Automated filament placement process optimization method and apparatus considering mold geometry
By constructing a two-dimensional feature tensor and a multi-objective optimization function, and combining the mold geometry, material properties and layup information, the problem of uneven layup quality in complex surface components during automatic filament layup was solved, achieving efficient and high-quality manufacturing results.
Patent Information
- Application Number
- CN202611325744.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
AI Technical Summary
Existing methods for optimizing automatic wire placement process parameters fail to fully consider the geometric characteristics of the mold, making it difficult to balance the placement quality of complex surface components in different areas. Furthermore, a single prediction model cannot effectively characterize the unique mechanisms of various defects, resulting in limited prediction accuracy and making it difficult for the optimization results to reflect the actual quality distribution of the components.
A two-dimensional feature tensor is constructed, which is combined with the three-dimensional geometric model of the mold, material properties and ply information. An independently trained defect prediction model is used to predict defects in each set of feature tensors. The target process parameters are determined by a multi-objective optimization function to achieve differentiated configuration for different geometric regions and ply positions.
It enables the laying quality of different areas in complex surface components to be taken into account, improving manufacturing efficiency and quality, and meeting the needs of high-quality and high-efficiency manufacturing.
Smart Images

Figure CN123120143A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent manufacturing technology for composite materials, and in particular to an automatic fiber placement process optimization method and apparatus that takes into account the geometric characteristics of the mold. Background Technology
[0002] Composite materials, due to their superior properties such as high specific strength, high specific stiffness, and strong designability, have been widely used in high-end manufacturing fields such as aerospace, rail transportation, and new energy equipment. Automated Fiber Placement (AFP), as an advanced automated molding process for composite materials, enables the efficient and high-quality manufacturing of complex-shaped components and is one of the core processes in the current field of composite component manufacturing. However, the automated fiber placement process involves the coupled effects of multiple process parameters such as placement speed, placement pressure, placement tension, and placement temperature. It is also affected by multiple factors such as mold geometry, layup design, and material properties, which can easily lead to various placement defects such as bubbles, wrinkles, overlaps, twists, and gaps, seriously affecting the mechanical properties and service life of the components.
[0003] Currently, the main shortcomings of automated fiber placement process parameter optimization methods are as follows: First, most existing process parameter optimization methods only use die type (such as flat plate, punch, and die) as discrete classification feature input models, failing to deeply characterize the three-dimensional geometric features of the die. In actual production, the curvature distribution varies significantly in different regions of the same component, especially for complex hyperbolic surfaces. Using globally uniform process parameters makes it difficult to balance the placement quality of flat and high-curvature areas. For example, in the placement process of complex surface components such as S-shaped air intakes, wing leading edges, and radomes, hyperbolic areas often become high-incidence areas for defects, and existing methods cannot configure differentiated process parameters for different geometric regions. Moreover, a single prediction model (such as GA-CNN, PI-GAN, etc.) is commonly used to predict multiple defects simultaneously. However, the formation mechanisms of different types of defects are fundamentally different: bubbles are mainly caused by the coupling effect of pressure and temperature, wrinkles are closely related to placement tension and local curvature, and overlaps are greatly affected by placement speed and layup angle. Using a single model makes it difficult to fully characterize the unique mechanisms of various defects, resulting in limited prediction accuracy. Furthermore, when constructing multi-objective optimization functions, most methods employ simple global summation or averaging, failing to fully consider the cumulative effect of defect propagation along the ply thickness direction, nor the differences in the contribution of different geometric regions to the overall component performance. This simplification makes it difficult for the optimization results to accurately reflect the actual quality distribution of the component. Additionally, the currently output process parameters are typically globally uniform values, meaning the entire component uses the same set of layup parameters. This approach has significant limitations when dealing with complex surface components: reducing the overall speed to ensure quality in high-curvature areas sacrifices production efficiency in flat areas; increasing the speed to improve efficiency in flat areas leads to an increase in defects in high-curvature areas. There is a lack of ability to configure differentiated parameters based on geometric regions and ply positions.
[0004] In summary, existing automated fiber placement process parameter optimization methods have shortcomings in geometric feature characterization, multi-defect collaborative prediction, objective function construction, and parameter optimization modes, making it difficult to meet the demands for high-quality and high-efficiency manufacturing of complex composite material components. Therefore, there is an urgent need for an automated fiber placement process parameter optimization method that can fully consider mold geometry, achieve multi-defect collaborative optimization, and support spatial-layer differentiated parameter configuration. Summary of the Invention
[0005] This invention provides an automatic wire placement process optimization method and apparatus that takes into account the geometric characteristics of the mold, in order to solve the problem of difficulty in meeting the high-quality and high-efficiency manufacturing requirements of complex surface composite material components, and can meet the high-quality and high-efficiency manufacturing requirements of complex surface composite material components.
[0006] This invention provides an automated wire placement process optimization method considering mold geometry, comprising: Obtain the three-dimensional geometric model of the mold, the material properties of the component to be optimized, and the layup information; Using the three-dimensional geometric model, the material property information, and the ply information, several sets of two-dimensional feature tensors are constructed. The features in each set of two-dimensional feature tensors include candidate process parameters for each geometric region in each ply of the component to be optimized, geometric feature vectors determined based on the three-dimensional geometric model of the mold, material property features determined based on the material property information, and layer orientation features determined based on the ply information. Several pre-built defect prediction models are used to predict defects for each set of two-dimensional feature tensors, so as to obtain the prediction of the component to be optimized for each defect category corresponding to each two-dimensional feature tensor. The defect categories that different defect prediction models can predict are different and each defect prediction model is trained independently. A multi-objective optimization function is constructed, and based on the multi-objective optimization function and the prediction of the component to be optimized for each defect category corresponding to each of the two-dimensional feature tensors, the target process parameters corresponding to the target optimization mode are determined. The target optimization mode is obtained based on the three-dimensional geometric model and the layup information.
[0007] The present invention also provides an automatic wire placement process optimization device that takes into account the geometric features of the mold, comprising: The data acquisition module is used to acquire the three-dimensional geometric model of the mold, the material property information of the component to be optimized, and the layup information; The feature tensor construction module is used to construct several sets of two-dimensional feature tensors using the three-dimensional geometric model, the material property information, and the ply information. The features in each set of two-dimensional feature tensors include candidate process parameters for each geometric region in each ply of the component to be optimized, geometric feature vectors determined based on the three-dimensional geometric model of the mold, material property features determined based on the material property information, and layer orientation features determined based on the ply information. The defect prediction module is used to predict defects for each set of two-dimensional feature tensors using several pre-built defect prediction models, and to obtain the prediction of the component to be optimized for each defect category for each two-dimensional feature tensor. The defect categories that different defect prediction models can predict are different and each defect prediction model is trained independently. The process parameter determination module is used to construct a multi-objective optimization function and, based on the multi-objective optimization function and the prediction of the component to be optimized for each defect category corresponding to each of the two-dimensional feature tensors, determine the target process parameters corresponding to the target optimization mode. The target optimization mode is obtained based on the three-dimensional geometric model and the layup information.
