric features include at least the local curvature and normal angle of the surface. Based on the fiber placement scheme, extract the layup design parameters; By combining the digital model, the geometric features of the mold, and the layup design parameters, an ideal trajectory of the center point of the filament placement tool in each layup is generated, and the layup number is marked for each ideal trajectory.

[0014] First, based on the digital model of the target mold, the geometric features of the mold at various locations on the mold surface are extracted. These geometric features include at least the local curvature and normal angle of the surface. Specifically, the digital model is a three-dimensional geometric representation of the target mold constructed in computer-aided design software. It is obtained by acquiring STEP or IGES format CAD files from the design department. This digital model is usually stored in the form of a NURBS surface. The curvature and normal information of each point can be obtained by calculating the partial derivatives of the surface in the parameter domain.

[0015] Local curvature includes principal curvature, Gaussian curvature, and mean curvature: Principal curvature describes the degree of bending of the surface in two mutually perpendicular principal directions; Gaussian curvature is used to identify the type of surface, for example, when the Gaussian curvature is greater than zero and the mean curvature is less than zero, it is a concave elliptic point; when the Gaussian curvature is greater than zero and the mean curvature is greater than zero, it is a convex elliptic point; when the Gaussian curvature is less than zero, it is a hyperbolic point, i.e., a saddle-shaped region; mean curvature is the average value of the principal curvatures and is used to quantitatively describe the overall bending strength of the surface.

[0016] The normal angle is the angle between the surface normal vector and a reference direction, which can be selected as the direction of gravity or the main axis of the filament placement head. The magnitude of the normal angle directly affects the attitude adjustment range of the filament placement head during the placement process. The aforementioned mold geometry features will be stored as a two-dimensional array associated with the surface parameter coordinates, serving as input for subsequent thickness model correction.

[0017] Secondly, based on the fiber placement scheme, layup design parameters are extracted. The fiber placement scheme includes information such as layup sequence, layup angle, layup path, fiber bundle width, and design thickness of each layer. Among these, the layup sequence refers to the order in which the layers are arranged in the thickness direction, determining the stacking method of the fiber bundles from the bottom to the top; the layup angle refers to the angle between the fiber orientation of each fiber bundle and the reference direction, and different layup combinations affect the mechanical properties and thickness accumulation law of the laminate; the layup path refers to the trajectory of the fiber placement head moving on the mold surface, determining the coverage area and layup direction of the fiber bundles; the fiber bundle width refers to the width of a single prepreg fiber bundle, affecting the coverage efficiency and the overlap method between fiber bundles during layup; and the design thickness of each layer refers to the theoretical thickness of each fiber bundle after layup, serving as the benchmark reference value for thickness deviation analysis. The layup design parameters are used to determine the design indicators of the current layup area under a specific number of layups, providing a benchmark for subsequent thickness deviation analysis.

[0018] Furthermore, by combining the digital model, mold geometry, and layup design parameters, the ideal trajectory of the tool center point in each layup is generated, and the layup number is marked for each ideal trajectory. Specifically, the ideal trajectory generation process is as follows: based on the current layup path planning, a series of discrete path points are determined on the parameter domain of the mold surface, each path point corresponding to a set of parameter coordinate values. For each path point, its position coordinates in three-dimensional space are calculated according to the parametric equation of the mold surface, and then the cumulative design thickness of the current layup is offset outward along the surface normal direction of that point to obtain the target spatial position of the tool center point. The tool center points of all path points are connected in the layup order to form the ideal trajectory of that layup. The layup number corresponding to each ideal trajectory is marked as the index basis for subsequent layer-by-layer compensation. For example, the ideal trajectory of the first layer is marked with a layup number of one, the ideal trajectory of the second layer is marked with a layup number of two, and so on.

[0019] This step yields the ideal spatial path, or ideal trajectory, that the tool center point should traverse in each layup. This ideal trajectory represents the reference path along the die surface where the tool center point should move under ideal conditions, assuming the die surface is an ideal rigid surface, the layup material is ideally uniformly distributed, and thickness deviation is not considered. It provides a geometric reference and deviation reference for the subsequent generation of correction trajectories.

[0020] S20: Establish a data-driven correction thickness model, using the mold geometry and the current layup number as input, determine the correction thickness corresponding to each position on the mold surface, and superimpose it onto the ideal trajectory to generate the correction trajectory; Secondly, after obtaining the ideal trajectory, it's crucial to recognize that in actual automated fiber placement, the actual layer thickness will deviate from the design thickness due to factors such as changes in the curvature of the die surface, the cumulative effect of the number of layers, and the thermochemical properties of the material. For example, in concave curved areas, the fiber bundles accumulate on the inner side of the concave surface, resulting in a localized actual thickness greater than the design thickness; in convex curved areas, the fiber bundles are stretched and dispersed, potentially leading to a localized thinner thickness; and as the number of layers increases, the deviation accumulates non-linearly. Therefore, placement cannot be simply performed according to the ideal trajectory; it requires pre-correction of the trajectory.

[0021] This step introduces a correction thickness model, which takes the mold geometry and the current number of layups as input and outputs the required correction thickness at each position on the mold surface. This correction thickness characterizes the expected deviation between the actual layup thickness and the design thickness. Its value is determined based on a multiphysics coupled analytical model and a data-driven method, taking into account various factors such as geometric deviation caused by curvature, cumulative layer deviation, temperature effect, curing degree effect, and residual stress springback.

[0022] Finally, the corrected thickness is superimposed onto the ideal trajectory along the normal direction of the die surface, thus obtaining the corrected trajectory. The corrected trajectory allows the tool center point to compensate for anticipated deviations during actual placement. For example, in concave areas, the tool center point is shifted outward to accommodate accumulated fibers, and in convex areas, it is shifted inward to compensate for stretching and thinning. This step enables an upgrade from passive response compensation to active predictive compensation, allowing the fiber placement head to adapt to surface geometry and layup accumulation effects.

[0023] First, a data-driven correction thickness model is established, including: Acquire process calibration experimental data, and combine the constructed input feature vector to perform feature encoding on the process calibration experimental data to form a sample dataset. The input feature vector includes geometric features, thermochemical features, and mechanical features. A regression function model of multiphysics coupling is constructed, which includes geometric components, thermochemical components and mechanical components; The regression function model is iteratively trained and evaluated based on the sample dataset, and the trained modified thickness model is output.

[0024] Specifically, firstly, process calibration experimental data is acquired, and then feature-encoded using a constructed input feature vector to form a sample dataset. The process calibration experimental data consists of real data collected during preliminary layup experiments, including actual corrected thickness values ​​under different curvature regions, different layup numbers, and different temperatures and degrees of cure. The input feature vector includes at least geometric, thermochemical, and mechanical features. Geometric features include the local curvature, normal angle, and current layup number of the mold surface, used to describe the influence of surface geometry on layup thickness. Thermochemical features include real-time temperature and resin cure degree during layup, used to describe the influence of material flowability changes with temperature and cure process on compaction behavior. Mechanical features include the residual stress state of previous layups and in-plane shear strain of fibers, used to describe the constraints of existing layups on the bonding behavior of subsequent layups. Through feature encoding, the original experimental data can be converted into a structured, dimensionlessly standardized sample dataset.

