Real-time control method for carbon fiber prepreg resin impregnation based on machine learning

Abstract

The application discloses a real-time regulation method for carbon fiber prepreg resin infiltration degree based on machine learning and relates to the technical field of industrial control.The method comprises the following steps: acquiring multi-source data in the production of carbon fiber prepreg, including fiber characteristic data, braiding structure data and process parameter data; inputting the multi-source data into a pre-constructed double-channel coupled prediction model to obtain a predicted infiltration distribution and calculate a predicted infiltration degree, wherein the model comprises a parallel infiltration kinetics model and an anisotropic model; and taking the minimum deviation between the predicted infiltration degree and a target infiltration degree as the target, iteratively optimizing based on the process parameters, outputting a regulation instruction and adjusting the process parameters. In this way, the technical effects of accurate prediction and prospective intervention before defect formation are achieved, and the product qualification rate and production efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, and in particular to a method for real-time control of resin wetting of carbon fiber prepreg based on machine learning. Background Technology

[0002] Carbon fiber prepreg, as a core material in high-end equipment fields such as aerospace and rail transportation, directly determines the mechanical properties and service reliability of composite materials through its impregnation quality. During prepreg production, the resin impregnation behavior of the fiber fabric involves a complex dynamic process involving multi-scale and multi-physics coupling, including microflow within the fiber bundle, macroflow between bundles, and interfacial interactions between the resin and fiber surfaces. Traditional production control mainly relies on operator experience and offline sampling inspection, optimizing the impregnation effect by adjusting process parameters such as temperature, pressure, and traction speed. In recent years, with the development of intelligent manufacturing technology, some researchers have attempted to use data-driven models to predict the impregnation process or simulate resin flow behavior based on physical simulations, aiming to achieve optimized control of process parameters.

[0003] While this method achieves quantitative prediction of wettability to some extent, its model is purely data-driven. When the process window shifts or raw material batches are changed, the model's prediction accuracy drops sharply, making it difficult to meet stable production requirements. Furthermore, it cannot proactively intervene before defects form, resulting in difficulty in effectively controlling the scrap rate. Summary of the Invention

[0004] This invention provides a real-time control method for resin wettability of carbon fiber prepreg based on machine learning, which solves the technical problems of poor process responsiveness and lack of forward-looking process adjustment and optimization capabilities in the prior art, resulting in low production efficiency. It achieves accurate prediction and forward-looking intervention before defects are formed, thereby improving the technical effect of product qualification rate and production efficiency.

[0005] The present invention provides a machine learning-based method for real-time control of resin wettability in carbon fiber prepregs, comprising: Acquire multi-source data during the production process of carbon fiber prepreg, wherein the multi-source data includes a first type of data reflecting fiber characteristics, a second type of data reflecting the weaving structure, and process parameter data; The multi-source data is transmitted to a pre-constructed dual-channel coupled prediction model to obtain the predicted infiltration distribution of the coupled output, and the predicted infiltration degree is calculated based on the predicted infiltration distribution. The dual-channel coupled prediction model includes a machine learning-based infiltration dynamics model and an anisotropic mode model set in parallel. The control objective is to minimize the deviation between the predicted wettability and the target wettability. The process parameter data is used as the state, and the process parameter adjustment amount is used as the action for iterative optimization. Control commands are output, and the process parameters are adjusted according to the control commands.

[0006] In one feasible implementation, multi-source data is acquired during the carbon fiber prepreg production process. This multi-source data includes a first type of data reflecting fiber characteristics, a second type of data reflecting the weaving structure, and process parameter data, including: Microscopic property data are obtained from a fiber property database and used as the first type of data; The original image is acquired through an online detection device, and the original image is preprocessed and its features are extracted to obtain weaving characteristic data as the second type of data; The process parameter data, including at least the immersion temperature, immersion pressure, and immersion time, are obtained through a process parameter sensor.

[0007] In one feasible implementation, multi-source data is acquired during the carbon fiber prepreg production process. This multi-source data includes a first type of data reflecting fiber characteristics, a second type of data reflecting the weaving structure, and process parameter data, including: The first type of data includes at least the monofilament diameter, surface energy, sizing agent type code, and roughness; The second type of data includes at least the weave type code, weave angle, yarn density, and yarn undulation.

[0008] In one feasible implementation, the construction of the dual-channel coupled prediction model includes: Obtain sample immersion data, wherein the sample immersion data includes sample first type data, sample second type data, sample process parameter data and corresponding label data; A first machine learning model is initialized and constructed, and constrained by a prior infiltration mechanism model. The first machine learning model is trained using the first type of sample data, the sample process parameter data, and the corresponding label data, and the training result is the infiltration dynamics model. Initialize and construct a second machine learning model, and train the second machine learning model using the sample second type data and the corresponding label data to obtain the anisotropic pattern model; The infiltration dynamics model and the anisotropic mode model are connected in parallel to the infiltration distribution coupling layer based on the attention mechanism to obtain the dual-channel coupled prediction model and perform joint training.

[0009] In one feasible implementation, a first machine learning model is initialized and constructed, constrained by a prior infiltration mechanism model. The first machine learning model is trained using the first type of sample data, the sample process parameter data, and the corresponding label data. The training result is the infiltration dynamics model, including: Construct the first machine learning model, which includes an input layer, a hidden layer, and an output layer; The input to the first machine learning model is defined as the first type of data and the process parameter data, and the output is a one-dimensional distribution of the theoretical wetting depth along the fiber direction. Obtain a priori infiltration mechanism model and define a mechanism loss factor. Define the residual between the output of the first machine learning model and the sample infiltration data as a data loss factor and construct a first loss function. Combining the first loss function, the first type of sample data and the sample process parameter data are used as inputs, and the corresponding label data is used as supervision to construct and train the first machine learning model, and the trained first machine learning model is output as the immersion dynamics model.