[0008] This invention provides an automatic wire placement process optimization method and apparatus that considers the geometric features of the mold. During the process optimization, the three-dimensional geometric features of the mold, material properties, layup positions, and candidate process parameters are all incorporated into a two-dimensional feature tensor. This allows the evaluation of process parameters to correspond to different geometric regions and different layup positions. Defect prediction models trained independently for different defect categories are used for prediction, enabling various defects to be characterized and evaluated based on their respective input features. Furthermore, a multi-objective optimization function integrates the prediction results of multiple defect types under different candidate process parameters, and the target process parameters are determined by combining the target optimization mode determined by the mold's three-dimensional geometric model and layup information. This allows the optimization of automatic wire placement process parameters to take into account the differences in geometric regions, layup differences, and prediction results of different defect categories of the component. It solves the problem that it is difficult to consider the layup quality of different regions when using a single global process parameter for complex surface components, and meets the demand for high-quality and high-efficiency manufacturing of complex surface composite material components. Attached Figure Description
[0009] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0010] Figure 1 This is one of the flowcharts of the automatic wire placement process optimization method that takes into account the geometric features of the mold provided by the present invention.
[0011] Figure 2 This is the second flowchart of the automatic wire placement process optimization method considering the geometric features of the mold provided by the present invention.
[0012] Figure 3 This is a schematic diagram of the partitioning of the three-dimensional geometric model of the mold provided by the present invention.
[0013] Figure 4 This is a schematic diagram of the dual-algorithm architecture based on the multi-defect independent prediction model and the process parameter optimization method based on NSGA-II provided by the present invention.
[0014] Figure 5 This is a schematic diagram of the automatic wire placement process optimization device that takes into account the geometric features of the mold provided by the present invention. Detailed Implementation
[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0016] The automatic wire placement process optimization method considering the geometric features of the mold provided by this invention can be applied to the automatic wire placement process planning of composite material components, such as components with complex shapes, such as wing leading edges, radomes, and composite fan blades of aero engines.
[0017] Please see Figure 1 The automatic wire placement process optimization method considering mold geometry provided by the present invention may include the following steps: Step 101: Obtain the three-dimensional geometric model of the mold, the material property information of the component to be optimized, and the layup information.
[0018] Among them, the three-dimensional geometric model can be imported through the computer-aided design interface, the material property information can be obtained from the material database or the human-computer interaction interface, and the ply information can be obtained from the ply design file.
[0019] Step 102: Construct several sets of two-dimensional feature tensors using the three-dimensional geometric model, material property information, and ply information.
[0020] The spatial dimension of the two-dimensional feature tensor corresponds to different geometric regions, and the layer dimension corresponds to different plies; each tensor element corresponds to a ply and a geometric region, and contains corresponding candidate process parameters, geometric feature vectors, material property features and layer features.
[0021] Candidate process parameters can be used to characterize the selectable process states of automated fiber placement equipment during the placement process. Geometric feature vectors can be used to describe the local geometry of the die surface. Material property features can be used to describe the material states of fibers, resins, and prepregs. Layer orientation features can be used to describe the position and placement direction of the current layup within the entire layup system. By filling the above information into a two-dimensional feature tensor, each tensor element can correspond to a specific layup and geometric region, and candidate process parameters, geometric features, material properties, and layer orientation features can be integrated into this element, allowing subsequent defect prediction to retrieve relevant information within the same data structure.
[0022] Step 103: Using several pre-built defect prediction models, perform defect prediction on each group of two-dimensional feature tensors to obtain the prediction of the component to be optimized for each defect category corresponding to each two-dimensional feature tensor.
[0023] Different defect prediction models can predict different defect categories, and each defect prediction model is trained independently. Specifically, the defect prediction model corresponding to a defect category is trained using the corresponding optimized feature tensor samples and their predicted value labels.
[0024] This allows for the creation of corresponding defect prediction models for different defect categories. These models can include: bubble prediction models, wrinkle prediction models, overlap prediction models, torsion prediction models, and gap prediction models. Each model can be trained and predicted independently to address the feature sensitivity and spatial propagation characteristics of different defect categories. The defect prediction models can be trained using optimized feature tensor samples and their predicted value labels for the corresponding defect category. These labels can be derived from historical production data, quality inspection data, or process test data, and are used to characterize the defect situation of the samples in the corresponding layup and geometric region.
[0025] Optionally, for defect prediction with strong spatial locality and weak cross-location coupling, such as bubbles and gaps, the extreme gradient boosting (XGBoost) algorithm can be used to build a defect prediction model. In this case, the optimized feature tensor samples of each geometric region under each layup can be used as training samples to form tabular training data. The objective function of the extreme gradient boosting algorithm model can be expressed as: in, This represents the objective function for model training. Represents the loss function. Indicates the first The predicted value for each sample, Indicates the first The label value of each sample, Indicates the first The regularization term for the tree, Indicates the number of decision trees. Indicates the number of samples. Indicates the first The function corresponding to each tree. The regularization term can be expressed as: in, This represents the total number of leaf nodes in the decision tree. This represents a hyperparameter used to control the number of leaf nodes. Represents the regularization coefficient. Indicates the first The weights of each leaf node in the decision tree. By combining the loss function and regularization term, the complexity of the tree structure can be controlled while fitting the defect prediction values.
[0026] For predicting defects that spread continuously across layers and regions, such as wrinkles, overlaps, and twists, a Convolutional Neural Network (CNN) can be used to build a defect prediction model. In this case, the complete spatial and layered structure of the optimized feature tensor can be preserved. The ply number and geometric region number are considered as two spatial dimensions of the feature map. A two-dimensional convolutional kernel is used to slide simultaneously along the ply direction and the region direction to capture the continuous changes in defects between layers and regions. The convolution calculation can be performed layer-by-layer and region-by-region convolution with bias, followed by summation, and then the result is obtained using the Rectified Linear Unit (ReLU) function.
[0027] In one possible implementation, an XGBoost-based regression sub-model can be trained independently for each bubble and gap. During the training phase, the parameters of each regression sub-model can be set as follows: the number of decision trees can be 200, the maximum tree depth can be 6, the learning rate can be 0.1, the sampling rate of subsamples can be 0.6, and the feature sampling rate can be 0.6. This setting is only an example and is not intended to limit this solution.