[0025] Feature encoding is a preprocessing step that transforms raw experimental data into a structured, dimensionlessly uniform sample dataset. In data-driven modeling, raw process calibration experimental data typically originates from different sampling channels and possesses different physical dimensions, orders of magnitude, and data types. For example, curvature is measured in millimeters, temperature in degrees Celsius, and the number of layers is a dimensionless integer. Directly inputting raw data into the model can lead to features with larger dimensions or wider numerical ranges dominating model training, while features with smaller dimensions or narrower numerical ranges are suppressed. Commonly used feature encoding methods in existing technologies include one-hot encoding, label encoding, normalized scaling encoding, and physical feature construction based on prior knowledge. One-hot encoding is suitable for converting discrete categorical variables into binary vectors; label encoding maps discrete categories to integer indices; normalized scaling encoding maps continuous features to the same numerical range through linear transformation; and physical feature construction combines raw measurements into dimensionless features with clear physical meaning based on domain knowledge.

[0026] Optionally, in this embodiment, the feature encoding is implemented by normalizing and scaling geometric features, thermochemical features, and mechanical features respectively, and constructing dimensionless features for some features. For local curvature and normal angle in geometric features, the minimum-maximum normalization method can be directly used, that is, subtracting the minimum value of this type of feature in the sample dataset from the original feature value, and then dividing by the difference between the maximum and minimum values ​​to map the numerical range to the interval of 0 to 1. For the number of layers, since it is a dimensionless integer, no normalization is performed, and it is directly used as the feature input. For temperature in thermochemical features, the same normalization method as for geometric features is used to map it to the interval of 0 to 1. For degree of curing, its value range is naturally 0 to 1, and no normalization is required. For residual stress in mechanical features, it is first divided by the equivalent modulus of the layers for dimensionless processing to obtain dimensionless residual strain features, and then normalized; for shear strain, it is a dimensionless quantity and is directly normalized. Through the feature encoding described above, all input features are unified to the same numerical range and dimensional system, eliminating the impact of dimensional differences on model training. Simultaneously, the relative magnitudes of each feature are maintained, enabling the regression model to fairly learn the contribution of each feature to the correction thickness. The structured sample dataset after encoding contains the input feature vectors and their corresponding correction thickness labels, serving as the standardized data foundation for subsequent regression model training.

[0027] Secondly, a multi-physics coupled regression function model is constructed, comprising geometric, thermochemical, and mechanical components. Specifically, the geometric components describe the independent effects of surface curvature and normal angle on layer thickness and their cumulative effect with the number of layups. This can be further decomposed into a thickness deviation term caused by curvature and a cumulative deviation term based on the number of layups. The curvature deviation term reflects the thickness change caused by fiber stacking on concave surfaces or stretching on convex surfaces, while the cumulative deviation term reflects the nonlinear amplification of deviation with increasing layup number. The thermochemical components describe the effects of temperature and curing degree fields on resin flowability, including temperature and curing degree influence factors. Increased temperature reduces resin viscosity, which is beneficial for compaction, while increased curing degree reduces resin flowability, requiring greater pressure compensation. The mechanical components describe the influence of residual stress rebound effect and shear deformation of already laid layers on the bonding behavior of subsequent layups.

[0028] Specifically, a multi-physics coupling regression function model is constructed, which includes geometric components, thermochemical components, and mechanical components, including: Construct geometric components, which include a thickness deviation term caused by curvature and a cumulative layer deviation term. The thickness deviation term is determined by regression analysis based on the local radius of curvature, the design thickness, and the current number of layers. The cumulative layer deviation term is determined by combining the average curvature and the current number of layers with a power law relationship. Thermochemical components are constructed, including a temperature influence factor and a cure degree influence factor. The temperature influence factor is determined based on the prior relationship between resin viscosity and temperature and is calibrated by offline experiments. The cure degree influence factor is determined based on the prior relationship of the resin curing kinetic model and is calibrated by offline experiments. Mechanical components are constructed, including a residual stress springback term and a shear deformation term. The residual stress springback term is determined based on the ratio of the residual stress of the preceding ply to the ply's equivalent modulus, combined with offline calibration. The shear deformation term is determined based on the shear strain within the fiber plane calibrated offline.

[0029] First, geometric components are constructed, including a thickness deviation term caused by curvature and a cumulative layer deviation term. The thickness deviation term caused by curvature is determined through regression analysis based on the local radius of curvature, the designed thickness, and the current number of layers. Its physical meaning is that when a layer of material with a thickness of h is laid on a concave surface with a radius of curvature R, the area difference between the curved surface and the flat surface leads to an increase in the areal density of the fibers at the concave surface, manifested as an increase in thickness. This deviation term is proportional to the square of the designed thickness, inversely proportional to the local radius of curvature, and accumulates linearly with the number of layers.

[0030] The cumulative deviation term is determined based on the average curvature and the current number of layers, combined with a power-law relationship. As the number of layers increases, the deviation of the preceding layer changes the local radius of curvature of the subsequent layer, causing the thickness deviation to accumulate superlinearly rather than simply linearly. Therefore, a power-law relationship is used to describe this nonlinear amplification effect.

[0031] Second, thermochemical components were constructed, including a temperature influence factor and a cure degree influence factor. The temperature influence factor was determined based on the prior relationship between resin viscosity and temperature, combined with offline experimental calibration. Resin viscosity decreases exponentially with increasing temperature. This decrease in viscosity makes the fiber layer easier to compact under pressure, thus requiring less compensation pressure in high-temperature laying environments. The cure degree influence factor was determined based on the prior relationship of the resin curing kinetic model and calibrated offline. As cure degree increases, the resin gradually crosslinks, its fluidity decreases sharply, and the compaction window narrows, requiring greater pressure to compensate. This cure degree influence factor increases sharply as the cure degree approaches the gel point. The relationship between cure degree and temperature was pre-calibrated using differential scanning calorimetry experiments, and the cure degree influence factor was calculated based on the ratio of cure degree to gel point.

[0032] Third, the mechanical components are constructed, including a residual stress springback term and a shear deformation term. The residual stress springback term is determined based on the ratio of the residual stress of the preceding layup to the equivalent modulus of the layup, combined with offline calibration. During the layup process, the residual stress accumulated in the preceding layup drives the laminate to spring back after the pressure rollers leave. The springback direction in the concave region is away from the die surface, resulting in an increase in the interlayer gap. This springback effect is directly proportional to the magnitude of the residual stress and inversely proportional to the equivalent modulus of the layup. The smaller the radius of curvature and the greater the residual stress, the more significant the springback. The shear deformation term is determined based on the in-plane shear strain of the fiber calibrated offline. When laid on a hyperboloid, the fiber is subjected to in-plane shear deformation, and the cross-sectional shape of the fiber bundle changes from rectangular to parallelogram, resulting in changes in local thickness. The greater the shear strain, the more severe the cross-sectional distortion.

[0033] The above components are weighted and summed to form a complete regression function model. During the weighted summation, the weight coefficients are set based on the contribution of each component to the actual corrected thickness and the dominance of various physical effects in the process. For example, the weight of the geometric component can be set to 0.5, the weight of the thermochemical component to 0.3, and the weight of the mechanical component to 0.2. For molds with drastic curvature changes, the weight of the geometric component can be appropriately increased; for high-temperature laying or rapid curing systems, the weight of the thermochemical component can be appropriately increased; and for thick plywood or high residual stress systems, the weight of the mechanical component can be appropriately increased.

