Optimization Method, System, Medium and Filament Winding Machine for Automatic Filament Winding Forming of Composite Materials

Abstract

The present application relates to the technical field of filament winding control, and discloses an optimization method, system, medium and filament winding machine for automatic filament winding forming of composite materials. The method includes: inputting dynamic tension control data into a fuzzy neural network calculation engine to obtain a take-up and pay-off speed dynamic compensation model; based on the take-up and pay-off speed dynamic compensation model, using a reinforcement learning engine to calculate the multi-filament bundle shearing sequence and generate an intelligent shearing instruction set; constructing a re-feed trajectory mapping relationship based on the intelligent shearing instruction set, and forming a re-feed path matrix after iterative calculation by a digital twin engine; establishing a temperature field distribution prediction model according to the re-feed path matrix, and constructing a temperature field dynamic control sequence after finite element analysis calculation; substituting the temperature field dynamic control sequence into a multi-objective constraint equation, and outputting a laying strategy data packet after iterative optimization by deep reinforcement learning. The present application realizes the precise coordination of multiple process parameters such as tension control, speed compensation, shearing control, and temperature control.

Description

Technical Field

[0001] This application relates to the field of filament winding control, and particularly to an optimization method, system, medium and filament winding machine for automatic filament winding forming of composite materials. Background Art

[0002] The automatic filament winding forming technology of composite materials has been widely used in the fields of aerospace, transportation, etc. The traditional automatic filament winding technology of composite materials mainly adopts manual operation mode, that is, the carbon fiber cloth is laid on the mold surface layer by layer by the operator, and then cured and formed. With the improvement of industrial automation level, an automatic filament winding system based on industrial robots has emerged, and the forming of carbon fiber composite structural parts is realized by controlling the filament winding head with a robotic arm. This technology controls a multi-degree-of-freedom robotic arm through a numerical control system, lays the prepreg tow on the mold surface according to a preset trajectory, and completes the laying process through processes such as heating and compaction.

[0003] However, there are some obvious deficiencies in the existing automatic filament winding technology of composite materials. First, when encountering molds with large curvatures or special shapes, traditional control strategies are difficult to ensure the laying quality, and problems such as wrinkles, overlaps or excessive gaps are likely to occur. Second, the coordination of the existing technology in key links such as tension control, shear positioning, and temperature control is insufficient, and it is difficult to achieve the precise coordination of various process parameters, affecting the forming quality. In addition, the optimization of traditional control methods for process parameters mainly relies on empirical settings, lacking intelligent adaptive adjustment capabilities, and it is difficult to meet the requirements of high-quality and high-efficiency production. Summary of the Invention

[0004] This application solves the problem of insufficient precision in multi-parameter coordinated control existing in the existing automatic filament winding technology of composite materials, and realizes the precise coordination of multiple process parameters such as tension control, speed compensation, shear control, and temperature control by establishing an intelligent control system based on deep learning.

[0005] In a first aspect, the present application provides an optimization method for automatic fiber placement forming of composite materials. The optimization method for automatic fiber placement forming of composite materials includes: the tension data of the prepreg tow input by a tension sensor undergoes LSTM deep learning operations to output dynamic tension control data; the dynamic tension control data is input into a fuzzy neural network calculation engine, and after speed curve analysis operations, a dynamic compensation model for the winding and unwinding speed is obtained; based on the dynamic compensation model for the winding and unwinding speed, a reinforcement learning engine is used to calculate the cutting sequence of multiple tows to generate an intelligent cutting instruction set; based on the intelligent cutting instruction set, a mapping relationship of the refeeding trajectory is constructed, and after iterative calculations by a digital twin engine, a refeeding path matrix is formed; a temperature field distribution prediction model is established based on the refeeding path matrix, and after finite element analysis calculations, a dynamic control sequence for the temperature field is constructed; the dynamic control sequence for the temperature field is substituted into a multi-objective constraint equation, and after iterative optimization by deep reinforcement learning, a laying strategy data packet is output.

[0006] In a second aspect, the present application provides an optimization system for automatic fiber placement forming of composite materials. The optimization system for automatic fiber placement forming of composite materials includes:

[0007] An input module, configured to input the tension data of the prepreg tow by a tension sensor, undergo LSTM deep learning operations, and output dynamic tension control data;

[0008] An operation module, configured to input the dynamic tension control data into a fuzzy neural network calculation engine, and after speed curve analysis operations, obtain a dynamic compensation model for the winding and unwinding speed;

[0009] A calculation module, configured to calculate the cutting sequence of multiple tows based on the dynamic compensation model for the winding and unwinding speed by using a reinforcement learning engine to generate an intelligent cutting instruction set;

[0010] A construction module, configured to construct a mapping relationship of the refeeding trajectory based on the intelligent cutting instruction set, and after iterative calculations by a digital twin engine, form a refeeding path matrix;

[0011] An establishment module, configured to establish a temperature field distribution prediction model based on the refeeding path matrix, and after finite element analysis calculations, construct a dynamic control sequence for the temperature field;

[0012] An output module, configured to substitute the dynamic control sequence for the temperature field into a multi-objective constraint equation, and after iterative optimization by deep reinforcement learning, output a laying strategy data packet.

[0013] In a third aspect of the present application, there is provided a computer-readable storage medium, in which instructions are stored, and when they run on a computer, cause the computer to execute the above-mentioned optimization method for automatic fiber placement forming of composite materials.

[0014] Fourthly, the present invention further provides a fiber placement machine, which comprises:

[0015] one or more processors;

[0016] a storage device for storing one or more programs;

[0017] When the one or more programs are executed by the one or more processors, the one or more processors implement the composite material automatic fiber placement forming optimization method as described above.

[0018] In the technical solution provided by this application, the tension data of the prepreg tow is processed through LSTM deep learning, and the precise analysis of the speed curve is realized by combining with the fuzzy neural network computing engine, improving the accuracy and stability of tension control. The reinforcement learning engine optimizes the calculation of the multi-tow shearing sequence, and the precise mapping of the refeeding trajectory is realized through the iterative calculation of the digital twin engine, greatly improving the shearing positioning accuracy and refeeding reliability. The temperature field distribution prediction model combines with the finite element analysis technology to realize the dynamic and precise control of the temperature field, ensuring the temperature uniformity and stability during the laying process. Through the multi-objective constraint optimization of the temperature field dynamic control sequence by deep reinforcement learning, the global optimization of the laying strategy is realized, significantly improving the quality stability of the fiber placement process. The entire solution realizes the collaborative optimization of multiple key parameters such as tension control, speed compensation, shearing positioning, and temperature control during the fiber placement process through multi-level data processing and intelligent algorithm optimization, effectively solving the problems of improper parameter coordination and process instability existing in traditional fiber placement technologies. At the same time, the intelligent optimization method based on deep learning and reinforcement learning enables the control strategy to be adaptively adjusted to meet the fiber placement requirements under different working conditions, improving the flexibility and intelligence level of the production process. In addition, the introduction of digital twin technology makes the laying trajectory planning more precise and the temperature field prediction more accurate, providing a reliable technical support for high-quality fiber placement manufacturing. Through multi-objective constraint optimization, while ensuring the product quality, the energy efficiency is improved and the production cost is reduced, having significant economic benefits and practical value. Description of the Drawings

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0020] Figure 1 It is a schematic diagram of an embodiment of the composite material automatic fiber placement forming optimization method in the embodiments of this application;

[0021] Figure 2It is a schematic diagram of an embodiment of the composite material automatic fiber placement forming optimization system in the embodiments of the present application.

[0022] Figure 3 It is the first structural schematic diagram of the fiber placement machine provided by the embodiments of the present application;

[0023] Figure 4 It is the second structural schematic diagram of the fiber placement machine provided by the embodiments of the present application. Detailed implementation manners

[0024] The embodiments of the present application provide a composite material automatic fiber placement forming optimization method, system, medium and fiber placement machine. Terms such as "first", "second", "third", "fourth", etc. (if any) in the specification, claims and the above drawings of the present application are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described here can be implemented in an order other than those illustrated or described here. In addition, the term "including" or "having" and any deformation thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those clearly listed steps or units, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0025] For ease of understanding, the following describes the specific process of the embodiments of the present application. Please refer to Figure 1 , an embodiment of the composite material automatic fiber placement forming optimization method in the embodiments of the present application includes:

[0026] Step S101: The tension data of the prepreg tow input by the tension sensor undergoes LSTM deep learning operation to output dynamic tension control data;

[0027] Step S102: Input the dynamic tension control data into the fuzzy neural network calculation engine. After speed curve analysis operation, a take-up and pay-off speed dynamic compensation model is obtained;

[0028] Step S103: Based on the take-up and pay-off speed dynamic compensation model, use the reinforcement learning engine to calculate the multi-tow cutting sequence and generate an intelligent cutting instruction set;

[0029] Step S104: Build a re-feed trajectory mapping relationship based on the intelligent cutting instruction set. After iterative calculation by the digital twin engine, a re-feed path matrix is formed;

[0030] Step S105: Establish a temperature field distribution prediction model according to the re-feed path matrix. After finite element analysis calculation, a temperature field dynamic control sequence is constructed;

[0031] Step S106: Substitute the temperature field dynamic control sequence into the multi-objective constraint equation. After iterative optimization by deep reinforcement learning, output the placement strategy data packet.