[0028] For wrinkles, overlaps, and twists, a separate CNN-based prediction model is trained. For example, the number of convolutional layers in the three prediction models can be 3, the kernel size can be 3×3, the number of channels in each layer can be [32, 64, 128], the convolution stride can be 1, and the dropout ratio can be 0.3. This prediction model structure is merely an example and is not intended to limit the structure of the prediction model provided in this invention. Specifically, the optimized feature tensor for each defect category can be determined first, and then the optimized feature tensor for the corresponding defect category can be processed using the respective defect prediction model to output the predicted value for each geometric region within each layup. The methods for determining the optimized feature tensor and the methods for determining the optimized feature tensor samples can be found in the methods for obtaining the optimized feature tensor described below, and will not be elaborated here.
[0029] Step 104: Construct a multi-objective optimization function, and based on the multi-objective optimization function and the prediction of the component to be optimized for each defect category corresponding to each of the two-dimensional feature tensors, determine the target process parameters corresponding to the target optimization mode.
[0030] The target optimization mode is obtained based on the three-dimensional geometric model and the layup information, or it can be determined according to the user's selection in other embodiments.
[0031] Each set of two-dimensional feature tensors corresponds to candidate process parameters. By comparing the defect prediction results corresponding to each candidate process parameter, the target process parameter is determined from the candidate process parameters.
[0032] In the above scheme, the three-dimensional geometric features of the mold, material properties, layup positions, and candidate process parameters are all incorporated into a two-dimensional feature tensor. This allows the evaluation of process parameters to correspond to different geometric regions and different layup positions. Defect prediction models trained independently for different defect categories are used for prediction, enabling various defects to be characterized and evaluated separately based on their corresponding input features. Furthermore, a multi-objective optimization function is used to synthesize the prediction results of multiple types of defects under different candidate process parameters. The target process parameters are determined by combining the target optimization mode determined by the three-dimensional geometric model of the mold and the layup information. This allows the optimization of automatic fiber placement process parameters to take into account the differences in geometric regions of the component, layup differences, and prediction results of different defect categories. It can solve the problem that it is difficult to take into account the layup quality of different regions when using a single global process parameter for complex surface components, and can meet the needs of high-quality and high-efficiency manufacturing of complex surface composite material components.
[0033] In some possible embodiments, step 102 described above may include, for example: Figure 2 The following steps are shown: Step 201: Based on the curvature characteristics at multiple locations in the three-dimensional geometric model, the three-dimensional geometric model is divided into regions to obtain multiple geometric regions.
[0034] First, a geometric analysis can be performed on the three-dimensional geometric model of the mold to obtain the curvature characteristics at multiple locations within the model. In actual production, the curvature distribution in different regions of the same component may vary significantly. For example, complex hyperbolic surface regions are generally more prone to laying quality fluctuations than flat regions. Therefore, the three-dimensional geometric model can be divided into regions based on the curvature characteristics at multiple locations, resulting in multiple geometric regions.
[0035] In one implementation, the three-dimensional geometric model can be triangulated to obtain a set of mesh vertices and a set of triangular faces. The curvature characteristics at the mesh vertices can then be calculated using discrete differential geometry methods. For example, the mesh vertices can be traversed, and the average curvature can be calculated using the discrete Laplace-Beltrami operator. mean curvature It can be used to reflect the average curvature of a surface at a local location; the angle deficit method can also be used to calculate Gaussian curvature. Gaussian curvature It can be used to reflect the comprehensive relationship between the degree of curvature of a local surface in two principal directions.
[0036] For example, for mesh vertices average curvature The calculation can be expressed as: in, Represents vertices The area of Voronoi. and Indicates the relationship with the edge Two opposite angles, Represents grid vertices The set of vertices in the ring neighborhood, This represents the adjacent vertices in a ring neighborhood. For grid vertices... Gaussian curvature The calculation can be expressed as: in, Represents grid vertices Gaussian curvature at that point Represents vertices The area of Voronoi. Represents grid vertices A set of triangular facets in the neighborhood of a ring. Represents grid vertices First The interior angles of adjacent triangular facets. Based on the mean curvature. and Gaussian curvature It can also calculate the principal curvature. and Principal curvature can be used to describe the degree of curvature of a surface in two principal directions, and its calculation can be expressed as: ; ; in, and These represent the principal curvatures in the two principal directions. Through the above curvature calculations, the three-dimensional geometric model can be transformed from a digital model that only represents the shape into a geometric expression with curvature information.
[0037] After obtaining the curvature characteristics, the mold surface can be divided into multiple geometric regions based on curvature thresholds. For example, a principal curvature threshold can be set. and according to and With principal curvature threshold The relationship identifies different types of geometric regions. For example, if... < If it is included in a flat area, and < If it is classified into the single curvature region, This can be classified into the hypercurvature region; for example, the classified geometric region can be found in [reference needed]. Figure 3 , Figure 3 The red area represents the hyperbolic region, the yellow area represents the monobolic region, and the green area at the lower left represents the flat region. This division method can separate the parts of the mold surface with large geometric differences, allowing subsequent tensor construction to preserve spatial regional differences.
[0038] Step 202: Establish a two-dimensional tensor framework according to the ply sequence number and the region number of multiple geometric regions in the ply information.
[0039] For example, a two-dimensional tensor framework is established using the ply number in the ply information as the first dimension and the region number of multiple geometric regions as the second dimension. If the total number of ply layers is... The total number of geometric regions in each layer is Then the two-dimensional tensor framework can correspond to The nth position, each position corresponds to the nth position. Layer, First Geometric regions, among which Indicates the ply number. This indicates the geometric region number. This framework allows each ply and each geometric region to have an independent data storage location, facilitating the representation of process differences for the same geometric region across different plies, and also facilitating the representation of process differences for the same ply across different geometric regions.
[0040] In some applications, ply information can include the total number of layers, the ply angle of each layer, and the ply sequence. The ply number reflects the position of the current ply in the thickness direction, and the region number reflects the mold surface region corresponding to the current tensor element.
[0041] Step 203: For each two-dimensional feature tensor, fill the two-dimensional tensor framework with one set of candidate process parameters, geometric feature vectors, material property features and layer orientation features of each geometric region within each ply, to obtain the two-dimensional feature tensor.
[0042] For each set of candidate process parameters, a corresponding two-dimensional feature tensor can be constructed. For each position in the two-dimensional tensor framework, a set of candidate process parameters under the corresponding ply and geometric region can be filled into that position, along with the geometric feature vector corresponding to that geometric region, the material property features corresponding to the component to be optimized, and the layer orientation features corresponding to that ply.
[0043] For example, in the automated fiber layup manufacturing of composite material fan blades for a certain type of aero-engine, the surface of the fan blade mold can be divided into multiple geometric regions, and multi-layer layup design information can be read. If the total number of layups is... The total number of geometric regions in each layer is The feature dimension is Then a two-dimensional feature tensor can be constructed. .