[0034] In summary, the final regression function model integrates geometric, thermochemical, and mechanical effects into the modified thickness calculation framework, providing a physically interpretable quantitative basis for subsequent modified trajectory generation and pressure compensation decisions.

[0035] Furthermore, the regression function model is iteratively trained and evaluated based on the sample dataset, and the trained modified thickness model is output.

[0036] For example, the thickness correction model adopts a multi-physics coupled regression function structure. The model input is an input feature vector containing geometric, thermochemical, and mechanical features, and the output is the corrected thickness at the corresponding position on the mold surface. Internally, the model consists of a weighted sum of four components: geometric components, thermochemical components, mechanical components, and coupling cross-terms. The geometric components are further decomposed into a thickness deviation term caused by curvature and a cumulative layer deviation term. The curvature deviation term is determined by multiplying the ratio of the square of the design thickness to the local radius of curvature by the number of layers and then by the normal angle correction factor. The cumulative layer deviation term is determined by a power-law function of the average curvature multiplied by the number of layers. The thermochemical components include a temperature influence factor and a cure degree influence factor, determined based on the resin viscosity-temperature relationship and the curing kinetics model, respectively. The mechanical components include a residual stress rebound term and a shear deformation term. The weighting coefficients of each component, as well as the material empirical coefficients, cumulative exponents, fitting parameters, material constants, and coupling coefficients in the model, are determined through regression analysis of process calibration experimental data.

[0037] Specifically, during training, the input feature vectors from the sample dataset are used as model input, and the corresponding experimentally measured corrected thickness is used as the supervision target. An optimization algorithm iteratively adjusts the undetermined parameters in the model. After each iteration, the model's prediction accuracy is evaluated using a validation set. When the model's prediction error on the validation set converges to below a preset threshold, training stops, and the trained corrected thickness model is output. The prediction error convergence threshold can be set according to the process requirements for corrected thickness accuracy; for example, it can be set to a root mean square error between the predicted and measured corrected thicknesses of less than 0.005 mm. By combining data-driven and multiphysics coupling, the model possesses both physical interpretability and the ability to learn complex patterns in actual processes from experimental data, compensating for the shortcomings of purely analytical models in characterizing actual working conditions.

[0038] Further, using the mold geometry and the current layup number as input, the corrected thickness corresponding to each position on the mold surface is determined and superimposed onto the ideal trajectory to generate the corrected trajectory, including: Based on the mold geometry and the current number of layups, and combined with the thermochemical and mechanical features of the current material extracted from the wire laying scheme, a real-time input vector is constructed. The corrected thickness model receives the real-time input vector and outputs the corrected thickness at multiple sampling points on the mold surface. Based on the corrected thickness of multiple sampling points, a continuous thickness correction field on the mold surface is determined by interpolation fitting; Combining physical boundary constraints, the thickness correction field is superimposed onto the ideal trajectory along the normal direction of the mold surface to generate the correction trajectory; The corrected thickness is a vector with a direction mark along the normal direction of the mold surface, and the spacing of the sampling points is determined based on the sampling accuracy of the displacement sensor.

[0039] First, based on the die geometry and the current layup number, and combined with the thermochemical and mechanical features of the current material extracted from the fiber placement scheme, a real-time input vector is constructed. Specifically, the real-time input vector is the set of input features required for the online operation of the thickness model, and its feature structure is consistent with the input feature vector used in the training phase. This real-time input vector includes information such as local curvature, normal angle, current layup number, real-time temperature, degree of cure, residual stress of previous layups, and fiber shear strain at the current layup position. Among them, local curvature and normal angle are extracted in real time from the die digital model, the current layup number is a known process parameter, real-time temperature and degree of cure are acquired online by an infrared thermal imager and dielectric sensor on the fiber placement head, and residual stress of previous layups and fiber shear strain are estimated through offline calibration or online simulation models.

[0040] Subsequently, the thickness correction model receives real-time input vectors and outputs the corrected thickness at multiple sampling points on the mold surface. Sampling points are discrete locations selected at certain intervals along the current layup path. The interval refers to the distance between two adjacent sampling points along the layup path. The spacing of the sampling points is determined based on the sampling accuracy of the displacement sensor. The higher the sensor's sampling accuracy, the smaller the sampling point spacing can be, resulting in a higher resolution of the thickness correction field. For example, when the sensor's sampling accuracy is 0.01 mm, the sampling point spacing can be set to 5 mm. The thickness correction model independently calculates the corrected thickness value at each sampling point based on the real-time input vector. This corrected thickness is a vector with a direction mark along the normal direction of the mold surface. A positive value indicates an outward offset, requiring an increase in the pressing force, while a negative value indicates an inward offset, requiring a decrease in the pressing force.

[0041] Furthermore, based on the corrected thickness at multiple sampling points, a continuous thickness correction field on the mold surface is determined through interpolation fitting. Since the regions between sampling points only have discrete corrected thickness values, an interpolation algorithm is needed to estimate the corrected thickness at intermediate positions to form a continuous thickness correction field. Optionally, radial basis function interpolation or kriging interpolation can be used, with the parameter coordinates of the sampling points and the corrected thickness values ​​as known points, to construct a continuous function covering the entire laying area.

[0042] For example, taking radial basis function interpolation as an example, the specific implementation is as follows: First, the coordinates of all sampling points on the parameter domain of the mold surface are marked as a point set. Each sampling point corresponds to a parameter coordinate pair and a correction thickness value at that point. Then, a radial basis function is selected, commonly in the form of a thin plate spline function. This function uses the distance between the sampling point and the point to be interpolated as the independent variable, with a larger weight for closer distances. For any position to be interpolated on the laying path, the Euclidean distance between that position and each sampling point in the parameter domain is calculated. Each distance value is substituted into the radial basis function to obtain the basis function value. Then, each basis function value is multiplied by the correction thickness weight coefficient of the corresponding sampling point and summed to obtain the correction thickness at the position to be interpolated. The weight coefficients are obtained by solving a system of linear equations, which ensures that the interpolation result at all sampling points is equal to the known correction thickness value. The advantage of radial basis function interpolation is that it can smoothly fit surfaces of arbitrary shapes and does not require a pre-set mesh structure, making it suitable for mold surfaces with complex shapes.

[0043] By using interpolation fitting, the corrected thickness values ​​of discrete sampling points are expanded into a continuous thickness correction field covering the entire laying area. Each point in this continuous thickness correction field corresponds to a corrected thickness value. Subsequent steps can superimpose these corrected thickness values ​​along the normal direction onto each point on the ideal trajectory to generate a continuous corrected trajectory. During interpolation, the spacing between sampling points determines the resolution of the correction field; a smaller spacing results in higher resolution but also higher computational cost, while a larger spacing increases computational efficiency but may lose local details. The sampling point spacing can be adaptively adjusted based on the sampling accuracy of the displacement sensor and the degree of curvature change of the surface. Sampling points are increased in areas with drastic curvature changes and decreased appropriately in areas with gentle curvature changes to balance accuracy and efficiency.