[0032] It can be understood that the execution subject of this application can be an optimization system for automatic fiber placement forming of composite materials, or a terminal or a server. Specifically, it is not limited here. In this embodiment of the application, the server is taken as the execution subject for illustration.

[0033] Specifically, the tension data of the prepreg tow collected by the tension sensor is input into the system. The tension sensor is arranged on the prepreg tow conveying path to collect the tension change data of the tow in real time, and the sampling frequency is 1000 Hz. The collected tension data is decomposed by wavelet transform at multiple scales to separate the tension fluctuation characteristics of different frequency bands, and the noise interference is eliminated by singular value decomposition filtering. The processed tension data is input into a three-layer LSTM network structure with 64 hidden neurons. The LSTM network learns the temporal characteristics of the tension data through the input gate, forget gate, and output gate, and captures the long-term dependence relationship of the tension change. The obtained dynamic tension control data is then input into the fuzzy neural network engine. The fuzzy neural network first compresses and reduces the dimension of the tension data through an autoencoder to reduce the data redundancy. The compressed feature data is fuzzily quantified through a triangular membership function to establish a fuzzy rule base. The fuzzy rule base is optimized by a genetic algorithm to select the optimal rule set. The optimized rule set is input into an improved BP neural network for online learning to obtain the speed compensation feature vector. The speed compensation feature vector extracts the speed curve features through a convolutional neural network, and finally outputs the take-up and pay-off speed dynamic compensation model.

[0034] Based on the take-up and pay-off speed dynamic compensation model, the state space of the shearing process is encoded by the PPO algorithm. The state space includes information such as the current tow position, tension, and speed, and is mapped through a multi-layer perceptron to generate the representation of the shearing action space. The shearing action space evaluates the value function of different shearing actions through a double Q network, and combines the Monte Carlo tree search strategy to output the shearing state transition probability matrix. Temporal difference learning is performed on the probability matrix to construct the shearing policy gradient vector. The policy gradient vector is input into the Actor-Critic network for optimization and update to form a multi-tow shearing decision sequence. Based on the intelligent shearing instruction set, the spatio-temporal features are extracted through a recursive neural tensor network to generate the re-feed trajectory feature point cloud. The feature point cloud is subjected to contrast learning through a siamese convolutional neural network, and combines with a three-dimensional space transformation network to output the trajectory encoding sequence. The encoding sequence is processed by a graph convolutional network to construct the re-feed trajectory topology map. The trajectory topology map is input into the physical simulation engine to calculate the re-feed dynamics parameters. The state estimation of the dynamics parameters is performed through a self-correcting particle filter algorithm, and finally the re-feed path matrix is formed.

[0035] According to the resending path matrix, the temperature field data is stratified by the three spatial dimensions of X, Y, and Z to generate a temperature data cube. The temperature data is downsampled and discretized to construct an array of temperature field node data. The node data undergoes thermodynamic feature extraction and energy conservation conversion to output a multi-dimensional feature data stream of temperature. The feature data stream is processed by spatio-temporal serialization to generate a four-dimensional evolution surface of the temperature field. The temperature gradient information is extracted from the evolution surface, and the key temperature control points are screened, and finally a dynamic control sequence of the temperature field is output. The dynamic control sequence of the temperature field is input into the multi-objective constraint optimization module. First, the control sequence is decoupled in multiple dimensions, and the feature vectors in the time dimension, spatial dimension, and temperature dimension are extracted respectively. The feature vectors are regularized, and the data is corrected according to the physical constraint conditions. Through multi-scale feature decomposition, the parameter change law is extracted, and an optimized feature space is constructed. Through the quantization process of the reward and punishment function, the weights of each parameter are allocated to form a multi-objective optimization data stream. Finally, through the strategy integration process, a laying strategy data packet is output.

[0036] Taking the manufacturing of a certain aviation composite component as an example: During the manufacturing process, the range of the initial tow tension data collected by the tension sensor is 5 - 15 N. After being processed by the LSTM network, the dynamic tension control accuracy is improved to ±0.2 N. After these tension data are input into the fuzzy neural network, the generated speed compensation model controls the retracting and feeding speed within the range of 10 - 50 mm / s, and the compensation accuracy reaches ±0.5 mm / s. Based on this compensation model, the reinforcement learning algorithm plans the optimal cutting sequence for 8 bundles of fibers, and the cutting positioning accuracy is controlled within the range of ±0.15 mm. The digital twin engine generates a resending trajectory according to the cutting instruction, and the trajectory positioning accuracy reaches ±0.1 mm. The temperature field control sequence maintains the heating temperature within the range of 350 ± 3 °C to ensure that the material is fully softened. The laying strategy data packet achieves a laying accuracy of ±0.2 mm, and the interlayer bonding strength reaches more than 95% of the required value. During the whole process, the collaborative optimization of various parameters significantly improves the product quality, and the production efficiency is increased by more than 3 times.

[0037] In the embodiments of the present application, the tension data of prepreg tows is processed through LSTM deep learning, and the precise analysis of the speed curve is realized by combining with a fuzzy neural network computing engine, improving the accuracy and stability of tension control. The reinforcement learning engine optimizes the calculation of the multi-tow shearing sequence, and realizes the precise mapping of the refeeding trajectory through the iterative calculation of the digital twin engine, greatly improving the shearing positioning accuracy and refeeding reliability. The temperature field distribution prediction model combines finite element analysis technology to realize the dynamic and precise control of the temperature field, ensuring the temperature uniformity and stability during the laying process. Through multi-objective constraint optimization of the temperature field dynamic control sequence by deep reinforcement learning, the global optimization of the laying strategy is realized, significantly improving the quality stability of the fiber placement process. The entire solution realizes the collaborative optimization of multiple key parameters such as tension control, speed compensation, shearing positioning, and temperature control during the fiber placement process through multi-level data processing and intelligent algorithm optimization, effectively solving the problems of improper parameter coordination and process instability existing in traditional fiber placement technologies. At the same time, the intelligent optimization method based on deep learning and reinforcement learning enables the control strategy to be adaptively adjusted to meet the fiber placement requirements under different working conditions, improving the flexibility and intelligence level of the production process. In addition, the introduction of digital twin technology makes the laying trajectory planning more accurate and the temperature field prediction more accurate, providing reliable technical support for high-quality fiber placement manufacturing. Through multi-objective constraint optimization, while ensuring product quality, it also realizes the improvement of energy efficiency and the reduction of production costs, with significant economic benefits and practical value.

[0038] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0039] (1) Perform multi-scale time-frequency decomposition on the tension data of prepreg tows through wavelet transform and Fourier analysis. After singular value decomposition filtering, generate a tension multi-dimensional spectrum feature matrix;

[0040] (2) Input the tension multi-dimensional spectrum feature matrix into a three-layer LSTM network structure with 64 hidden neurons. After forward recursive calculation and long short-term memory unit state update, output a set of tension spatio-temporal state vectors;

[0041] (3) According to the set of tension spatio-temporal state vectors, perform deep temporal correlation mining through a bidirectional gated recurrent unit based on the Transformer architecture, and combine with a residual connection mechanism to form a tension prediction deep feature map;

[0042] (4) Calculate the tension prediction deep feature map through a multi-head attention mechanism, fuse the self-attention and cross-attention weight distributions, and screen out a high-dimensional tension feature sequence;

[0043] (5) The high-dimensional tension feature sequence is iteratively optimized in batches through an improved stochastic gradient descent algorithm, and combined with L1 and L2 regularization constraints to construct a tension prediction loss function surface.

[0044] (6) Substitute the tension prediction loss function surface into the Adam optimizer with momentum term. Through adaptive learning rate adjustment and gradient clipping techniques, after batch normalization processing, dynamic tension control data is output.

[0045] Specifically, multi-scale time-frequency decomposition analysis is performed on the prepreg tow tension data. Wavelet transform is used to decompose the tension signal at different scales, and Fourier analysis converts the signal into the frequency domain space. After singular value decomposition filtering, a tension spectrum feature matrix containing multiple frequency band features is generated.

[0046] The calculation process of the multi-head attention mechanism can be expressed as:

[0047]

[0048] Where:

[0049] Q, K, and V represent the query matrix, key matrix, and value matrix respectively, h represents the number of attention heads, n represents the sequence length, d k represents the key vector dimension, α ( represents the weight coefficient of the i-th attention head, ψ(V + ) represents the non-linear transformation function of the value matrix, ω ( represents the output projection matrix of the i-th attention head, j represents the sequence position index being processed, m represents the summation index during normalization, T represents the matrix transpose operation, Q i represents the query vector of the i-th attention head, K j represents the key vector at the j-th position, K m represents the key vector at the m-th position, V j represents the value vector at the j-th position.