[0044] As can be seen, this scheme divides regions based on curvature features and establishes a two-dimensional tensor framework according to the ply number and region number. This allows the spatial differences of complex surface components and the differences in ply thickness direction to be incorporated into a unified feature expression, thereby improving the completeness of the two-dimensional feature tensor in representing the state of the automated fiber placement process.
[0045] In some embodiments, for each geometric region within each layup, candidate process parameters include candidate layup speed, candidate layup pressure, candidate layup tension, and candidate layup temperature; layer orientation features include layup angle and normalized layup number; material property features include fiber volume fraction, resin viscosity, prepreg viscosity coefficient, glass transition temperature, and single-layer thickness; geometric feature vectors include local average curvature, local Gaussian curvature, and the angle between the layup direction and the principal curvature direction, where the principal curvature direction may be the principal curvature. or The corresponding principal direction, for example, the principal curvature direction could be The corresponding main direction.
[0046] Candidate process parameters can be preset within a range based on equipment capacity, material processing window, and historical process experience. For example, the laying speed can be selected from 0.5 m / s to 1 m / s, the laying pressure from 300 Newtons to 1200 Newtons, the laying tension from 5 Newtons to 15 Newtons, and the laying temperature from 30°C to 80°C. The above ranges are only for illustrating one possible implementation; the specific ranges can be configured according to the material system and equipment capacity.
[0047] The ply angle represents the fiber placement direction relative to a reference direction in the current ply, and can be, for example, 0°, ±45°, or 90°, which can be determined by the ply design document. The normalized ply number represents the relative position of the current ply within the total number of plies. If the current ply number is... The total number of plies is Then the normalized ply number can be expressed as / Normalized layup numbers can be used to reflect the positional differences of the bottom, intermediate, and top layers during the layup process, enabling defect prediction and process parameter selection to perceive the impact of layer position on layup quality.
[0048] Taking T800 / epoxy prepreg as an example, material properties can include fiber volume fraction. =60%, resin viscosity Prepreg viscosity coefficient =0.85, glass transition temperature =190℃ and single-layer thickness =0.125mm. The above material properties can be used as constant features to fill in the tensor elements of the corresponding components.
[0049] Local average curvature can be statistically obtained from the average curvature of the mesh vertices within the geometric region, reflecting the average degree of surface curvature in that region. Local Gaussian curvature can be statistically obtained from the Gaussian curvature of the mesh vertices within the geometric region, reflecting the relationship between the curvatures of the two principal directions in that region. The angle between the layup direction and the principal curvature direction can be calculated by comparing the layup direction of the current layup with the principal curvature direction within the geometric region. Its value can be used to describe the relative relationship between the fiber layup direction and the principal curvature direction of the surface; this angle can be denoted as... ,in Indicates the first The layer in the first The angle between the layup direction and the principal curvature direction in a geometric region. By filling the local average curvature, local Gaussian curvature, and the angle between the layup direction and the principal curvature direction into a two-dimensional feature tensor, the tensor elements can express the local constraints related to the die geometry during the automatic filament placement process.
[0050] In some embodiments, for each geometric region within each ply, the two-dimensional feature tensor may further include: a velocity-curvature coupling feature determined based on the candidate ply velocity and the local average curvature. For example, for the first... Layer, First Each geometric region can be determined based on its candidate laying speed. and local mean curvature Construction of velocity and curvature coupling characteristics , Represents the absolute value of the local average curvature; the pressure and curvature compensation characteristics are determined based on the candidate laying pressure and the local Gaussian curvature. For example, for the first... Layer, First Each geometric region can be determined based on its candidate layup pressure. and local Gaussian curvature Structural pressure and curvature compensation characteristics / (1+ ), Represents the absolute value of the local Gaussian curvature; based on the tension and curvature coupling characteristics determined by the candidate layup tension and the local Gaussian curvature, for example, for the... Layer, First Each geometric region can be determined based on its candidate layup tension. and local Gaussian curvature Construction of tension and curvature coupling characteristics Based on the temperature and velocity coupling characteristics determined by the candidate placement temperature and candidate placement speed, for example, for the first... Layer, First Each geometric region can be selected based on its candidate layup temperature. and candidate deployment speed Construct temperature and velocity coupling characteristics / .
[0051] In some embodiments, determining the target optimization mode may include the following steps: First, based on the number of layers in the layup information and the curvature characteristics at multiple locations in the 3D geometric model, the target optimization mode is determined. This target optimization mode can be a global parameter mode, a partitioned parameter mode, a layer-by-layer parameter mode, or a joint parameter mode. For example, if the curvature at each location is high and the number of layers is high, a joint parameter mode can be selected; if the curvature at each location is low and the number of layers is low, a global parameter mode can be selected. That is, the mapping relationship between the curvature characteristics and the number of layers at each location and the target optimization mode can be pre-constructed, and the target optimization mode is determined based on this mapping relationship.
[0052] Among them, the number of layers in the layup information can reflect the layup scale of the component to be optimized along the thickness direction, and the curvature features at multiple locations in the three-dimensional geometric model can reflect the geometric complexity of the mold surface in the spatial direction.
[0053] The global parameter mode indicates that the entire component to be optimized uses a set of candidate process parameters; the partition parameter mode indicates that each geometric region corresponds to a set of candidate process parameters; the layer-by-layer parameter mode indicates that each ply corresponds to a set of candidate process parameters; and the joint parameter mode indicates that each geometric region within each ply corresponds to a separate set of candidate process parameters.
[0054] The methods for determining the candidate process parameters in each of the two-dimensional feature tensors include: Based on the target optimization model, candidate process parameters for each geometric region within each layup are determined. The dimensions of the candidate process parameters differ depending on the target optimization model.
[0055] After determining the target optimization mode, candidate process parameters for each geometric region within each ply can be generated or configured in this manner. If a global parameter mode is used, the same set of candidate layup speeds, candidate layup pressures, candidate layup tensions, and candidate layup temperatures can be filled into the tensor elements corresponding to all plies and all geometric regions. If a partitioned parameter mode is used, the same geometric region can use the candidate process parameters corresponding to that geometric region in different plies, and the candidate process parameters between different geometric regions can be different. If a layer-by-layer parameter mode is used, different geometric regions within the same ply can use a set of candidate process parameters corresponding to that ply, and the candidate process parameters between different plies can be different. If a joint parameter mode is used, then the... Layer Each geometric region can have independent candidate process parameters, thus enabling differentiated configurations in both spatial and layer directions.