[0044] Finally, combining physical boundary constraints, the thickness correction field is superimposed onto the ideal trajectory along the normal direction of the die surface to generate the correction trajectory. Physical boundary constraints include equipment limit constraints and process safety constraints. For example, the correction trajectory cannot cause the tool center point to exceed the mechanical stroke range of the yarn placement head, cannot cause hard interference between the pressure roller and the die, and the rate of change of the correction amount cannot exceed the dynamic response capability of the yarn placement head. The correction value of each point in the thickness correction field is superimposed onto the ideal trajectory along the normal direction of that point to obtain the correction trajectory point for that point. All correction trajectory points are connected according to the placement sequence to form a complete correction trajectory. Through this step, the theoretical ideal trajectory is corrected into a correction trajectory that can compensate for actual thickness deviations, enabling the yarn placement head to actively adapt to the surface geometry and layup accumulation effect during actual placement.

[0045] S30: Combining the geometric features of the mold and the deviation between the ideal trajectory and the corrected trajectory, the dynamic compensation air pressure of the filament placement head is determined by a compensation air pressure decision model based on Gaussian process regression. Furthermore, after obtaining the corrected trajectory, it is necessary to determine the dynamic compensation air pressure that should be applied by the placement head pressure roller during the actual placement process. The corrected trajectory solves the spatial positioning problem of the center point of the placement head tool. However, even if the center point of the tool moves along the corrected trajectory, an appropriate bonding pressure is still required to compact the fiber bundle to the target thickness. The compaction requirements vary at different spatial locations, so it is necessary to dynamically determine the air pressure value at each moment based on the current placement conditions.

[0046] This step inputs information such as the die geometry and the deviation between the ideal trajectory and the corrected trajectory into a compensation air pressure decision model based on Gaussian process regression. Specifically, the deviation between the ideal trajectory and the corrected trajectory reflects the expected thickness compensation requirement. A larger deviation indicates a more severe thickness deviation at that location due to curvature or layer accumulation, requiring greater air pressure to compact the fibers. Die geometry includes local curvature and normal angle. The greater the curvature or the further the normal angle deviates from the vertical direction, the stronger the uneven constraint on the fibers during compaction, and the air pressure needs to be adjusted accordingly.

[0047] Gaussian process regression is a nonparametric probabilistic model whose output is a Gaussian distribution rather than a single numerical value, including a predicted mean and a predicted variance. The predicted mean serves as the recommended air pressure value, while the predicted variance represents the model's confidence in the prediction. When the predicted variance of the model output is large, it indicates a significant difference between the current input conditions and the training data distribution, suggesting the model is in an extrapolation or curvature abrupt change region. In this case, a robust compensation strategy should be adopted, adding a certain multiple of the standard deviation to the predicted mean as the actual applied compensation air pressure. This ensures adequate fit even with high model uncertainty and triggers an alarm to alert operators to monitor the laying quality in that area.

[0048] This step enables pressure compensation decisions based on a combination of physical priors and data-driven approaches, ensuring both the interpretability of the model and the ability to learn complex process rules from experiments.

[0049] Specifically, the construction of the compensation pressure decision model based on Gaussian process regression includes: By using Hertzian contact theory as a physical prior and embedding it into the mean function of a Gaussian process regression model, a Gaussian process regression model with embedded physical information is constructed. A training dataset is obtained through process calibration experiments, and the Gaussian process regression model is trained based on the training dataset, wherein each data point in the training dataset includes the experimentally verified optimal air pressure value under given feature conditions. Based on the prediction variance output by the trained Gaussian process regression model, and combined with the distribution characteristics of the training data, an uncertainty safety factor is determined, which is used to compensate for the uncertainty of the prediction. The characteristic conditions include at least process characteristics, thermochemical characteristics, geometric characteristics, and deviation evaluation characteristics.

[0050] First, Hertzian contact theory is used as a physical prior and embedded into the mean function of a Gaussian process regression model to construct a Gaussian process regression model with embedded physical information. Gaussian process regression is a nonparametric Bayesian regression method whose output is a Gaussian distribution, including the predicted mean and predicted variance. The mean function of the model represents the prior expectation of the output when there is no observed data. In this step, Hertzian contact theory is used as the mathematical form of the mean function. Specifically, Hertzian contact theory describes the nonlinear relationship between contact pressure and indentation depth when an elastic sphere contacts an elastic plane, i.e., the pressure is proportional to the cube of the indentation depth.

[0051] In automated yarn placement scenarios, the pressure roller is analogous to an elastic sphere, the layup surface to an elastic plane, and the pressing depth is the pressure roller compression amount corresponding to the expected corrected thickness. By embedding this physical relationship into the mean function, even in regions with limited training data, the model can provide reasonable prior predictions based on physical laws, avoiding unreasonable outputs from purely data-driven models in extrapolation regions.

[0052] Secondly, a training dataset was obtained through process calibration experiments, and a Gaussian process regression model was trained based on the training dataset. Each data point in the training dataset includes the experimentally validated optimal gas pressure value under given characteristic conditions. The characteristic conditions include at least process-related features, thermochemical features, geometric features, and deviation evaluation features. Specifically, process-related features include the current number of layers and the layup speed; thermochemical features include temperature influence factors and curing degree influence factors; geometric features include local curvature and normal angle; deviation evaluation features include thickness correction ratio and trajectory deviation ratio, where the thickness correction ratio is the ratio of the corrected thickness to the design thickness, and the trajectory deviation ratio is the ratio of the spatial deviation between the target trajectory and the corrected trajectory to the reference displacement. The process calibration experiments were conducted before or during layup, collecting layup samples under different characteristic conditions. By gradually adjusting the gas pressure and measuring the actual layer thickness, the optimal gas pressure value that makes the layer thickness closest to the design thickness was determined.

[0053] Furthermore, based on the prediction variance output of the trained Gaussian process regression model and combined with the distribution characteristics of the training data, an uncertainty safety factor is determined. The prediction variance of the model output reflects the degree of uncertainty in the prediction under the current input conditions; a larger variance indicates that the model is less confident in the prediction. The distribution characteristics of the training data include the density distribution of each feature in the feature space. In areas with dense training data, the model uncertainty is lower, while in sparse training data in marginal or extrapolation regions, the model uncertainty is higher. The uncertainty safety factor is typically between 1.5 and 2.0. The safety factor is multiplied by the prediction standard deviation and then superimposed on the prediction mean to serve as the actual applied compensation pressure. In extrapolation regions or regions with abrupt curvature changes, a larger safety factor is used to ensure sufficient pressure for adequate fiber adhesion; in areas with sufficient training data, a smaller safety factor is used to avoid overpressure. This uncertainty safety factor allows for a conservative strategy when the model confidence is low, improving the robustness and reliability of the process.

[0054] Furthermore, combining the mold geometry and the deviation between the ideal trajectory and the corrected trajectory, the dynamic compensation air pressure of the fiber placement head is determined using a compensation air pressure decision model based on Gaussian process regression, including: Calculate the ratio of the design thickness to the corrected thickness under the current number of layers to obtain the thickness correction ratio; Calculate the ratio of the spatial deviation between the ideal trajectory and the corrected trajectory to the reference displacement to obtain the trajectory deviation ratio; Obtain the process characteristics and thermochemical characteristics of the current laying area, wherein the process characteristics include at least the current number of layers and the laying speed, and the thermochemical characteristics include at least the temperature influence factor and the cure degree influence factor; The mold geometry, thickness correction ratio, trajectory deviation ratio, process features, and thermochemical features are concatenated into an input feature vector, which is then input into the compensation pressure decision model to output dynamic compensation pressure.