[0050] Extract the temporal characteristics of the tension data. When the sampling frequency is 1000 Hz, 1000 original tension data points are generated per second. The signal is decomposed into 4 scale levels through wavelet transform, and each level contains the characteristics of different frequency bands. The singular value decomposition filter sets the threshold to 0.1 to filter out the noise components with singular values less than the threshold. Next, the obtained multi-dimensional spectrum feature matrix of tension is input into a three-layer LSTM network for processing. The first layer of the LSTM network contains 64 hidden neurons for preliminary feature extraction; the second layer also contains 64 neurons for feature fusion; the 64 neurons in the third layer are responsible for generating the final temporal feature representation. Each LSTM cell contains an input gate, a forget gate, and an output gate, and controls the information flow through the gating mechanism to capture long-term dependencies.

[0051] The obtained set of tension spatio-temporal state vectors is then input into a bidirectional gated recurrent unit based on the Transformer architecture. The Transformer architecture performs deep feature extraction on the tension data through multiple layers of self-attention mechanisms and feed-forward neural networks. The bidirectional gated recurrent unit processes the sequence data in both the forward and backward directions to capture bidirectional temporal dependencies. The residual connection mechanism retains the original feature information through skip connections to alleviate the vanishing gradient problem in deep networks. The deep feature map for tension prediction is calculated through the multi-head attention mechanism to generate a richer feature representation. The multi-head attention mechanism projects the input features into multiple subspaces, calculates the attention scores separately, and then combines the results. The self-attention mechanism focuses on the correlations within the sequence, while the cross-attention mechanism focuses on the associations between different features. By allocating attention weights, important features are highlighted and irrelevant information is suppressed.

[0052] The obtained high-dimensional tension feature sequence is optimized through an improved stochastic gradient descent algorithm. The algorithm uses the mini-batch method, with each batch containing 128 samples, and the initial learning rate is set to 0.001. The L1 regularization constraint is used to generate a sparse solution, and the L2 regularization constraint is used to prevent overfitting. The regularization coefficients are set to 0.01 and 0.001 respectively. The gradient is calculated through backpropagation, the model parameters are iteratively updated, and the prediction loss function surface is constructed. Finally, the tension prediction loss function surface is input into the Adam optimizer. The Adam optimizer combines the advantages of the momentum method and RMSprop and has the ability to adaptively adjust the learning rate. The initial momentum decay rate is set to 0.9, and the second-order momentum decay rate is 0.999. The gradient clipping threshold is set to 5.0 to prevent gradient explosion. The batch normalization layer normalizes the features to accelerate the training convergence.

[0053] Taking the manufacturing of the wing skin of a large airliner as an example, the originally collected tension data fluctuates within the range of 5 - 15 N, with a sampling frequency of 1000 Hz, generating a large amount of noise interference. After being decomposed into 4 scale levels by wavelet transform, the high-frequency noise is effectively separated. Singular value decomposition filtering reduces the singular values from the original 127 to 43, retaining the main feature information. After being processed by the LSTM network, the tension prediction error is reduced from the initial ±0.5 N to ±0.08 N. After further processing by the multi-head attention mechanism (8 attention heads), the prediction lead is increased to 200 ms when it comes to mutation working conditions. Finally, through optimization by the Adam optimizer, the dynamic tension control error is stabilized within the range of ±0.05 N. The entire processing process realizes the noise reduction, feature extraction, time series prediction and precise control of the tension data, providing a reliable data basis for subsequent speed planning and shear control.

[0054] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0055] (1) Denoise and extract features from the dynamic tension control data through an autoencoder, and output a tension feature dimensionality reduction data set;

[0056] (2) Fuzzily quantify the tension feature dimensionality reduction data set through a triangular membership function, and through Mamdani inference mechanism operation, construct a fuzzy rule base matrix;

[0057] (3) Optimize and screen the rules from the fuzzy rule base matrix through a genetic algorithm, and fuse crossover and mutation operations to generate an optimal fuzzy inference rule set;

[0058] (4) Perform online learning calculation on the optimal fuzzy inference rule set through an improved BP neural network, and combine the Dropout random inactivation strategy to form a speed compensation feature vector;

[0059] (5) Input the speed compensation feature vector into a convolutional neural network for speed curve feature extraction, and through multi-layer pooling and fully connected operations, construct a speed control response sequence;

[0060] (6) Perform parameter optimization calculation on the speed control response sequence through a particle swarm optimization-based backpropagation algorithm, and through error feedback iteration, output a retraction speed dynamic compensation model.

[0061] Specifically, when the autoencoder processes the dynamic tension control data, the data is input into the encoder. The encoder consists of multiple layers of neural networks. The first layer contains 256 neurons, the second layer has 128 neurons, and finally the data is compressed into a 64-dimensional latent space. The decoder, on the contrary, reconstructs the data through a three-layer neural network and finally outputs data with the same dimension as the input. By minimizing the reconstruction error, the autoencoder learns the core features of the data while removing noise interference and outputs a tension feature dimensionality reduction dataset. The obtained tension feature dimensionality reduction dataset is then fuzzily quantified through a triangular membership function. The triangular membership function maps continuous numerical values into a fuzzy set, and each variable is divided into three fuzzy subsets: low, medium, and high. The Mamdani inference mechanism constructs IF-THEN rules based on these fuzzy subsets to form an initial fuzzy rule base. The antecedent part of the rule describes the state of the input variable, and the consequent part corresponds to the control output. The entire rule base is stored in matrix form. The fuzzy rule base matrix is optimized by a genetic algorithm. The initial population size is set to 100, and each individual encoding represents a set of rule combinations. The crossover operation uses two-point crossover with a crossover probability of 0.8; the mutation operation uses uniform mutation with a mutation probability of 0.1. The performance of each rule combination is evaluated through a fitness function, and excellent individuals are selected to enter the next generation. After 500 generations of evolutionary optimization, the optimal fuzzy inference rule set is selected. The processing of the optimal fuzzy inference rule set uses an improved BP neural network. The network structure includes an input layer (rule number nodes), two hidden layers (128 and 64 nodes respectively), and an output layer (control variable nodes). The Dropout random inactivation strategy randomly shuts down 50% of the neurons during the training process to prevent overfitting. The network adopts an online learning mode to update the weight parameters in real time and outputs a speed compensation feature vector.

[0062] The speed compensation feature vector is input into a convolutional neural network for further processing. The network contains three convolutional layers, using 32, 64, and 128 convolutional kernels respectively, with a kernel size of 3×3. Each convolutional layer is followed by a max-pooling layer with a pooling window size of 2×2. Finally, feature integration is performed through two fully connected layers (containing 512 and 256 nodes respectively), and a speed control response sequence is output. In the final stage, a backpropagation algorithm based on particle swarm optimization is used for parameter optimization. The particle swarm size is set to 200, and each particle represents a set of network parameters. The position update of the particle is based on the individual optimal position and the global optimal position. The inertia weight linearly decreases from 0.9 to 0.4, and the learning factors are both set to 2.0. Through error feedback iterative optimization, a dynamic compensation model for the winding and unwinding speed is finally output.

[0063] Taking the manufacturing of a certain type of composite wind turbine blade as an example, the original dynamic tension control data contains 500 feature dimensions, and the data volume is 1000 sample points per second. After being processed by the autoencoder, the feature dimensions are reduced to 64 dimensions, and the reconstruction error is less than 3%. In the fuzzy quantization stage, the tension value is divided into 7 fuzzy subsets, and the initial rule base contains 343 rules. After being optimized by the genetic algorithm, the number of rules is reduced to 87, and the control accuracy is improved by 30%. During the online learning process of the improved BP neural network, the network parameters are updated every 500 ms, and the prediction accuracy reaches 95%. After being processed by the convolutional neural network, the time resolution of the speed control response reaches 10 ms. Finally, through the parameter adjustment optimized by the particle swarm optimization, the dynamic control accuracy of the winding and unwinding speed reaches ±0.2 mm / s, and the adaptive compensation time is shortened to 50 ms. In the actual production process, this processing flow effectively solves the problem of unstable tension caused by speed fluctuations during the filament winding process of composites, ensuring the stability and consistency of the filament winding quality. Each processing link is based on actual engineering requirements, and the effectiveness of the algorithm is verified through quantitative optimization indicators.