[0056] In one implementation, the feasible range of candidate process parameters can be set according to the process window of the automated fiber placement equipment. For example, the candidate placement speed can range from 0.5 m / s to 1 m / s, the candidate placement pressure can range from 300 N to 1200 N, the candidate placement tension can range from 5 N to 15 N, and the candidate placement temperature can range from 30°C to 80°C. In some applications, in hypercurvature regions and in layups near the surface, the candidate placement speed can be set to a relatively low candidate value to enhance the fit of the fiber bundle to the die surface; in flat regions and intermediate layers, a relatively high candidate placement speed can be set to improve manufacturing efficiency. In hypercurvature regions, the candidate placement pressure can be set to a relatively high candidate value to enhance interlayer compaction; in flat regions, a relatively low candidate placement pressure can be set. In hypercurvature regions, the candidate placement tension can be set to a relatively low candidate value to reduce internal stress in the fiber bundle at sharp bends; in flat regions, a relatively high candidate placement tension can be set to ensure fiber orientation accuracy. In the hypercurvature region, the candidate layup temperature can be set to a relatively high value to reduce resin viscosity and improve spreadability; in the flat region, a relatively low candidate layup temperature can be set. These candidate value configurations can serve as a source of candidate parameters for generating the two-dimensional feature tensor, and the specific values can be determined based on equipment capabilities, material system, and process documentation.
[0057] By generating candidate process parameters using different target optimization modes, the dimensionality of the candidate parameters corresponding to the two-dimensional feature tensor can be matched with the complexity of the component. For example, the global parameter mode has a lower dimensionality of candidate parameters, which is suitable for situations where parameter configuration requirements are relatively uniform; the joint parameter mode has a higher dimensionality of candidate parameters, which is suitable for situations where complex surfaces and multi-layer laying require differentiated configurations. In other words, when constructing several sets of two-dimensional feature tensors, several sets of candidate process parameters can be formed according to the actual selected target optimization mode, and these parameters can be filled into the two-dimensional tensor framework accordingly.
[0058] As can be seen, this scheme determines the global parameter mode, partition parameter mode, layer-by-layer parameter mode or joint parameter mode based on the number of layers and curvature characteristics. This allows the configuration granularity of candidate process parameters to be adapted to the geometric complexity and layer structure of the component to be optimized, thereby improving the flexibility of the subsequent determination of target process parameters.
[0059] In some embodiments, step 103 above may include the following steps: First, by using a pre-defined weighted reorganization corresponding to several defect categories, the features in each two-dimensional feature tensor are weighted to obtain the optimized feature tensor of each two-dimensional feature tensor with respect to each defect category.
[0060] The defect categories can include at least two of the following: bubbles, wrinkles, overlaps, twists, and gaps. Different defect categories have different formation mechanisms, and therefore their sensitivity to the features within the two-dimensional feature tensor can also differ. Based on the causes of each defect category, a set of preset weighted reassemblies can be configured for each defect category, and these preset weighted reassemblies can be used to weight the features within the two-dimensional feature tensor. By enhancing process parameters, material properties, and geometric coupling features that are more strongly correlated with different defect causes, and weakening auxiliary features with weaker correlations, optimized feature tensors for different defect categories can be obtained. The optimized feature tensor can still retain the two-dimensional structure formed by the ply number and geometric region number, but the feature values have been adjusted according to the weights of the corresponding defect categories, enabling subsequent defect prediction models to read input data that better matches the characteristics of the defect category's causes.
[0061] The weights of some features in the two-dimensional feature tensor can be found in Table 1: Table 1: Weights of certain features in predicting five types of defects
[0062] In Table 1 above, rows represent different features, and columns represent different defect categories.
[0063] Secondly, for each defect category, the defect prediction model corresponding to the defect category is used to process each optimized feature tensor corresponding to the defect category to obtain the predicted value of each geometric region in each ply of the component to be optimized regarding the defect category.
[0064] Then, for each two-dimensional feature tensor, the predicted value of each geometric region within each ply for each defect category is statistically analyzed to obtain the defect prediction tensor, and the defect prediction tensor is used as the defect prediction status corresponding to the two-dimensional feature tensor.
[0065] For example, after obtaining the predicted value corresponding to each defect category, statistics can be performed according to the ply number and geometric region number to obtain the defect prediction tensor. For the th Class defects can be used to obtain defect prediction tensors. ,in, Indicates the total number of plies. This represents the total number of geometric regions in each layer. If five types of defects are included—bubbles, wrinkles, overlaps, twists, and gaps—then five defect prediction tensors can be obtained. , =1,2,3,4,5. This defect prediction tensor can be used as the defect prediction result of the corresponding two-dimensional feature tensor to represent the defect prediction distribution of the component to be optimized under different plies and different geometric regions under this set of candidate process parameters.
[0066] In practice, the original training data can be cleaned first to remove abnormal samples caused by equipment malfunctions or recording errors. Furthermore, the features of each process parameter can be normalized to the minimum and maximum values, mapping the feature values to the [0,1] interval to reduce the impact of dimensional differences on model training. The processed dataset can also be divided into training and testing sets according to a preset ratio to train and validate each defect prediction model. Through these processes, each defect prediction model can be trained based on optimized feature tensor samples corresponding to the corresponding defect category, and can predict the two-dimensional feature tensors corresponding to candidate process parameters during the process parameter determination stage.
[0067] In some embodiments, step 104 above may include the following steps: First, obtain the layer weights and spatial weights of each geometric region within each ply.
[0068] Layer weights can be used to reflect the differences in the influence of ply along the thickness direction on defect propagation or component performance, while spatial weights can be used to reflect the differences in the influence of different geometric regions on defect formation or overall quality evaluation.
[0069] Optionally, the layer weights can be implemented using an increasing, decreasing, or intermediate weighted function. An increasing function assigns a larger weight to plies closer to the surface; a decreasing function assigns a larger weight to plies closer to the bottom layer; and an intermediate weighted function assigns a larger weight to intermediate layers. Spatial weights can be implemented using an equal-weighting or area-weighting model. An equal-weighting model assigns the same weight to each geometric region in the objective function, while an area-weighting model determines the spatial weight based on the proportion of the geometric region's area to the total area of the current ply region. The specific weighting method used can be configured based on the service requirements of the component to be optimized, the focus of quality considerations, and production experience.
[0070] Secondly, based on the layer weights and spatial weights of each geometric region within each ply, the predicted values corresponding to each defect category in the defect prediction tensor are weighted and summed to obtain the objective function value corresponding to each defect category.
[0071] For each set of two-dimensional feature tensors, its corresponding defect prediction tensor can be read. If the defect categories include bubbles, wrinkles, overlaps, twists, and gaps, the predicted values for each of the five defect categories can be weighted and summed to obtain five objective function values. For the k-th defect category, its objective function value can be expressed as: in, Indicates the first corresponding to the candidate process parameter The objective function value of the class defect. Indicates the total number of plies. This indicates the total number of geometric regions in each layer. Indicates the first Layer weights, Indicates the first Spatial weights of geometric regions Indicates the first Class defects in the first The first layer The predicted values for each geometric region. Using this formula, the predicted values of local defects in each ply and each geometric region can be aggregated into the objective function value corresponding to the defect category.