[0055] First, calculate the ratio of the design thickness to the corrected thickness for the current number of plies to obtain the thickness correction ratio. The design thickness is the theoretical thickness of a single layer specified in the ply design parameters, and the corrected thickness is the expected thickness deviation compensation amount output by the thickness correction model in step S20. The thickness correction ratio reflects the relative degree of thickness deviation at the current ply position caused by factors such as curvature, cumulative layer count, and thermochemical effects. A larger ratio indicates a more severe expected thickness deviation, requiring greater air pressure to compact the fibers.

[0056] Secondly, the trajectory deviation ratio is obtained by calculating the ratio of the spatial deviation between the ideal trajectory and the corrected trajectory to the reference displacement. The spatial deviation between the ideal trajectory and the corrected trajectory refers to the offset distance of the corrected trajectory relative to the ideal trajectory along the normal direction of the die surface at the same spatial point in the same layup. This offset distance is directly determined by the correction thickness. The reference displacement is a preset normalized reference value, representing a dimensionless reference scale used to normalize the spatial deviation. This reference displacement is set comprehensively based on the working stroke range of the filament placement head and the typical magnitude of the correction thickness; for example, it can be set to 1 mm. The trajectory deviation ratio normalizes the spatial deviation to a range of 0 to 1 or larger, facilitating the model's learning of input features of different magnitudes. This trajectory deviation ratio reflects the magnitude of adjustment to the tool center point position to achieve thickness compensation; a larger ratio indicates a larger adjustment magnitude and a corresponding greater demand for air pressure compensation.

[0057] Furthermore, the process characteristics and thermochemical characteristics of the current layup area are obtained. Process characteristics include at least the current number of layups and the layup speed. The current number of layups determines the extent of the cumulative effect of thickness deviation; the larger the number of layups, the more significant the impact of residual stress and cure degree of the preceding layers. The layup speed affects the resin flow time during the pressing process; the faster the speed, the shorter the resin flow time, which may require higher air pressure to achieve sufficient compaction. Thermochemical characteristics include at least a temperature influence factor and a cure degree influence factor. The temperature influence factor reflects the effect of the current temperature on resin flowability, while the cure degree influence factor reflects the effect of the degree of resin crosslinking on the compaction window.

[0058] Furthermore, the mold's geometric features, thickness correction ratio, trajectory deviation ratio, process features, and thermochemical features are concatenated into an input feature vector, which is then input into the compensation pressure decision model to output dynamic compensation pressure. Specifically, the input feature vector is a multi-dimensional vector, composed of the aforementioned features concatenated in a fixed order, and serves as the input to the Gaussian process regression model. The Gaussian process regression model has been trained offline, establishing a mapping relationship between the input feature vector and the optimal pressure value. Based on the current value of the input feature vector and combined with pressure experience under similar feature conditions in the training data, the model outputs dynamic compensation pressure.

[0059] The final dynamic compensation air pressure is the recommended air pressure value predicted by the Gaussian process regression model based on the current input feature vector. It represents the optimal bonding pressure under the current laying position, current number of layers, current geometric features and thermochemical conditions, and is used to drive the pneumatic actuator to adjust the compaction strength of the pressure roller on the fiber bundle.

[0060] In addition, the compensation pressure decision model simultaneously outputs the prediction variance. When the variance exceeds a preset threshold, a robust compensation strategy is triggered. This strategy adds a safety factor multiple of the standard deviation to the predicted mean as the actual applied compensation pressure, ensuring sufficient fit even in areas with low model confidence. The preset threshold is a critical variance value used to determine the reliability of the Gaussian process regression model's predictions under the current input conditions. When the model's output prediction variance exceeds this threshold, it indicates a significant difference between the current input feature vector and the distribution of samples in the training dataset. The model is in an extrapolation region or a region with scarce training data, such as a curvature abrupt change, and the model's prediction results are unreliable. The specific value of the preset threshold can be determined based on the statistical distribution of the posterior variance of each sample point in the training dataset. For example, the 90th quantile of the posterior variance of the training set samples can be used as the threshold; that is, when the prediction variance is higher than 90% of the posterior variance of the training samples, a low confidence judgment is triggered. By setting this threshold, the system can automatically switch to the robust compensation strategy when model uncertainty is high, avoiding under- or over-pressure problems caused by model extrapolation.

[0061] S40: The bonding pressure of the filament placement head at the current number of lay-up layers is controlled according to the dynamic compensation air pressure, and the measured thickness data is collected by the displacement sensor arranged inside the pressure roller of the filament placement head, and the corrected thickness model is updated in a closed loop.

[0062] Finally, after determining the dynamic compensation air pressure, it needs to be applied to the fiber layup head pressure roller, and the model needs to be continuously optimized using measured data. This step sends the dynamic compensation air pressure output by the compensation air pressure decision model to the pneumatic actuator. The pneumatic actuator adjusts its output air pressure according to this instruction, driving the pressure roller to compact the fiber bundle onto the mold or the already laid surface with the corresponding bonding pressure. During the layup process, the displacement sensor array arranged inside the pressure roller collects real-time data on the physical distance between the working surface of the pressure roller and the mold surface. Based on the thickness error, the parameters of the thickness correction model are updated in a closed loop.

[0063] Specifically, the bonding pressure of the yarn-laying head at the current layup layer is controlled according to the dynamically compensated air pressure, and the measured thickness data is collected by displacement sensors arranged inside the pressure roller of the yarn-laying head. The corrected thickness model is then updated in a closed loop, including: Establish a calibration relationship between the measurement readings of the displacement sensor and the physical distance; Based on the calibration relationship, the real-time measurement readings collected by the displacement sensor during the wire laying process are converted into actual physical distances, and the actual thickness is calculated by combining the digital model to construct the actual thickness field; By combining the expected thickness corresponding to the corrected trajectory with the current number of layers and the measured thickness field, the thickness error at each location is calculated; Based on the thickness error, the parameters of the corrected thickness model are updated in a closed loop.

[0064] First, a calibration relationship is established between the displacement sensor's measurement readings and the physical distance. After being installed inside the pressure roller, the displacement sensor's initial output is a voltage or current signal, which needs to be converted into a true physical distance. The calibration process uses a set of standard gauge blocks of known thickness. The sensor probe is sequentially aligned with the surfaces of gauge blocks of different thicknesses, and the corresponding sensor readings are recorded, thus fitting a linear or non-linear mapping relationship between the readings and the distance. Considering that temperature changes can cause impedance drift in the sensor coil, a temperature drift coefficient also needs to be calibrated at different temperatures for online compensation. After calibration, the sensor readings can be converted into millimeter-level physical distances using the calibration formula.

[0065] Secondly, based on the calibration relationship, the real-time measurement readings collected by the displacement sensor during the yarn laying process are converted into measured physical distances, and the measured thickness is calculated by combining the digital model to construct a measured thickness field. During the laying process, the readings of multiple sensors at each sampling moment are calibrated and converted to obtain the physical distance of each measurement point. The actual thickness of the current layup is obtained by subtracting the current measured physical distance from the zero-point calibration distance before laying (i.e., the reference distance on the surface of the bare die), and then subtracting the constant of the pressure roller structure. The thickness data of all measurement points are organized according to spatial coordinates, and a continuous thickness field covering the entire laying area, i.e., the measured thickness field, is reconstructed by interpolation methods such as radial basis function interpolation.