[0064] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0065] (1) Encoding the state space of the winding and unwinding speed dynamic compensation model through the PPO algorithm, and combining with the multi-layer perceptron mapping to generate the shear action space representation;

[0066] (2) Estimating the value function of the shear action space representation through the double Q network, and fusing the Monte Carlo tree search strategy to output the shear state transition probability matrix;

[0067] (3) Performing temporal difference learning on the shear state transition probability matrix through the TD-Lambda algorithm, and combining with the experience replay buffer mechanism to construct the shear policy gradient vector;

[0068] (4) Inputting the shear policy gradient vector into the Actor-Critic network architecture, and through soft update policy optimization, forming a multi-filament bundle shear decision sequence;

[0069] (5) Performing hierarchical decomposition calculation on the multi-filament bundle shear decision sequence through the hierarchical reinforcement learning algorithm, and combining with the priority sampling technique to screen out the key shear timing data;

[0070] (6) Exploring and optimizing the key shear timing data through the maximum entropy reinforcement learning algorithm, and through reward function reshaping, outputting the intelligent shear instruction set.

[0071] Specifically, during the optimization process of automatic fiber placement for composite materials, the dynamic compensation model of the winding and unwinding speed is first input into the Proximal Policy Optimization (PPO) algorithm. The PPO algorithm updates the policy in a way of probability ratio truncation, encoding continuous speed compensation data into a discrete state space. The state space includes key parameters such as fiber bundle position, speed, and tension, and is subjected to feature mapping through a multi-layer perceptron with a three-layer structure (each layer contains 256, 128, and 64 neurons respectively), and finally generates a representation of the shear action space.

[0072] The formula for hierarchical decomposition calculation is expressed as:

[0073]

[0074] Where: S represents the input state space, HDC represents the hierarchical decomposition calculation function, L represents the number of hierarchies, n 2 represents the number of nodes in the l-th layer, β 2 represents the weight coefficient of the l-th layer, γ i represents the importance factor of node i, represents the state function of node i in the l-th layer, represents the transfer function of node i, δ 2 represents the inter-layer coupling coefficient, i represents the node index in the current layer, j represents the node index for the summation operation, represents the state vector of the i-th node in the l-th layer, represents the state vector of the j-th node in the l-th layer, represents the transfer vector of the i-th node in the l-th layer.

[0075] The representation of the shear action space is then input into a Double Q-Network for processing. The Double Q-Network contains two networks with the same structure but independent parameters, which respectively evaluate the action value, and suppress overestimation by taking the minimum value. At the same time, the Monte Carlo Tree Search strategy is introduced. Through 1000 sampling simulations of future state transitions, a search tree with a depth of 10 is constructed, and the shear state transition probability matrix is output. The TD-Lambda algorithm is used for temporal difference learning on the shear state transition probability matrix. The Lambda parameter is set to 0.8, which is used to balance the weights of the current reward and the future reward. The size of the experience replay buffer is set to 10000, and 256 samples are randomly sampled each time for batch learning, gradually constructing the shear policy gradient vector.

[0076] The shear policy gradient vector is input into the Actor-Critic network architecture. The Actor network is responsible for generating the action policy, and the Critic network evaluates the state value. The soft update policy adopts an update rate of 0.001, and the network parameters are smoothly updated through exponential moving average, finally forming a multi-fiber bundle shear decision sequence.

[0077] Perform hierarchical decomposition calculation on the multi-filament bundle shearing decision sequence, and decompose the complex shearing task into multiple subtasks. The priority sampling technique assigns weights to samples based on the magnitude of TD error, preferentially learning experiences with high TD error, and screening out key shearing time-series data. Finally, the maximum entropy reinforcement learning algorithm is used to explore and optimize the key shearing time-series data. The entropy weight coefficient is set to 0.01, maintaining moderate policy randomness while maximizing the expected return. The reward function comprehensively considers factors such as shearing accuracy, timing rationality, and energy consumption, and outputs the final intelligent shearing instruction set after reshaping and optimization.

[0078] Taking the manufacturing of a certain type of helicopter rotor blade as an example, the original speed compensation model contains the laying data of 8 fiber bundles, and a 256-dimensional state space representation is obtained after encoding by the PPO algorithm. After evaluation by the double Q network, 10 candidate actions are generated for each shearing position, and the dimension of the state transition probability matrix is 256×10. Through 1000 rounds of iterative learning by the TD-Lambda algorithm, the dimension of the policy gradient vector is optimized to 64 dimensions. After processing by the Actor-Critic network, the generated multi-filament bundle shearing decision sequence contains the optimal shearing parameters of each fiber bundle at different times. The hierarchical decomposition divides the entire shearing process into three levels: preparation, execution, and verification, and 752 key time-series points are screened out from the original 8000 decision points. Finally, after optimization by the maximum entropy reinforcement learning algorithm, the intelligent shearing instruction set achieves a shearing accuracy of ±0.15mm, and the shearing response time is reduced to 20ms, greatly improving the manufacturing efficiency and quality stability of composite components.

[0079] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0080] (1) Extract spatio-temporal features of the intelligent shearing instruction set through a recursive neural tensor network, and combine attention mechanism operations to generate a refeeding trajectory feature point cloud;

[0081] (2) Perform contrast learning calculation on the refeeding trajectory feature point cloud through a siamese convolutional neural network, and combine a three-dimensional space transformation network to output a refeeding trajectory encoding sequence;

[0082] (3) Process the refeeding trajectory encoding sequence through a graph convolutional network, fuse edge feature extraction operations, and construct a refeeding trajectory topology map;

[0083] (4) Input the refeeding trajectory topology map into a simulation environment based on a physical engine, and calculate through the Verlet integration algorithm to form a refeeding dynamics parameter set;

[0084] (5) Perform state estimation on the refeeding dynamics parameter set through a self-correcting particle filter algorithm, and combine Kalman smoothing processing to screen out the optimal solution of the refeeding trajectory;

[0085] (6) Uncertainty quantification of the optimal solution of the retransmission trajectory is performed through a variational inference algorithm. After Bayesian posterior distribution calculation, a retransmission path matrix is output.

[0086] Specifically, the intelligent shear instruction set is input into a recursive neural tensor network for processing. The recursive neural tensor network adopts a three-dimensional tensor structure, and each neuron contains a weight tensor of 128×128×128 to capture high-order feature correlations. Temporal features are extracted through time-recursive connections, and spatial features are extracted through spatial convolution operations. The attention mechanism adopts a multi-head structure with 8 attention heads, and the attention dimension of each head is 64. Features at different positions are weighted and aggregated to finally generate a retransmission trajectory feature point cloud containing spatial position and temporal information. The retransmission trajectory feature point cloud is then input into a siamese convolutional neural network. The siamese network consists of two convolutional branches with shared weights. Each branch contains 5 convolutional layers, and the kernel size decreases from 7×7 to 3×3, while the number of channels increases from 32 to 256. Contrastive learning uses a triplet loss function to pull similar trajectories closer and push different trajectories farther apart. The three-dimensional spatial transformation network contains three transformation modules: rotation, translation, and scaling. Each module predicts transformation parameters through a fully connected layer to achieve spatial alignment of features, and finally outputs a standardized retransmission trajectory encoding sequence.

[0087] The retransmission trajectory encoding sequence undergoes deep feature extraction through a graph convolutional network. The graph convolutional network constructs the trajectory sequence into a directed graph structure, where nodes represent trajectory points and edges represent the connection relationships between points. The network contains 3 graph convolutional layers, and the output channel numbers of each layer are 64, 128, and 256 respectively. Edge feature extraction adopts dynamic edge convolution operations to adaptively learn the connection weights between nodes and construct a complete retransmission trajectory topology map. The retransmission trajectory topology map is input into a simulation environment based on a physics engine. The physics engine considers physical constraints such as gravity, friction, and elasticity, and calculates the motion state of trajectory points through the Verlet integration algorithm. The integration time step is set to 0.001 seconds, and numerical stability is ensured through iterative updates of position and velocity at each time step, and a set of retransmission dynamics parameters containing information such as position, velocity, and acceleration is output.

[0088] The resending kinetic parameter set is used for state estimation through the self - calibrating particle filter algorithm. The particle filter algorithm uses 1000 particles to represent the state distribution, and each particle contains complete kinetic parameters. The self - calibrating mechanism dynamically adjusts the particle weights according to the observed data to increase the sampling density in the high - probability region. The Kalman smoother uses the forward - backward algorithm to perform two - way smoothing on the state sequence, eliminate noise fluctuations, and select the optimal solution of the resending trajectory. The optimal solution of the resending trajectory is used for uncertainty quantification through the variational inference algorithm. The variational inference approximates the posterior distribution as a Gaussian distribution family and iteratively optimizes the distribution parameters by minimizing the KL divergence. The Bayesian posterior calculation uses the Monte Carlo sampling method with 10000 samples, calculates the uncertainty interval of the trajectory parameters, and finally outputs a resending path matrix with confidence information.