[0072] In one implementation, if the spatial weights adopt an area weighting model, then regions with larger areas can have an influence in the overall objective function that matches their area.
[0073] Then, based on the objective function values corresponding to each two-dimensional feature tensor, the target process parameters of the component to be optimized are determined from the candidate process parameters corresponding to each two-dimensional feature tensor.
[0074] Each set of two-dimensional feature tensors corresponds to a set of candidate process parameters, and multiple objective function values can be obtained for each set of candidate process parameters. For the five types of defects—bubbles, wrinkles, overlaps, twists, and gaps—multiple objective functions can be formed. The form of the multiple objective function can be: in, Indicates candidate process parameters The corresponding set of objective function values, This represents the objective function value about the bubble. This represents the objective function value related to the folds. This represents the objective function values with respect to overlap. This represents the objective function value with respect to torsion. This represents the objective function value regarding the gap. The above objective typically aims to minimize the number or severity of defects. Candidate process parameters. This can include layup speed, layup pressure, layup tension, and layup temperature. For different target optimization modes, candidate process parameters can correspond to global parameters, partitioned parameters, layer-by-layer parameters, or combined parameters. Global parameters can be expressed as... The partitioning parameters can be expressed as The layer-by-layer parameters can be expressed as The joint parameters can be expressed as ,in, Indicates the number of floors. This indicates the number of areas on each floor.
[0075] In some embodiments, obtaining the layer weights and spatial weights of each geometric region within each ply may include the following steps: First, for each geometric region within each ply, the layer weight of each geometric region within the ply is determined based on the position of the ply along the ply thickness direction.
[0076] Layer weights can be determined based on the position of the ply along its thickness. Since defects in automated fiber placement components may propagate, accumulate, or be covered by subsequent ply layers during the stacking process, different ply layers can have varying impacts on the final quality evaluation. If the current project prioritizes the fit between the bottom layer and the mold, a larger layer weight can be assigned to the bottom layer; if the focus is on the surface quality of the outer layers or the quality of subsequent assembly interfaces, a larger layer weight can be assigned to the surface layers; if the intermediate layers are considered to contribute significantly to overall performance based on component stress and ply structure, an intermediate-weighted layer weighting approach can be used.
[0077] In one implementation, an intermediate weighted function can be used to determine the layer weights, reflecting engineering experience that intermediate layers have a significant impact on the overall component performance. Layer-wise weighting of ply It can be represented as: ,in, =1,2…… ,in, In this embodiment, the total number of layers is [number]. It can take the value 128.
[0078] Secondly, for each geometric region within each ply, the spatial weight of the geometric region is determined based on the ratio between the area of the geometric region and the total area of the ply in which the geometric region is located.
[0079] No. Spatial weight of each geometric region It can be represented as: in, Indicates the first The area of each geometric region This represents the total area of the layer containing the geometric region. Using an area-weighted model, different geometric regions can participate in the aggregation of the objective function according to their area proportion.
[0080] In one implementation, the area of a geometric region can be obtained by accumulating the triangular facets of the triangular meshed three-dimensional geometric model. For each geometric region, the total area of the triangular facets it contains can be calculated as the area of that geometric region.
[0081] In this embodiment, the layer weight is determined based on the position of the ply in the thickness direction, and the spatial weight is determined based on the area ratio of the geometric region. This allows the objective function construction process to simultaneously reflect the differences in ply position and geometric region area.
[0082] In some embodiments, the method for determining the target process parameters of the component to be optimized from the candidate process parameters corresponding to each two-dimensional feature tensor based on the objective function values corresponding to each two-dimensional feature tensor can be as follows: First, based on the objective function values corresponding to each two-dimensional feature tensor, the dominance relationship of the candidate process parameters corresponding to each two-dimensional feature tensor is determined, and the set of non-dominated candidate parameters, i.e., the Pareto optimal solution set, is obtained.
[0083] The candidate process parameters in this solution set can form different trade-offs among multiple defect categories.
[0084] A non-dominated sorting genetic algorithm (NSGA-II) can be used to solve the multi-objective optimization function. This algorithm can generate a candidate population within the feasible region of process parameters and maintain the quality and distribution of candidate solutions through fast non-dominated sorting and crowding distance. When laying speed, laying pressure, laying tension, and laying temperature are used as optimization variables, a feasible region can be set. For example, the range of laying speed can be 0.5 m / s to 1 m / s with a step size of 0.01; the range of laying pressure can be 300 N to 1200 N with a step size of 10; the range of laying tension can be 5 N to 15 N with a step size of 0.1; and the range of laying temperature can be 30℃ to 80℃ with a step size of 1. In each generation of evolution, the process parameters corresponding to the candidate solutions can be substituted into the two-dimensional feature tensor construction process and the defect prediction process to calculate multiple objective function values, and the population is updated based on the dominance relationship. The population size can be set to 200, and the maximum number of iterations can be set to 500, or adjusted according to the complexity of the components and computing resources.
[0085] The non-dominated sorting genetic algorithm can include population initialization, individual evaluation, fast non-dominated sorting, crowding distance calculation, parent selection, crossover and mutation, environment selection and population update, and iteration termination. Population initialization can generate initial candidate process parameters within the feasible domain of process parameters; individual evaluation can call the defect prediction tensor and objective function construction process to calculate the values of each objective function; fast non-dominated sorting can divide different non-dominated levels according to the dominance relationship; crowding distance calculation can evaluate the distribution density of candidate solutions within the same non-dominated level; parent selection can adopt a binary tournament mechanism; crossover and mutation can adopt simulated binary crossover and polynomial mutation; environment selection and population update can prioritize retaining individuals with lower non-dominated levels and larger crowding distances; after the iteration reaches the maximum number of iterations or meets the convergence condition, a set of non-dominated candidate parameters can be output.
[0086] Secondly, the objective function values corresponding to each candidate process parameter in the non-dominated candidate parameter set are normalized, and the evaluation weights corresponding to each defect category are determined based on the results of the normalization process.
[0087] Normalization can be performed using the minimum-maximum normalization method to map each objective function value to a preset interval.
[0088] When determining the evaluation weights corresponding to each defect category, the entropy weight method can be used. The entropy weight method can determine the objective weights based on the dispersion of each objective function value in the non-dominated candidate parameter set.
[0089] Then, based on the evaluation weights, a comprehensive evaluation is performed on each group of candidate process parameters in the non-dominated candidate parameter set to obtain the evaluation results corresponding to each group of candidate process parameters.