[0066] Furthermore, by combining the expected thickness corresponding to the corrected trajectory for the current number of plies with the measured thickness field, the thickness error at each location is calculated. Specifically, the expected thickness corresponding to the corrected trajectory refers to the thickness value expected to be achieved after superimposing the corrected thickness along the normal direction of the ideal trajectory at the current number of plies, i.e., the design thickness plus the corrected thickness. Subtracting the expected thickness from the measured thickness at the same location in the measured thickness field yields the thickness error at that location. The sign of the thickness error reflects the direction of deviation in the model prediction; a positive value indicates that the actual thickness is greater than the expected thickness, and a negative value indicates that the actual thickness is less than the expected thickness.

[0067] Finally, based on the thickness error, the parameters of the corrected thickness model are updated in a closed loop. Specifically, a recursive least squares algorithm is used, taking the thickness error as information, to iteratively adjust the undetermined parameters in the corrected thickness model. The algorithm first calculates the regression feature vector at the current sampling point, which consists of the partial derivatives of the input features with respect to the corrected thickness. Then, the Kalman gain is calculated to determine the correction weights for the parameter estimates. Finally, the parameter estimates and covariance matrix are updated. After multiple iterations, the model parameters gradually converge to the optimal value that minimizes the thickness error. The updated parameters are immediately used to calculate the corrected thickness for the remaining position of the current layup and subsequent layups, achieving online optimization through learning while laying.

[0068] Through this closed-loop update mechanism, the thickness correction model can continuously learn from the measured data, gradually eliminate prediction bias, and make the correction trajectory increasingly accurate.

[0069] Meanwhile, the measured thickness data can also be used for online updates of the compensation pressure decision model.

[0070] Specifically, after performing a closed-loop update on the parameters of the modified thickness model, the process further includes: Based on the updated modified thickness model, update the modified thickness; Using the updated corrected thickness as the new input feature, and combining it with the measured optimal air pressure value, the data of the current number of layers is added to the training dataset of the compensated air pressure decision model; The posterior distribution of the compensated air pressure decision model is updated online, and alternating iterative optimization of the modified thickness model and the compensated air pressure decision model is performed.

[0071] First, based on the updated corrected thickness model, the corrected thickness of the current or subsequent layup is recalculated. Since the empirical parameters, cumulative exponents, and weighting coefficients in the corrected thickness model have been adjusted according to the measured thickness error, the model's response to surface geometry, thermochemical effects, and mechanical effects is more accurate. Substituting the updated model parameters into the corrected thickness calculation process, the corrected thickness for areas not yet laid is recalculated, ensuring that the corrected trajectory reflects the latest learned process rules in a timely manner.

[0072] Specifically, using the updated corrected thickness as the new input feature, and combining it with the measured optimal air pressure value, the data for the current number of layers is added to the training dataset of the compensation air pressure decision model. The measured optimal air pressure value refers to the air pressure value that, after actual verification, makes the layer thickness closest to the design thickness at the current laying position. By combining the input feature vectors at each sampling point of the current layer with the corresponding measured optimal air pressure values ​​to form new data points, these points are added to the training dataset of the compensation air pressure decision model, allowing the model to gradually accumulate more process experience.

[0073] Furthermore, the posterior distribution of the compensation pressure decision model is updated online, and alternating iterative optimization of the modified thickness model and the compensation pressure decision model is performed. Specifically, the input feature vector of the newly added data points and the measured optimal pressure value are substituted into the posterior update formula of the Gaussian process regression to update the covariance matrix and weight vector in the expressions for the predicted mean and predicted variance. Compared with the batch processing method during offline training, the online update has a smaller computational load and can meet the real-time requirements of the wire laying process.

[0074] Specifically, the thickness correction model and the pressure compensation decision model form a closed-loop relationship of alternating iterative optimization. On one hand, the thickness correction model outputs corrected thickness features for trajectory correction and pressure decision-making; on the other hand, thickness error based on measured thickness is fed back to the thickness correction model, driving its parameter updates; data points based on the measured optimal pressure value are added to the training dataset of the pressure compensation decision model, driving its posterior distribution updates. The two models mutually promote each other in alternating iterations. The thickness correction model provides more accurate corrected thickness features for pressure decision-making, while the pressure compensation decision model provides better pressure control for the layup process, thereby generating higher-quality measured data for continuous learning by both models. Through multiple alternating iterations of layup layers, the prediction accuracy and adaptability of both models continuously improve, ultimately achieving high-quality and stable control of complex curved surface wire layup processes.

[0075] In summary, through this closed-loop update mechanism, the thickness correction model and the pressure compensation decision model can continuously learn from the actual deployment data, gradually eliminate prediction biases, make the correction trajectory and pressure decision more and more accurate, and ultimately control the interlayer thickness deviation within the allowable range.

[0076] In summary, the embodiments of this application have at least the following technical effects: This invention first acquires a digital model of the target mold and the filament placement scheme, extracts the mold's geometric features, and determines the ideal trajectory of the filament placement head tool's center point, providing accurate geometric prior information for subsequent trajectory correction and pressure compensation. Second, a data-driven correction thickness model is established. Using the mold's geometric features and the current number of layups as input, the correction thickness is determined and superimposed on the ideal trajectory to generate the correction trajectory. This achieves quantitative analysis of geometric factors such as curvature and layup number in relation to thickness deviation, replacing traditional trial-and-error methods. Third, combining the mold's geometric features, the deviation between the ideal trajectory and the correction trajectory, a compensation air pressure decision model based on Gaussian process regression is used to determine the dynamic compensation air pressure. This model embeds Hertzian contact theory into the Gaussian process prior, achieving high-precision air pressure compensation decision-making under small sample conditions, while outputting predictive variance to achieve robust control with uncertainty awareness. Finally, the bonding pressure of the filament placement head is controlled according to the dynamic compensation air pressure, and measured thickness data is collected by displacement sensors arranged inside the pressure roller to perform closed-loop updates to the correction thickness model, achieving a leap in information dimension.

[0077] This invention solves the technical problems in the prior art, such as poor thickness consistency between layers on complex curved surfaces, inability to measure in situ in real time, neglect of multi-physics coupling effects, and long adaptation cycle across working conditions. It improves the layer thickness control accuracy and rapid adaptation capability across molds in the automatic wire laying process.

[0078] Example 2, as Figure 2As shown, based on the same inventive concept as the TCP dynamic compensation method for the filament placement head based on an internal displacement sensor provided in Embodiment 1, this embodiment of the invention also provides a TCP dynamic compensation system for the filament placement head based on an internal displacement sensor, comprising: The trajectory acquisition module 11 is used to acquire the digital model and filament placement scheme of the target mold, extract the geometric features of the mold, and determine the ideal trajectory of the center point of the filament placement head tool accordingly. The trajectory correction module 12 is used to establish a data-driven correction thickness model. Taking the geometric features of the mold and the current number of layers as input, it determines the correction thickness corresponding to each position on the mold surface and superimposes it onto the ideal trajectory to generate the correction trajectory. The air pressure decision module 13 is used to determine the dynamic compensation air pressure of the filament placement head by combining the geometric features of the mold, the deviation between the ideal trajectory and the corrected trajectory, and through a compensation air pressure decision model based on Gaussian process regression. The closed-loop update module 14 is used to control the bonding pressure of the filament placement head at the current number of layers according to the dynamic compensation air pressure, and to collect measured thickness data through the displacement sensor arranged inside the pressure roller of the filament placement head, and to perform closed-loop update of the corrected thickness model.