[0089] Taking the manufacturing of the fuselage skin of a large civil airliner as an example, the intelligent cutting instruction set contains the cutting information of 8 bundles of fibers, and each bundle of fibers has 500 key timing points. After being processed by the recursive neural tensor network, point cloud data of 50000 feature points is generated. Through the contrast learning of the Siamese network, the feature dimension is compressed to 1024 dimensions, and the spatial transformation network achieves an alignment accuracy of ±0.05mm. The topological graph constructed by the graph convolutional network contains 2000 nodes and 8000 edges, and the edge feature dimension is 256. The physical engine simulation completes 10000 steps of integral calculation within 1 second to generate a complete kinetic trajectory. After particle filter processing, the average error of state estimation is reduced to 0.1mm, and the Kalman smoother controls the trajectory jitter amplitude within 0.05mm. The finally output resending path matrix has a dimension of 8×500×6, contains the three - dimensional position and attitude information of each timing point, and is accompanied by a 95% confidence interval.

[0090] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0091] (1) Multidimensionally segment the temperature field data through the resending path matrix in three - dimensional space, generate a temperature data cube according to the X, Y, and Z axes, and then perform downsampling conversion on the temperature data cube to output the basic temperature field dataset;

[0092] (2) Perform high - dimensional numerical discretization processing on the basic temperature field dataset, convert the continuous temperature surface into a node matrix, and then perform sparsification processing on the node matrix to construct a multi - level node data array of the temperature field;

[0093] (3) Pass the multi - level node data array of the temperature field through thermodynamic feature extraction operations, and after energy conservation conversion processing, output a multi - dimensional temperature feature data stream;

[0094] (4) Perform spatiotemporal serialization processing on the temperature multidimensional characteristic data stream, rearrange the data according to the time axis, superimpose the spatial position information labels, and generate a four-dimensional evolution surface of the temperature field;

[0095] (5) Perform temperature gradient analysis on the four-dimensional evolution surface of the temperature field through core feature extraction and threshold screening to output the temperature key control point cloud;

[0096] (6) The temperature key control point cloud is converted into a piecewise continuous control signal sequence, and after smoothing and outlier correction, the temperature field dynamic control sequence is output.

[0097] Specifically, the temperature field information in the retransmission path matrix is ​​processed in three-dimensional space. Specifically, data segmentation is performed in three orthogonal directions, X, Y, and Z, and the resolution of each direction is set to 0.1 mm. A temperature data cube is constructed within the actual workpiece size range (such as 500 mm × 300 mm × 50 mm). For each spatial point, its corresponding temperature value is recorded to form a high-dimensional temperature distribution data set. In order to reduce the computational burden, the temperature data cube is downsampled, and the sampling interval is set to 1 mm. While maintaining the temperature field distribution characteristics, the data volume is greatly reduced, and the temperature field basic data set is finally output. The temperature field basic data set is then processed by high-dimensional numerical discretization. First, the continuous temperature distribution surface is converted into a discrete node matrix with a node spacing of 1 mm. Each node records the temperature value and its spatial coordinates. Considering the continuity and smoothness of the actual temperature field distribution, the node matrix is ​​sparsely processed, and the relative temperature gradient threshold method is used to delete the nodes with insignificant temperature changes (areas with temperature gradients less than 2 ° C / mm), and the dense sampling points in the area of ​​​​drastic temperature changes are retained, and finally a multi-level node data array of the temperature field is constructed. The multi-level node data array of the temperature field is processed by thermodynamic feature extraction operations. First, the heat flux and thermal diffusion coefficient of each node are calculated, taking into account the thermal conductivity of the material (such as the transverse thermal conductivity of carbon fiber is 0.8W / (m·K), and the longitudinal thermal conductivity is 8W / (m·K)). At the same time, the influence of thermal radiation and convection heat transfer is considered, and the energy balance equation under the boundary conditions is calculated. Under the premise of ensuring energy conservation, the temperature field is feature extracted to obtain a temperature multidimensional feature data stream containing information such as heat flux density, temperature gradient, and thermal resistance. The temperature multidimensional feature data stream is time-space serialized, with a sampling time interval of 0.1 seconds, and the complete three-dimensional temperature field distribution is recorded at each time point. The data is rearranged according to the time axis, and a spatial position label is added to each data point, including absolute coordinates and relative position information. Through the combination of time and space dimensions, a four-dimensional evolution surface of the temperature field is generated, which clearly shows the change law of the temperature field over time and space.

[0098] After the four-dimensional evolution surface of the temperature field undergoes core feature extraction operations, the focus is on analyzing the temperature gradient changes. Calculate the temperature gradient values in each direction, and identify regions with significant temperature changes (gradient greater than 5 °C / mm) and regions with stable temperatures (gradient less than 1 °C / mm). By setting a temperature gradient threshold (3 °C / mm) for screening, extract the temperature control key points that have a significant impact on the filament winding quality, and form a temperature key control point cloud.

[0099] Finally, convert the temperature key control point cloud into a real-time control signal. Use the cubic spline interpolation algorithm to connect the discrete control points to generate a smooth temperature control curve. Correct the abnormal temperature points (deviating from the mean by more than 3 standard deviations) to ensure the continuity and reliability of the control signal. The finally output dynamic control sequence of the temperature field contains complete timing control instructions for guiding the real-time adjustment of the heating system.

[0100] For example: During the filament winding process of the cylindrical part (diameter 2 meters, length 3 meters), first divide the workpiece into 200,000 spatial grid points, and record the temperature data at each point. After downsampling, 20,000 temperature sampling points are retained to form the basic data set of the temperature field. The discretization process converts the continuous temperature field into 4,000 key nodes, and each node contains temperature values and three-dimensional coordinate information. During the thermodynamic feature extraction process, the thermal conductivity (0.8 - 8 W / (m·K)) and heat flux density (0 - 500 W / m 2 ) of each node are calculated. The spatio-temporal serialization process records the temperature changes during the 8-hour filament winding process, with a sampling interval of 0.1 second, generating 288,000 time-series data points. Through temperature gradient analysis, it is found that the temperature gradient reaches 6 °C / mm in the fiber crossing area, while the temperature gradient in the flat area is only 0.5 °C / mm. Finally, 500 temperature control key points are selected, and a smooth control sequence is generated through cubic spline interpolation, with a control accuracy of ±2 °C.

[0101] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0102] (1) Process the dynamic control sequence of the temperature field through multi-dimensional constraint decoupling operations, perform hierarchical decomposition on the time dimension, space dimension, and temperature dimension, and output a multi-objective feature vector group;

[0103] (2) Perform regularization processing on the multi-objective feature vector group, correct the data according to the physical constraint conditions, and generate a constraint condition boundary matrix;

[0104] (3) Perform multi-scale feature decomposition operations on the constraint condition boundary matrix, extract the dynamic parameter change rules, and construct a strategy optimization feature space;

[0105] (4) Quantify the policy optimization feature space through the reward and punishment function, assign numerical values to the weights of each parameter, and form a multi-objective optimization data stream;

[0106] (5) Perform hierarchical processing on the multi-objective optimization data stream, and output the laying optimization parameter set through the dynamic threshold screening mechanism;

[0107] (6) Integrate the laying optimization parameter set, reorganize the timing of the control instructions, and output the laying strategy data packet.

[0108] Specifically, perform multi-dimensional constraint decoupling processing on the dynamic temperature field control sequence. In the time dimension, sample the continuous control sequence at intervals of 0.1 seconds, and record the temperature control value and change rate at each time point. In the space dimension, divide the three-dimensional space into grid cells of 10mm×10mm×1mm, and record the position coordinates and temperature distribution of each cell. In the temperature dimension, discretize the temperature values at intervals of 5°C to form a temperature level sequence. Through data dimensionality reduction and feature extraction, integrate the information of the three dimensions into a multi-objective feature vector group. Perform regularization processing on the multi-objective feature vector group. First, perform standardization transformation on the data to make the data distributions of different dimensions within the same numerical range. Physical constraint conditions include the temperature upper limit (350°C), the temperature change rate limit (±10°C / s), the spatial temperature gradient constraint (≤5°C / mm), etc. According to these constraint conditions, correct the data points outside the range to ensure that all data meet physical realizability. The corrected data is reorganized into a constraint condition boundary matrix, where each row of the matrix represents a set of constraint conditions, and each column corresponds to a control parameter.

[0109] The constraint condition boundary matrix undergoes multi-scale feature decomposition operations. Use wavelet transform to decompose the data into different frequency components. For high-frequency components, extract the features of rapid temperature changes; for low-frequency components, focus on the long-term temperature change trend. By analyzing the variation laws of parameters at different scales, summarize control laws, such as the heating power curve during the temperature rise stage and the temperature fluctuation range during the stable stage. These laws are encoded into the feature space to form the basis of policy optimization. The policy optimization feature space is then quantified through the reward and punishment function. Set the reward factor and punishment factor. The reward factor targets ideal control effects (such as good temperature stability and high bonding strength between adjacent layers), while the punishment factor targets adverse conditions (such as material damage caused by too high temperature and poor interlayer bonding due to too low temperature). According to the influence degree of different parameters on the final quality, assign weight coefficients to generate a multi-objective optimization data stream. The weight assignment uses the analytic hierarchy process to quantify the importance of each parameter into specific numerical values.