[0090] The candidate process parameters in the non-dominated candidate parameter set can be comprehensively evaluated based on the evaluation weights of each defect category. The comprehensive evaluation can employ the entropy weight and superior-inferior solution distance method (Technique for Order Preference by Similarity to Ideal Solution, TOPSIS). This method first determines the positive and negative ideal solutions based on the normalized objective function values and evaluation weights, and then calculates the weighted Euclidean distance from each candidate process parameter to the positive and negative ideal solutions.
[0091] For the defect minimization objective, a positive ideal value corresponds to a smaller normalized objective function value, while a negative ideal value corresponds to a larger normalized objective function value.
[0092] Then, based on the evaluation results, the target process parameters are determined from the set of non-dominated candidate parameters.
[0093] Candidate process parameters in the non-dominated candidate parameter set can be sorted based on relative proximity or weighted comprehensive score, and the candidate process parameter whose evaluation result meets preset conditions can be selected as the target process parameter. For example, when using a combination of entropy weight and TOPSIS, the candidate process parameter with the highest relative proximity can be selected as the target process parameter; when using a preference-based weighting method, the candidate process parameter with the best weighted comprehensive score can be selected as the target process parameter. The target process parameter can include the layup speed, layup pressure, layup tension, and layup temperature in the corresponding optimization mode. When using a joint parameter mode, the target process parameter can include a set of parameters corresponding to each geometric region of each ply; when using a partitioned parameter mode or a layer-by-layer parameter mode, the target process parameter can correspond to a set of parameters for each geometric region or each ply; when using a global parameter mode, the target process parameter can correspond to a set of general parameters for the entire component.
[0094] Ultimately, the target process parameters can be output to the automated fiber placement equipment for production, or they can be output to the operator for confirmation first. The automated fiber placement equipment can perform variable parameter placement according to the target process parameters, allowing different geometric regions and different layup positions to use placement speed, placement pressure, placement tension, and placement temperature that match their process requirements. For complex surface components, this method can achieve spatial and layer-wise differentiated process parameter configurations within an overall optimization framework.
[0095] As can be seen, this scheme obtains a set of non-dominated candidate parameters by judging the dominance relationship of candidate process parameters, and determines the target process parameters by combining normalization, evaluation weights and comprehensive evaluation. It can make compromises and screen among multiple defect categories, thereby improving the adaptability of the target process parameters to the requirements of multi-defect collaborative control.
[0096] Please see Figure 4 The automatic wire placement process optimization method considering mold geometry provided by the present invention may include the following steps: The filament-laying experimental data were classified, and spatial-layer dual-dimensional feature tensors were constructed. For bubbles and gaps, the corresponding feature tensors were expanded into tabular input data according to the positions of the layup and geometric regions, and the XGBoost algorithm was used to construct bubble prediction models and gap prediction models, respectively. For wrinkles, overlaps, and twists, the spatial-layer structure of the feature tensors was preserved, and the CNN algorithm was used to construct wrinkle prediction models, overlap prediction models, and twist prediction models, respectively. During the construction of each defect prediction model, corresponding feature weights were set according to the sensitivity of different defects to each input feature, thus obtaining five independently trained defect prediction models.
[0097] Furthermore, a multi-objective optimization function with spatial-layer dual-dimensional weighted aggregation is constructed, and the value range and optimization mode of the process parameters to be optimized are determined. Population initialization is performed within the feasible region of the process parameters. The candidate process parameters corresponding to each individual in the population are used to construct the corresponding dual-dimensional feature tensor, and the aforementioned defect prediction model is called to obtain the prediction results for each defect. Then, the objective function value corresponding to each individual is calculated. When the preset termination condition is not met, the population is selected, crossovered, and mutated to generate the next generation population, and the defect prediction and objective function value calculation are repeated. When the preset termination condition is met, the iteration ends and the Pareto optimal solution set is output for subsequent determination of the target process parameters.
[0098] The automatic wire placement process optimization device considering mold geometry features provided by the present invention will be described below. The automatic wire placement process optimization device considering mold geometry features described below can be referred to in correspondence with the automatic wire placement process optimization method considering mold geometry features described above. Figure 5 As shown, the automated wire placement process optimization device 400, which considers the geometry of the mold, includes the following modules: The data acquisition module 401 is used to acquire the three-dimensional geometric model of the mold, the material property information of the component to be optimized, and the layup information; The feature tensor construction module 402 is used to construct several sets of two-dimensional feature tensors using the three-dimensional geometric model, the material property information, and the ply information. The features in each set of two-dimensional feature tensors include candidate process parameters for each geometric region in each ply of the component to be optimized, geometric feature vectors determined based on the three-dimensional geometric model of the mold, material property features determined based on the material property information, and layer orientation features determined based on the ply information. The defect prediction module 403 is used to perform defect prediction on each set of two-dimensional feature tensors using several pre-built defect prediction models, and to obtain the prediction of the component to be optimized for each defect category corresponding to each two-dimensional feature tensor. The defect categories that different defect prediction models can predict are different and each defect prediction model is trained independently. The process parameter determination module 404 is used to construct a multi-objective optimization function and, based on the multi-objective optimization function and the prediction of the component to be optimized for each defect category corresponding to each of the two-dimensional feature tensors, determine the target process parameters corresponding to the target optimization mode. The target optimization mode is obtained based on the three-dimensional geometric model and the layup information.
[0099] The implementation methods of the above modules can refer to the above-mentioned automatic wire placement process optimization method that takes into account the geometric features of the mold, which will not be elaborated here.
Claims
1. A method for automatic fiber placement process optimization considering mold geometry features, characterized in that, include: Obtain the three-dimensional geometric model of the mold, the material properties of the component to be optimized, and the layup information; Using the three-dimensional geometric model, the material property information, and the ply information, several sets of two-dimensional feature tensors are constructed. The features in each set of two-dimensional feature tensors include candidate process parameters for each geometric region in each ply of the component to be optimized, geometric feature vectors determined based on the three-dimensional geometric model of the mold, material property features determined based on the material property information, and layer orientation features determined based on the ply information. Several pre-built defect prediction models are used to predict defects for each set of two-dimensional feature tensors, so as to obtain the prediction of the component to be optimized for each defect category corresponding to each two-dimensional feature tensor. The defect categories that different defect prediction models can predict are different and each defect prediction model is trained independently. A multi-objective optimization function is constructed, and based on the multi-objective optimization function and the prediction of the component to be optimized for each defect category corresponding to each of the two-dimensional feature tensors, the target process parameters corresponding to the target optimization mode are determined. The target optimization mode is obtained based on the three-dimensional geometric model and the layup information.
2. The method of claim 1, wherein, Using the three-dimensional geometric model, the material property information, and the ply information, several sets of two-dimensional feature tensors are constructed, including: Based on the curvature characteristics at multiple locations in the three-dimensional geometric model, the three-dimensional geometric model is divided into regions to obtain multiple geometric regions. A two-dimensional tensor framework is established based on the ply sequence number in the ply information and the region number of the multiple geometric regions. For each of the two-dimensional feature tensors, a set of candidate process parameters, geometric feature vectors, material property features, and layer orientation features of each geometric region within each layup are filled into the two-dimensional tensor framework to obtain the two-dimensional feature tensor.