[0079] The trajectory acquisition module 11 is specifically used for: Obtain the digital model and filament placement scheme of the target mold, extract the geometric features of the mold, and determine the ideal trajectory of the center point of the filament placement head tool accordingly, including: Based on the digital model of the target mold, extract the mold geometric features at each position on the mold surface. The mold geometric features include at least the local curvature and normal angle of the surface. Based on the fiber placement scheme, extract the layup design parameters; By combining the digital model, the geometric features of the mold, and the layup design parameters, an ideal trajectory of the center point of the filament placement tool in each layup is generated, and the layup number is marked for each ideal trajectory.

[0080] Specifically, the trajectory correction module 12 is used for: Establish a data-driven corrected thickness model, including: Acquire process calibration experimental data, and combine the constructed input feature vector to perform feature encoding on the process calibration experimental data to form a sample dataset. The input feature vector includes geometric features, thermochemical features, and mechanical features. A regression function model of multiphysics coupling is constructed, which includes geometric components, thermochemical components and mechanical components; The regression function model is iteratively trained and evaluated based on the sample dataset, and the trained modified thickness model is output.

[0081] Specifically, a multi-physics coupling regression function model is constructed, which includes geometric components, thermochemical components, and mechanical components, including: Construct geometric components, which include a thickness deviation term caused by curvature and a cumulative layer deviation term. The thickness deviation term is determined by regression analysis based on the local radius of curvature, the design thickness, and the current number of layers. The cumulative layer deviation term is determined by combining the average curvature and the current number of layers with a power law relationship. Thermochemical components are constructed, including a temperature influence factor and a cure degree influence factor. The temperature influence factor is determined based on the prior relationship between resin viscosity and temperature and is calibrated by offline experiments. The cure degree influence factor is determined based on the prior relationship of the resin curing kinetic model and is calibrated by offline experiments. Mechanical components are constructed, including a residual stress springback term and a shear deformation term. The residual stress springback term is determined based on the ratio of the residual stress of the preceding ply to the ply's equivalent modulus, combined with offline calibration. The shear deformation term is determined based on the shear strain within the fiber plane calibrated offline.

[0082] Further, using the mold geometry and the current layup number as input, the corrected thickness corresponding to each position on the mold surface is determined and superimposed onto the ideal trajectory to generate the corrected trajectory, including: Based on the mold geometry and the current number of layups, and combined with the thermochemical and mechanical features of the current material extracted from the wire laying scheme, a real-time input vector is constructed. The corrected thickness model receives the real-time input vector and outputs the corrected thickness at multiple sampling points on the mold surface. Based on the corrected thickness of multiple sampling points, a continuous thickness correction field on the mold surface is determined by interpolation fitting; Combining physical boundary constraints, the thickness correction field is superimposed onto the ideal trajectory along the normal direction of the mold surface to generate the correction trajectory; The corrected thickness is a vector with a direction mark along the normal direction of the mold surface, and the spacing of the sampling points is determined based on the sampling accuracy of the displacement sensor.

[0083] Specifically, the air pressure decision module 13 is used for: The construction of a compensated pressure decision model based on Gaussian process regression includes: By using Hertzian contact theory as a physical prior and embedding it into the mean function of a Gaussian process regression model, a Gaussian process regression model with embedded physical information is constructed. A training dataset is obtained through process calibration experiments, and the Gaussian process regression model is trained based on the training dataset, wherein each data point in the training dataset includes the experimentally verified optimal air pressure value under given feature conditions. Based on the prediction variance output by the trained Gaussian process regression model, and combined with the distribution characteristics of the training data, an uncertainty safety factor is determined, which is used to compensate for the uncertainty of the prediction. The characteristic conditions include at least process characteristics, thermochemical characteristics, geometric characteristics, and deviation evaluation characteristics.

[0084] Furthermore, combining the mold geometry and the deviation between the ideal trajectory and the corrected trajectory, the dynamic compensation air pressure of the fiber placement head is determined using a compensation air pressure decision model based on Gaussian process regression, including: Calculate the ratio of the design thickness to the corrected thickness under the current number of layers to obtain the thickness correction ratio; Calculate the ratio of the spatial deviation between the ideal trajectory and the corrected trajectory to the reference displacement to obtain the trajectory deviation ratio; Obtain the process characteristics and thermochemical characteristics of the current laying area, wherein the process characteristics include at least the current number of layers and the laying speed, and the thermochemical characteristics include at least the temperature influence factor and the cure degree influence factor; The mold geometry, thickness correction ratio, trajectory deviation ratio, process features, and thermochemical features are concatenated into an input feature vector, which is then input into the compensation pressure decision model to output dynamic compensation pressure.

[0085] Specifically, the closed-loop update module 14 is used for: The bonding pressure of the yarn-laying head at the current layup layer is controlled by the dynamically compensated air pressure, and the measured thickness data is collected by displacement sensors arranged inside the pressure roller of the yarn-laying head. The corrected thickness model is then updated in a closed loop, including: Establish a calibration relationship between the measurement readings of the displacement sensor and the physical distance; Based on the calibration relationship, the real-time measurement readings collected by the displacement sensor during the wire laying process are converted into actual physical distances, and the actual thickness is calculated by combining the digital model to construct the actual thickness field; By combining the expected thickness corresponding to the corrected trajectory with the current number of layers and the measured thickness field, the thickness error at each location is calculated; Based on the thickness error, the parameters of the corrected thickness model are updated in a closed loop.

[0086] Furthermore, after performing a closed-loop update on the parameters of the modified thickness model, the method further includes: Based on the updated modified thickness model, update the modified thickness; Using the updated corrected thickness as the new input feature, and combining it with the measured optimal air pressure value, the data of the current number of layers is added to the training dataset of the compensated air pressure decision model; The posterior distribution of the compensated air pressure decision model is updated online, and alternating iterative optimization of the modified thickness model and the compensated air pressure decision model is performed.

Claims

1. A TCP dynamic compensation method for wire placement heads based on internal displacement sensors, characterized in that, include: Obtain the digital model and filament placement scheme of the target mold, extract the geometric features of the mold, and determine the ideal trajectory of the center point of the filament placement head tool accordingly; A data-driven correction thickness model is established. Using the mold geometry and the current layup number as input, the correction thickness corresponding to each position on the mold surface is determined and superimposed on the ideal trajectory to generate the correction trajectory. Combining the mold geometry and the deviation between the ideal trajectory and the corrected trajectory, the dynamic compensation air pressure of the filament placement head is determined using a compensation air pressure decision model based on Gaussian process regression. The bonding pressure of the filament placement head at the current number of lay-up layers is controlled by the dynamic compensation air pressure, and the measured thickness data is collected by the displacement sensor arranged inside the pressure roller of the filament placement head to perform closed-loop update of the corrected thickness model.