[0110] The multi-objective optimization data flow undergoes hierarchical processing and is divided into multiple levels according to the priority of control objectives. The highest priority includes temperature stability and material safety, the secondary priority includes energy efficiency and process rhythm, and finally, there are auxiliary optimization indicators. Corresponding dynamic thresholds are set for each level, and the thresholds are adaptively adjusted according to the process stage. Through this multi-level screening mechanism, each parameter is gradually optimized, and finally, a set of laying optimization parameter sets that meet all requirements is output. The laying optimization parameter sets finally undergo strategy integration processing. The discrete optimization parameters are rearranged in chronological order to construct a complete control strategy sequence. The control instructions at adjacent time points are smoothed to eliminate mutation points and ensure the continuity of control. At the same time, considering the coupling relationship between different control parameters, the control instructions are coordinated and optimized, and finally, a laying strategy data packet is output.

[0111] Taking the manufacturing of a certain type of composite material aerospace cabin section as an example: The original dynamic temperature control sequence contains 8 hours of temperature data with a sampling frequency of 10 Hz, generating 288,000 temperature control points. After multi-dimensional decoupling, 2,880 key time points (sampling interval of 10 seconds) are retained in the time dimension, the space dimension is divided into 1,000 control units, and the temperature dimension is divided into 70 levels (0 - 350 °C, interval of 5 °C). Regularization processing normalizes all parameters to the [-1, 1] interval. Through screening by physical constraint conditions, 85% of the valid data points are retained. Multi-scale feature decomposition extracts typical control modes, such as the preheating stage (0 - 50 °C / min), the working stage (±2 °C / min), and the cooling stage (-20 °C / min). The reward and punishment function quantifies and gives a weight of 0.4 to temperature stability, a weight of 0.3 to material properties, a weight of 0.2 to energy efficiency, and a weight of 0.1 to other factors. Dynamic threshold screening optimizes 1,000 control units into 200 key control points. The finally generated laying strategy data packet achieves a temperature control accuracy of ±2 °C, the interlayer bonding strength reaches 98% of the design requirements, and the energy utilization efficiency is increased by 15%.

[0112] The above describes the composite material automatic fiber placement forming optimization method in the embodiments of the present application. Next, the composite material automatic fiber placement forming optimization system in the embodiments of the present application will be described. Please refer to Figure 2 One embodiment of the composite material automatic fiber placement forming optimization system in the embodiments of the present application includes:

[0113] An input module, configured to output dynamic tension control data after the tension data of the prepreg tow input by the tension sensor undergoes LSTM deep learning operations;

[0114] An operation module, configured to input the dynamic tension control data into a fuzzy neural network computing engine, and obtain a retracting and paying-off speed dynamic compensation model after speed curve analysis operations;

[0115] A calculation module, configured to calculate the multi-filament shear sequence by using a reinforcement learning engine based on the retraction speed dynamic compensation model, and generate an intelligent shear instruction set;

[0116] A construction module, configured to construct a retraction trajectory mapping relationship based on the intelligent shear instruction set, and form a retraction path matrix after iterative calculation by a digital twin engine;

[0117] An establishment module, configured to establish a temperature field distribution prediction model according to the retraction path matrix, and construct a temperature field dynamic control sequence after finite element analysis calculation;

[0118] An output module, configured to substitute the temperature field dynamic control sequence into a multi-objective constraint equation, and output a placement strategy data packet after iterative optimization by deep reinforcement learning.

[0119] Through the collaborative cooperation of the above-mentioned various components, the tension data of the prepreg filaments is processed by LSTM deep learning, and the precise analysis of the speed curve is realized by combining a fuzzy neural network calculation engine, improving the accuracy and stability of tension control. The reinforcement learning engine optimizes the calculation of the multi-filament shear sequence, and realizes the precise mapping of the retraction trajectory through the iterative calculation of the digital twin engine, greatly improving the shear positioning accuracy and retraction reliability. The temperature field distribution prediction model combines finite element analysis technology to realize the dynamic and precise control of the temperature field, ensuring the temperature uniformity and stability during the laying process. Through multi-objective constraint optimization of the temperature field dynamic control sequence by deep reinforcement learning, the global optimization of the placement strategy is realized, significantly improving the quality stability of the filament laying process. The entire solution realizes the collaborative optimization of multiple key parameters such as tension control, speed compensation, shear positioning, and temperature control during the filament laying process through multi-level data processing and intelligent algorithm optimization, effectively solving the problems of improper parameter matching and unstable process existing in traditional filament laying technology. At the same time, the intelligent optimization method based on deep learning and reinforcement learning enables the control strategy to be adaptively adjusted to meet the filament laying requirements under different working conditions, improving the flexibility and intelligence level of the production process. In addition, the introduction of digital twin technology makes the placement trajectory planning more accurate and the temperature field prediction more accurate, providing reliable technical support for high-quality filament laying manufacturing. Through multi-objective constraint optimization, while ensuring the product quality, the energy efficiency is improved and the production cost is reduced, having significant economic benefits and practical value.

[0120] This application also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions run on a computer, the computer is made to execute the steps of the composite material automatic filament laying forming optimization method.

[0121] The present application also provides a fiber placement machine, which includes: one or more processors; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement an optimization method for automatic fiber placement molding of composite materials.

[0122] The fiber placement machine provided by the present application, as Figure 3 and Figure 4 shown, are the first schematic diagram and the second schematic diagram of the fiber placement machine provided by the present application, Figure 3 which includes: a wire guiding wheel 31, a clamping device 84, a refeeding device 75, and an infrared heating lamp 96. Figure 4 which includes: a connecting flange 153, a control system and a tension motor device 166, a carbon fiber prepreg 129, and a fiber placement device 53.

[0123] The fiber placement machine can only place thermosetting and thermoplastic carbon fiber prepregs. 8 rolls of thermosetting fiber prepregs are installed on a storage roll. Each bundle of fibers is led out from the storage yarn roll. After passing through a film collecting device, the release liner on the surface of the prepreg tow is recovered by a separate film collecting roll, and at the same time each tow enters a tension control mechanism. The tow led out from the tension control mechanism has a certain amount of tension. Subsequently, the tow passes through an independent clamping mechanism, a separate refeeding mechanism, and a separate shearing mechanism, and finally is placed on the surface of a mold under the action of a heating device and a compaction mechanism to achieve heating and compaction for shaping, completing the placement process.

[0124] The fiber prepreg tow itself has a certain stickiness. During the storage process of the tows, to prevent the tows from sticking to each other, a release liner made of polyethylene material is covered on one side of each tow. If the fiber tow fails to separate from the release liner in time before entering the placement process, the release liner involved in the placement process will affect the tight adhesion between the tow layers, causing quality defects such as pores, seriously reducing the molding quality. Therefore, it is necessary to recover the release liner. The difficulty in the design lies in that the feeding of the fiber tow and the recovery of the release liner are a dynamic synchronous process, and the feeding speed of the fiber tow and the recovery speed of the release liner should be kept consistent. As the placement process progresses, the shaft diameter of the storage roll becomes smaller and smaller, and the shaft diameter of the film collecting shaft becomes larger and larger. Therefore, how to maintain the real-time matching of the feeding speed and the film collecting speed is the difficulty in the design.

[0125] The film winding action is a real-time dynamic matching action, and the unwinding speed of the fiber tow and the film winding speed should be basically the same. If the film winding speed cannot keep up with the unwinding speed, the release liner paper that has not had time to separate from the fiber tow will be involved in the laying process, which will affect the fiber shearing performance to a lesser extent and cause serious product quality defects in severe cases. If the unwinding speed cannot keep up with the film winding speed, the fiber tow will accumulate before entering the laying process, affecting the laying efficiency and quality. For the unwinding-film winding, our solution is as follows: Use friction to achieve real-time dynamic matching of the film winding speed and the unwinding speed. The fiber tow is stored on the roll in the form of winding, and the release liner paper adheres to the upper surface of the fiber tow. The friction roller and the film winding roller are in interference fit. Under the action of tension, the fiber tow covered with the release liner paper is led out from the unwinding roll and divided into two paths. One path is the fiber tow passing through the friction roller, and the other path is the release liner paper wound around the film winding roller. The fiber tow passing through the friction roller drives the film winding roller to rotate in the form of friction force, so that the release liner paper separated from the fiber tow can be gradually wound around the film winding roller to complete the film winding action. The invention has the advantages of simple structure, convenient disassembly and maintenance; it can achieve real-time speed matching of yarn unwinding-film winding with small fluctuations.