3. The method of claim 2, wherein, For each geometric region within each of the aforementioned ply: The candidate process parameters include candidate layup speed, candidate layup pressure, candidate layup tension, and candidate layup temperature; The layer orientation features include the ply angle and normalized ply number of the ply; The material properties include fiber volume fraction, resin viscosity, prepreg viscosity coefficient, glass transition temperature, and single-layer thickness. The geometric feature vector includes the local mean curvature, the local Gaussian curvature, and the angle between the ply direction and the principal curvature direction.
4. The method according to claim 3, characterized in that, For each geometric region within each of the aforementioned ply, the two-dimensional feature tensor further includes: The velocity-curvature coupling feature is determined based on the candidate laying speed and the local average curvature. The velocity-curvature coupling feature is... , wherein Indicates the first Layer, First Candidate tile placement speed for each geometric region Indicates the first Layer, First Local average curvature of a geometric region; The pressure and curvature compensation features are determined based on the candidate layup pressure and the local Gaussian curvature, and these features are expressed as follows: / (1+ ), wherein, the Indicates the first Layer, First The candidate layup pressure of each geometric region, the Indicates the first Layer, First Local Gaussian curvature of a geometric region; The tension-curvature coupling characteristics, determined based on the candidate layup tension and the local Gaussian curvature, are expressed as follows: , wherein Indicates the first Layer, First The candidate layup tension of a geometric region, the Indicates the first Layer, First Local Gaussian curvature of a geometric region; The temperature-velocity coupling characteristic determined based on the candidate placement temperature and the candidate placement velocity is expressed as follows: , Indicates the first Layer, First The candidate layup temperature for each geometric region, the Indicates the first Layer, First The candidate tile placement speed for each geometric region.
5. The method of claim 2, wherein, The methods for determining the target optimization mode include: Based on the number of layers in the ply information and the curvature features at multiple locations in the three-dimensional geometric model, a target optimization mode is determined. The target optimization mode is a global parameter mode, a partition parameter mode, a layer-by-layer parameter mode, or a joint parameter mode. The methods for determining candidate process parameters in each of the two-dimensional feature tensors include: Based on the target optimization mode, candidate process parameters for each geometric region within each layup are determined, wherein the dimensions of the candidate process parameters corresponding to different target optimization modes are different.
6. The method according to any one of claims 1 to 5, characterized in that, Defect prediction is performed on each set of two-dimensional feature tensors using several pre-built defect prediction models, resulting in the prediction of the component to be optimized for each defect category, including: By using a preset weighted recombination corresponding to several defect categories, the features in each of the two-dimensional feature tensors are weighted to obtain the optimized feature tensor of each of the two-dimensional feature tensors with respect to each defect category. For each defect category, the defect prediction model corresponding to the defect category is used to process each optimized feature tensor corresponding to the defect category to obtain the predicted value of each geometric region in each ply of the component to be optimized with respect to the defect category. The defect prediction model corresponding to the defect category is trained using the corresponding optimized feature tensor samples and their predicted value labels. For each of the two-dimensional feature tensors, the predicted values of each geometric region within each of the ply are statistically analyzed for each defect category to obtain a defect prediction tensor, and the defect prediction tensor is used as the defect prediction status corresponding to the two-dimensional feature tensor. Among them, multiple defect prediction models include bubble prediction model, wrinkle prediction model, overlap prediction model, torsion prediction model and gap prediction model.
7. The method of claim 6, wherein, Based on the multi-objective optimization function and the predictions of the components to be optimized for each defect category corresponding to each of the two-dimensional feature tensors, the target process parameters corresponding to the target optimization mode are determined, including: Obtain the layer weight and spatial weight of each geometric region within each of the plies; Based on the layer weights and spatial weights of each geometric region within each ply, the predicted values corresponding to each defect category in the defect prediction tensor are weighted and summed to obtain the objective function value corresponding to each defect category. Based on the objective function values corresponding to each of the two-dimensional feature tensors, the target process parameters of the component to be optimized are determined from the candidate process parameters corresponding to each of the two-dimensional feature tensors.
8. The method of claim 7, wherein, Obtain the layer weights and spatial weights of each geometric region within each of the aforementioned plies, including: For each geometric region within each of the aforementioned ply: The layer weight of each geometric region within the ply is determined based on the position of the ply along the thickness direction. The spatial weight of the geometric region is determined based on the ratio between the area of the geometric region and the total area of the layer in which the geometric region is located.
9. The method of claim 7, wherein, Based on the objective function values corresponding to each of the two-dimensional feature tensors, the target process parameters of the component to be optimized are determined from the candidate process parameters corresponding to each of the two-dimensional feature tensors, including: Based on the objective function values corresponding to each of the two-dimensional feature tensors, the dominance relationship of the candidate process parameters corresponding to each of the two-dimensional feature tensors is determined to obtain a set of non-dominated candidate parameters. The objective function values corresponding to each candidate process parameter in the non-dominated candidate parameter set are normalized, and the evaluation weights corresponding to each defect category are determined based on the results of the normalization process. Based on the evaluation weights, a comprehensive evaluation is performed on each group of candidate process parameters in the non-dominated candidate parameter set to obtain the evaluation results corresponding to each group of candidate process parameters. Based on the evaluation results, the target process parameters are determined from the set of non-dominated candidate parameters.
10. An automated fiber placement process optimization device that takes into account mold geometry features, characterized in that, include: The data acquisition module is used to acquire the three-dimensional geometric model of the mold, the material property information of the component to be optimized, and the layup information; The feature tensor construction module is used to construct several sets of two-dimensional feature tensors using the three-dimensional geometric model, the material property information, and the ply information. The features in each set of two-dimensional feature tensors include candidate process parameters for each geometric region in each ply of the component to be optimized, geometric feature vectors determined based on the three-dimensional geometric model of the mold, material property features determined based on the material property information, and layer orientation features determined based on the ply information. The defect prediction module is used to predict defects for each set of two-dimensional feature tensors using several pre-built defect prediction models, and to obtain the prediction of the component to be optimized for each defect category for each two-dimensional feature tensor. The defect categories that different defect prediction models can predict are different and each defect prediction model is trained independently. The process parameter determination module is used to construct a multi-objective optimization function and, based on the multi-objective optimization function and the prediction of the component to be optimized for each defect category corresponding to each of the two-dimensional feature tensors, determine the target process parameters corresponding to the target optimization mode. The target optimization mode is obtained based on the three-dimensional geometric model and the layup information.