2. The TCP dynamic compensation method for the filament placement head based on an internal displacement sensor as described in claim 1, characterized in that, Obtain the digital model and filament placement scheme of the target mold, extract the geometric features of the mold, and determine the ideal trajectory of the center point of the filament placement head tool accordingly, including: Based on the digital model of the target mold, extract the mold geometric features at each position on the mold surface. The mold geometric features include at least the local curvature and normal angle of the surface. Based on the fiber placement scheme, extract the layup design parameters; By combining the digital model, the geometric features of the mold, and the layup design parameters, an ideal trajectory of the center point of the filament placement tool in each layup is generated, and the layup number is marked for each ideal trajectory.

3. The TCP dynamic compensation method for the filament placement head based on an internal displacement sensor as described in claim 1, characterized in that, Establish a data-driven corrected thickness model, including: Acquire process calibration experimental data, and combine the constructed input feature vector to perform feature encoding on the process calibration experimental data to form a sample dataset. The input feature vector includes geometric features, thermochemical features, and mechanical features. A regression function model of multiphysics coupling is constructed, which includes geometric components, thermochemical components and mechanical components; The regression function model is iteratively trained and evaluated based on the sample dataset, and the trained modified thickness model is output.

4. The TCP dynamic compensation method for the filament placement head based on an internal displacement sensor as described in claim 3, characterized in that, A multiphysics coupling regression function model is constructed, comprising geometric, thermochemical, and mechanical components, including: Construct geometric components, which include a thickness deviation term caused by curvature and a cumulative layer deviation term. The thickness deviation term is determined by regression analysis based on the local radius of curvature, the design thickness, and the current number of layers. The cumulative layer deviation term is determined by combining the average curvature and the current number of layers with a power law relationship. Thermochemical components are constructed, including a temperature influence factor and a cure degree influence factor. The temperature influence factor is determined based on the prior relationship between resin viscosity and temperature and is calibrated by offline experiments. The cure degree influence factor is determined based on the prior relationship of the resin curing kinetic model and is calibrated by offline experiments. Mechanical components are constructed, including a residual stress springback term and a shear deformation term. The residual stress springback term is determined based on the ratio of the residual stress of the preceding ply to the ply's equivalent modulus, combined with offline calibration. The shear deformation term is determined based on the shear strain within the fiber plane calibrated offline.

5. The TCP dynamic compensation method for the filament placement head based on an internal displacement sensor as described in claim 1, characterized in that, Using the mold geometry and the current layup number as input, the correction thickness corresponding to each position on the mold surface is determined and superimposed onto the ideal trajectory to generate the correction trajectory, including: Based on the mold geometry and the current number of layups, and combined with the thermochemical and mechanical features of the current material extracted from the wire laying scheme, a real-time input vector is constructed. The corrected thickness model receives the real-time input vector and outputs the corrected thickness at multiple sampling points on the mold surface. Based on the corrected thickness of multiple sampling points, a continuous thickness correction field on the mold surface is determined by interpolation fitting; Combining physical boundary constraints, the thickness correction field is superimposed onto the ideal trajectory along the normal direction of the mold surface to generate the correction trajectory; The corrected thickness is a vector with a direction mark along the normal direction of the mold surface, and the spacing of the sampling points is determined based on the sampling accuracy of the displacement sensor.

6. The TCP dynamic compensation method for the filament placement head based on an internal displacement sensor as described in claim 1, characterized in that, The construction of a compensated pressure decision model based on Gaussian process regression includes: By using Hertzian contact theory as a physical prior and embedding it into the mean function of a Gaussian process regression model, a Gaussian process regression model with embedded physical information is constructed. A training dataset is obtained through process calibration experiments, and the Gaussian process regression model is trained based on the training dataset, wherein each data point in the training dataset includes the experimentally verified optimal air pressure value under given feature conditions. Based on the prediction variance output by the trained Gaussian process regression model, and combined with the distribution characteristics of the training data, an uncertainty safety factor is determined, which is used to compensate for the uncertainty of the prediction. The characteristic conditions include at least process characteristics, thermochemical characteristics, geometric characteristics, and deviation evaluation characteristics.

7. The TCP dynamic compensation method for the filament placement head based on an internal displacement sensor as described in claim 6, characterized in that, Combining the mold geometry and the deviation between the ideal trajectory and the corrected trajectory, the dynamic compensation air pressure of the fiber placement head is determined using a compensation air pressure decision model based on Gaussian process regression, including: Calculate the ratio of the design thickness to the corrected thickness under the current number of layers to obtain the thickness correction ratio; Calculate the ratio of the spatial deviation between the ideal trajectory and the corrected trajectory to the reference displacement to obtain the trajectory deviation ratio; Obtain the process characteristics and thermochemical characteristics of the current laying area, wherein the process characteristics include at least the current number of layers and the laying speed, and the thermochemical characteristics include at least the temperature influence factor and the cure degree influence factor; The mold geometry, thickness correction ratio, trajectory deviation ratio, process features, and thermochemical features are concatenated into an input feature vector, which is then input into the compensation pressure decision model to output dynamic compensation pressure.

8. The TCP dynamic compensation method for the filament placement head based on an internal displacement sensor as described in claim 1, characterized in that, The bonding pressure of the yarn-laying head at the current layup layer is controlled by the dynamically compensated air pressure, and the measured thickness data is collected by displacement sensors arranged inside the pressure roller of the yarn-laying head. The corrected thickness model is then updated in a closed loop, including: Establish a calibration relationship between the measurement readings of the displacement sensor and the physical distance; Based on the calibration relationship, the real-time measurement readings collected by the displacement sensor during the wire laying process are converted into actual physical distances, and the actual thickness is calculated by combining the digital model to construct the actual thickness field; By combining the expected thickness corresponding to the corrected trajectory with the current number of layers and the measured thickness field, the thickness error at each location is calculated; Based on the thickness error, the parameters of the corrected thickness model are updated in a closed loop.

9. The TCP dynamic compensation method for the filament placement head based on an internal displacement sensor as described in claim 8, characterized in that, After performing a closed-loop update on the parameters of the modified thickness model, the method further includes: Based on the updated modified thickness model, update the modified thickness; Using the updated corrected thickness as the new input feature, and combining it with the measured optimal air pressure value, the data of the current number of layers is added to the training dataset of the compensated air pressure decision model; The posterior distribution of the compensated air pressure decision model is updated online, and alternating iterative optimization of the modified thickness model and the compensated air pressure decision model is performed.

10. A TCP dynamic compensation system for a wire placement head based on an internal displacement sensor, characterized in that, The method for performing the TCP dynamic compensation method for the filament placement head based on an internal displacement sensor as described in any one of claims 1-9 includes: The trajectory acquisition module is used to acquire the digital model and filament placement scheme of the target mold, extract the geometric features of the mold, and determine the ideal trajectory of the center point of the filament placement head tool accordingly. The trajectory correction module is used to establish a data-driven correction thickness model. Taking the mold geometry features and the current layup number as input, it determines the correction thickness corresponding to each position on the mold surface and superimposes it onto the ideal trajectory to generate the correction trajectory. The air pressure decision module is used to determine the dynamic compensation air pressure of the filament placement head by combining the geometric features of the mold, the deviation between the ideal trajectory and the corrected trajectory, and through a compensation air pressure decision model based on Gaussian process regression. The closed-loop update module is used to control the bonding pressure of the filament placement head at the current number of layers according to the dynamic compensation air pressure, and to collect measured thickness data through displacement sensors arranged inside the pressure roller of the filament placement head, and to perform closed-loop update of the corrected thickness model.

Images