[0126] During the fiber laying process, there are mainly two tension stages: the refeeding tension in the tow refeeding stage and the laying tension in the fiber laying stage. If the refeeding tension is too small, the tow will be loose during the refeeding process and easily break away from the guiding device; if the refeeding tension is too large, slippage between the tow and the roller is likely to occur, reducing the refeeding accuracy, and a large feeding tension means a greater pressing force, which will more seriously cause fiber wear. If the laying tension is too small, the product lacks prestress and its mechanical properties decline; if the laying tension is too large, bridging is likely to occur when laying on the concave surface, and in more serious cases, the fiber will be broken. At the same time, if the laying tension fluctuates violently, the prestress of each part of the product will be inconsistent, and the overall mechanical properties will be affected. Therefore, maintaining an appropriate and sufficiently stable tension has an important impact on the fiber laying quality.

[0127] The function of the tension control mechanism is to keep the fiber tow under a certain tension during the automatic fiber laying process to reduce the tension fluctuation during the laying process. The tension control method can adopt a combination of two schemes: active control and passive adjustment. The first active method is to provide damping for the unwinding roll during the fiber unwinding process, and the resistance is provided by a torque motor or an electronic damper. The resistance of the unwinding roll is closed-loop controlled through a tension sensor combined with the system preset value. The second passive method is to apply tension to the tension shaft through a tension cylinder. The cylinder tension compensates for the tension accuracy of the tension shaft through electronic control proportionality combined with the system preset value. At the same time, precise closed-loop control is implemented for the application and reduction of tension in cooperation with shearing and refeeding.

[0128] The fiber laying device bundles multiple prepreg bundles into different adjustable lengths. The length adjustment is achieved by controlling the cutting action of any fiber bundle. Therefore, the shearing mechanism must have the function of independently cutting any fiber bundle. The function of the shearing mechanism is to receive the command of the control system to cut the fiber. When cutting, the fiber bundle remains taut under the coordination of tension. To cut the fiber bundle quickly and neatly, the shearing mechanism should meet the following requirements: fast response, quick action, and sufficient shearing force.

[0129] The cutter is the core part of the shearing mechanism. There are two cutting methods: ultrasonic cutting and mechanical cutting. The ultrasonic cutter uses ultrasonic energy to locally heat and melt the material to be cut to achieve the purpose of cutting the material. Ultrasonic cutting can seal the cutting edge, which better avoids the shortcomings of fuzzing and wire drawing at the incision, but it is large in size, difficult to install, and expensive. Mechanical cutting uses a sharp-edged tool to cut the fiber bundle. It is low-cost, easy to install, and can operate in a narrow space. At the same time, the cutting performance can be improved by the reasonable design of the tool and the control of the cutting speed. Considering the strict requirements of the wire laying head for compact structure, pneumatic mechanical cutting is used to minimize the pipeline setting value.

[0130] The shape of the cutting edge has an important influence on the cutting performance. Common cutting edge design schemes include single-sided cutting and double-sided cutting. Single-sided cutting is to grind a bevel on one side of the blade tip and keep the other side flat. The biggest advantage of single-sided cutting is that it can grind a very small bevel angle and the cutting edge is sharp. However, the blade surface is thin and very easy to wear. Double-sided cutting is to grind a bevel on both sides of the blade tip to form a cutting edge. The blade surface has higher strength than a single blade and is more suitable for straight cutting occasions. The shearing of fiber filaments is a process dominated by straight cutting, so a single-sided cutting design is adopted. At the same time, the cutting edge is designed to have a certain slope. Based on a large number of experiments and experience, we set this angle to 5°. It effectively avoids the phenomenon of knife collision and plays a role in protecting the cutter. When 8 tows are laid, the width of each bundle is 6.35mm. If single-filament cutting is to be achieved, a highly integrated cutting device is required. We use cylinders as power and customize double-force cylinders to achieve smooth cutting of tows with a thickness of less than 0.2mm. At the same time, each bundle of fibers is driven by an independent cylinder for shearing.

[0131] The re-feeding mechanism can send the cut ends of the fiber bundles back to the compacting mechanism when laying is required, and then enter the re-laying process. The reliability of the re-feeding mechanism directly affects the re-feeding accuracy. The function of the re-feeding mechanism is to re-transmit the cut fiber ends to the compacting mechanism when the width of the laid fiber bundle needs to be increased. The reliability of the re-feeding mechanism has an important impact on the laying accuracy. In order to meet the requirements of multi-channel re-feeding in a narrow space: ensure that the re-feeding force is appropriate, the fiber bundles neither slip nor break; at the same time, the action response is fast and the reliability is high. The fiber re-feeding scheme is as follows:

[0132] The coaxial multi-position form is adopted to drag the fiber to complete refeeding. In this invention, whenever the fiber is sheared by the fiber laying head and the fiber needs to be retransported to the compaction mechanism, the refeeding motor drives the refeeding roller by driving the internal and external meshing gears. With the cooperation of the pressure wheel, the end of the sheared fiber is dragged to the pressure roller mechanism to complete one refeeding, and the refeeding accuracy of ±0.15 mm can be ensured. The refeeding pressure wheel keeps the fiber ribbon clamped with the refeeding wheel in real time under pneumatic drive, and cooperates with the system to transport and stop the fiber ribbon according to the analysis of the laid model. This invention uses the method of roller extrusion for refeeding, which can meet the requirements of fiber refeeding in a limited space. And by selecting a reasonable material for the refeeding wheel, the refeeding performance of the fiber can be effectively enhanced. By adjusting the rotation speed and time of the driving wheel, the refeeding speed and length of the fiber can be controlled. The control is simple and the structure is compact.

[0133] The difference between this invention and the traditional compaction roller is that a whole pressure roller is used to replace the independent pressing block. Two groups of micro pressure sensors are installed on the fixed substrate and the moving plate of the pressure roller. The range of each group of sensors is 0 - 400 N, and the accuracy is 0.02%. The final pressure is 800 N. This sensor can feedback the pressure of the pressure roller in real time, and accurately adjust the required compaction force of the equipment according to the preset value of the system in combination with the gap between the pressure roller and the mold or the laying layer.

[0134] When the fiber laying device is laying, each bundle of fibers has a certain tension. When a single fiber is cut by the cutting device, the fiber bundle at the rear end needs to be retransported to the pressure roller mechanism as required. However, this bundle of fibers always maintains a certain tension, and when the refeeding pressure wheel is lifted, it will cause the fiber bundle to retract. In this way, the refeeding will fail the next time. Therefore, a clamping device is added at the upstream end of the refeeding mechanism. After the fiber bundle passes through the clamping system, no matter how the downstream refeeding mechanism and the cutting mechanism cooperate, the fiber bundle cannot retract automatically, which ensures the next refeeding. The clamping system is composed of a group of fixed wheels and a group of moving wheels. A one-way bearing is installed inside the fixed wheel, which causes the fiber bundle to move in only one direction. When the moving wheel group clamps the fiber bundle, the one-way bearing plays a role in stopping, which ensures that the cut fiber does not shrink.

[0135] Heating the carbon fiber prepreg during the laying process is an indispensable part of the process. Without heating, the prepregs cannot be closely adhered to each other, there are protrusions on the laying surface, there are gaps between layers, and the laying surface is very uneven. After heating, the laying quality is significantly improved and the laying surface is flat. Infrared heating refers to the process in which the heated body absorbs radiant energy and converts it into heat energy after being irradiated by infrared rays. Infrared heating is actually a special form of resistance heating. In the present invention, 3 infrared heating tubes are used as light sources. Due to limited installation space, we use double-tube infrared tubes with a length of 100 mm, and the power of a single tube is 700 W. At a fixed position, its heating temperature can reach 350 °C, and the temperature error is ±3 °C. At the same time, because the temperature of the carbon fiber rises significantly under the irradiation of the infrared heater.

[0136] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than limiting them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A composite material automatic wire placement molding optimization method, characterized in that: The composite material automatic wire placement forming optimization method comprises: The prepreg tow tension data input by the tension sensor is processed through LSTM deep learning to output dynamic tension control data; The dynamic tension control data is input into a fuzzy neural network calculation engine, and after speed curve analysis and calculation, a dynamic compensation model for the retraction and release speed is obtained; Based on the retracting and releasing speed dynamic compensation model, a reinforcement learning engine is used to calculate the multi-tow cutting sequence and generate an intelligent cutting instruction set; A re-delivery trajectory mapping relationship is constructed based on the intelligent cutting instruction set, and a re-delivery path matrix is ​​formed after iterative calculation by the digital twin engine; A temperature field distribution prediction model is established according to the retransmission path matrix, and a temperature field dynamic control sequence is constructed after finite element analysis and calculation; The temperature field dynamic control sequence is substituted into the multi-objective constraint equation, and after deep reinforcement learning iterative optimization, a placement strategy data package is output.

2. The composite material automatic wire placement molding optimization method according to claim 1, characterized in that: The prepreg tow tension data input by the tension sensor is processed by LSTM deep learning operation to output dynamic tension control data, including: Performing multi-scale time-frequency decomposition on the prepreg tow tension data by wavelet transform and Fourier analysis, and generating a tension multi-dimensional frequency spectrum feature matrix after singular value decomposition filtering; The tension multi-dimensional spectrum feature matrix is ​​input into a three-layer LSTM network structure containing 64 hidden neurons, and after forward recursive calculation and long short-term memory unit state update, a tension spatiotemporal state vector set is output; According to the tension spatiotemporal state vector set, deep temporal correlation mining is performed through a bidirectional gated recurrent unit based on the Transformer architecture, and a tension prediction deep feature map is formed in combination with a residual connection mechanism; The tension prediction deep feature map is calculated by a multi-head attention mechanism, and the self-attention and cross-attention weight distributions are integrated to screen out a high-dimensional tension feature sequence; The high-dimensional tension feature sequence is iteratively optimized in batches by using an improved stochastic gradient descent algorithm, and a tension prediction loss function surface is constructed by combining L1 and L2 regularization constraints; The tension prediction loss function surface is substituted into the Adam optimizer with momentum term, and the dynamic tension control data is output after batch normalization processing through adaptive learning rate adjustment and gradient clipping technology.

3. The composite material automatic wire placement molding optimization method according to claim 1, characterized in that: The dynamic tension control data is input into the fuzzy neural network calculation engine, and after speed curve analysis and calculation, a dynamic compensation model for the retracting and releasing speed is obtained, including: Performing noise reduction and feature extraction operations on the dynamic tension control data through an autoencoder to output a tension feature dimension reduction data set; The tension feature dimension reduction data set is fuzzy quantized by a triangular membership function, and a fuzzy rule base matrix is ​​constructed by Mamdani reasoning mechanism operation; The fuzzy rule base matrix is ​​optimized and screened by a genetic algorithm, and crossover and mutation operations are integrated to generate an optimal fuzzy reasoning rule set; The optimal fuzzy inference rule set is subjected to online learning calculation by an improved BP neural network, combined with a Dropout random deactivation strategy, to form a speed compensation feature vector; The speed compensation feature vector is input into a convolutional neural network to extract speed curve features, and a speed control response sequence is constructed through multi-layer pooling and full connection operations; The speed control response sequence is subjected to parameter optimization calculation by a back propagation algorithm based on particle swarm optimization, and the retraction and extension speed dynamic compensation model is output after error feedback iteration.

4. The composite material automatic wire placement molding optimization method according to claim 1, characterized in that: The method uses a reinforcement learning engine to calculate the multi-tow cutting sequence based on the retracting and releasing speed dynamic compensation model and generate an intelligent cutting instruction set, including: The retracting and releasing speed dynamic compensation model is encoded in state space by PPO algorithm, combined with multi-layer perceptron mapping, to generate a shearing action space representation; The shear action space representation is passed through a double Q network to estimate the value function, and a Monte Carlo tree search strategy is integrated to output a shear state transition probability matrix; The shear state transition probability matrix is ​​subjected to temporal difference learning through the TD-Lambda algorithm, and a shear strategy gradient vector is constructed in combination with an experience replay buffer mechanism; The cutting strategy gradient vector is input into the Actor-Critic network architecture, and a multi-tow cutting decision sequence is formed through soft update strategy optimization; The multi-tow cutting decision sequence is subjected to hierarchical decomposition calculation by a hierarchical reinforcement learning algorithm, and the key cutting time series data is screened out by combining the priority sampling technology; The key cutting time series data is explored and optimized through the maximum entropy reinforcement learning algorithm, and reshaped through the reward function to output the intelligent cutting instruction set.

5. The composite material automatic wire placement molding optimization method according to claim 1, characterized in that: The re-delivery trajectory mapping relationship is constructed based on the intelligent cutting instruction set, and a re-delivery path matrix is ​​formed after iterative calculation by the digital twin engine, including: The intelligent cutting instruction set is subjected to spatiotemporal feature extraction through a recursive neural tensor network, and combined with attention mechanism operation to generate a feature point cloud of the retransmission trajectory; The feature point cloud of the re-delivery trajectory is subjected to comparative learning calculation by a twin convolutional neural network, and combined with a three-dimensional space transformation network, a re-delivery trajectory coding sequence is output; The re-delivery trajectory coding sequence is processed by a graph convolutional network, and an edge feature extraction operation is integrated to construct a re-delivery trajectory topology map; The re-delivery trajectory topology map is input into a simulation environment based on a physical engine, and is calculated by a Verlet integration algorithm to form a re-delivery dynamics parameter set; The state of the re-delivery dynamics parameter set is estimated by a self-correcting particle filter algorithm, and the optimal solution of the re-delivery trajectory is screened out by combining Kalman smoothing. The uncertainty of the optimal solution of the re-delivery trajectory is quantified through a variational inference algorithm, and the re-delivery path matrix is ​​output after Bayesian posterior distribution calculation.

6. The composite material automatic wire placement molding optimization method according to claim 1, characterized in that: The temperature field distribution prediction model is established based on the retransmission path matrix, and after finite element analysis and calculation, a temperature field dynamic control sequence is constructed, including: The retransmission path matrix is ​​multi-dimensionally segmented into temperature field data through three-dimensional space stratification, a temperature data cube is generated according to the X, Y, and Z axes, and then the temperature data cube is downsampled and converted to output a temperature field basic data set; Performing high-dimensional numerical discretization processing on the basic data set of the temperature field, converting the continuous temperature surface into a node matrix, and then performing sparse processing on the node matrix to construct a multi-level node data array of the temperature field; The temperature field multi-level node data array is subjected to thermodynamic feature extraction operation, and energy conservation conversion processing to output a temperature multi-dimensional feature data stream; Performing spatiotemporal serialization processing on the temperature multidimensional characteristic data stream, rearranging the data according to the time axis, superimposing spatial position information labels, and generating a four-dimensional evolution surface of the temperature field; The temperature field four-dimensional evolution surface is subjected to core feature extraction operation, temperature gradient analysis is performed, and then threshold screening is performed to output a temperature key control point cloud; The temperature key control point cloud is converted into a segmented continuous control signal sequence, and after smoothing and outlier correction, the temperature field dynamic control sequence is output.

7. The composite material automatic wire placement molding optimization method according to claim 1, characterized in that: Substituting the temperature field dynamic control sequence into the multi-objective constraint equation, and after deep reinforcement learning iterative optimization, outputting a placement strategy data package, including: The temperature field dynamic control sequence is processed by multi-dimensional constraint decoupling operation, the time dimension, space dimension and temperature dimension are layered and disassembled, and a multi-objective feature vector group is output; Regularizing the multi-objective feature vector group, performing data correction according to physical constraints, and generating a constraint boundary matrix; Performing multi-scale feature decomposition operation on the constraint condition boundary matrix, extracting the dynamic parameter change law, and constructing a strategy optimization feature space; The strategy optimization feature space is quantized through a reward and punishment function, and numerical values ​​are assigned to the weights of each parameter to form a multi-objective optimization data stream; hierarchically processing the multi-objective optimization data stream, and outputting a set of placement optimization parameters through a dynamic threshold screening mechanism; The placement optimization parameter set is subjected to strategic integration processing, the control instructions are reorganized in time sequence, and the placement strategy data packet is output.

8. A composite material automatic wire placement forming optimization system, used to implement the composite material automatic wire placement forming optimization method according to any one of claims 1 to 7, characterized in that: The composite material automatic wire placement forming optimization system comprises: An input module is used to process the prepreg tow tension data input by the tension sensor through LSTM deep learning calculation and output dynamic tension control data; A calculation module, used for inputting the dynamic tension control data into a fuzzy neural network calculation engine, and obtaining a dynamic compensation model for the retraction and release speed after speed curve analysis and calculation; A calculation module, for calculating the multi-tow cutting sequence based on the retraction and release speed dynamic compensation model using a reinforcement learning engine to generate an intelligent cutting instruction set; A construction module is used to construct a re-delivery trajectory mapping relationship based on the intelligent cutting instruction set, and form a re-delivery path matrix after iterative calculation by the digital twin engine; Establishing a module for establishing a temperature field distribution prediction model according to the retransmission path matrix, and constructing a temperature field dynamic control sequence after finite element analysis and calculation; The output module is used to substitute the temperature field dynamic control sequence into the multi-objective constraint equation, and output the placement strategy data package after deep reinforcement learning iterative optimization.

9. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the composite material automatic wire placement forming optimization method according to any one of claims 1 to 7 is implemented.

10. A wire laying machine, characterized in that: The wire laying machine comprises: one or more processors; A storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the composite material automatic wire placement optimization method as described in any one of claims 1-7.

Images