[0010] In one feasible implementation, a second machine learning model is initialized and constructed, and then trained using the second type of sample data and the corresponding label data to obtain the anisotropic pattern model, including: A graph neural network is constructed as the second machine learning model, and the weaving structure in the second type of sample data is converted into graph data, wherein nodes correspond to yarn intersections, edges correspond to yarn segments, and node features and edge features are extracted from the second type of sample data. The output of the second machine learning model is defined as an anisotropic pattern characterizing the permeability components of the resin in several directions, wherein the anisotropic pattern is characterized as a permeability anisotropic coefficient matrix. Using the graph data as input, the corresponding anisotropic patterns of the samples in the label data as supervision, and minimizing the residuals of the anisotropic patterns as the objective, the second machine learning model is trained to obtain the anisotropic pattern model.

[0011] In one feasible implementation, the multi-source data is transmitted to a pre-built dual-channel coupled prediction model to obtain the predicted infiltration distribution of the coupled output, and the predicted infiltration degree is calculated based on the predicted infiltration distribution, including: The first type of data and the process parameter data are input into the wetting kinetics model to obtain the predicted one-dimensional theoretical wetting depth. The second type of data is input into the anisotropic pattern model to obtain the predicted anisotropic pattern; In the coupling layer of the dual-channel coupled prediction model, the predicted anisotropic mode is used as a correction factor to perform spatial weighted fusion on the predicted one-dimensional theoretical wetting depth to generate the predicted wetting distribution of the prepreg. The predicted infiltration distribution is statistically calculated to obtain the predicted infiltration level at the current time.

[0012] In one feasible implementation, the control objective is to minimize the deviation between the predicted wettability and the target wettability. Iterative optimization is performed using process parameter data as the state and process parameter adjustment amounts as actions. Control commands are output, and process parameters are adjusted according to these commands, including: Adjust the process parameter data according to the adjustment amount of the process parameter; The adjusted process parameter data is updated to the dual-channel coupled prediction model to obtain the updated predicted wettability; Calculate the deviation between the updated predicted wettability and the target wettability, and iteratively optimize the process parameter adjustment amount based on the deviation and the preset deviation target until the deviation target is met; The control command is output according to the adjustment amount of the process parameters.

[0013] In one feasible implementation, the acquisition of the second type of data further includes: The three-dimensional topography point cloud of the prepreg surface is acquired in real time by an online detection device, and the weaving angle, yarn undulation and weaving density are extracted and output as online weaving characteristic data. Obtain the prepreg inventory database and retrieve offline batch-specific weaving characteristic data; The second type of data is obtained by combining the online knitting characteristic data with the inherent knitting characteristic data of the batch.

[0014] This invention discloses a machine learning-based method for real-time control of resin wettability in carbon fiber prepregs. The method includes: collecting multi-source information during the production process of carbon fiber prepregs, including first-type data characterizing fiber properties, second-type data reflecting weaving structure characteristics, and corresponding process parameter data; inputting the multi-source information into a pre-constructed dual-channel coupled prediction model to obtain a coupled output wettability distribution prediction result, and calculating the predicted wettability based on this distribution. The dual-channel coupled prediction model includes a parallel-running machine learning wettability dynamics model and an anisotropic mode analysis model; minimizing the deviation between the predicted wettability and the set target wettability as the optimization objective, using process parameter data as system state variables, and adjusting process parameters as control actions for iterative optimization to generate control commands, and dynamically adjusting the production process parameters according to the control commands. This machine learning-based method for real-time control of resin wettability in carbon fiber prepregs achieves accurate prediction and proactive intervention before defects form, improving product qualification rate and production efficiency. Attached Figure Description

[0015] Figure 1 This is a schematic flowchart of the real-time control method for resin wettability of carbon fiber prepreg based on machine learning according to the present invention. Detailed Implementation

[0016] The above technical solutions will now be described in detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of them. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments used only to explain the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the drawings, not all of them.

[0017] Example, Figure 1 This is a flowchart illustrating the real-time control method for resin wettability of carbon fiber prepreg based on machine learning according to the present invention. The method includes: Acquire multi-source data during the production process of carbon fiber prepreg, wherein the multi-source data includes a first type of data reflecting fiber characteristics, a second type of data reflecting the weaving structure, and process parameter data.

[0018] In some embodiments, multi-source data is acquired during the production process of carbon fiber prepreg, wherein the multi-source data includes a first type of data reflecting fiber characteristics, a second type of data reflecting the weaving structure, and process parameter data, including: The first type of data includes at least the monofilament diameter, surface energy, sizing agent type code, and roughness; the second type of data includes at least the weaving type code, weaving angle, yarn density, and yarn undulation.

[0019] Specifically, multi-source data refers to a heterogeneous set of data originating from different detection methods and reflecting different physical dimensions in the carbon fiber prepreg production process. The first type of data reflecting fiber characteristics includes monofilament diameter, surface energy, sizing agent type code, and roughness. These data characterize the fiber's physicochemical properties and directly affect the resin's spreading and wetting behavior on the fiber surface. The second type of data includes weaving type code, weaving angle, yarn density, and yarn undulation. These data characterize the geometric topology of the fiber fabric and determine the directional differences in resin flow. Process parameter data includes at least wetting temperature, wetting pressure, and wetting time. These data characterize the external conditions applied during the production process and directly drive the resin flow and curing process.

[0020] For example, in a carbon fiber prepreg production line, the first type of data is obtained through a fiber characteristic database. Taking a batch of T700 grade carbon fiber as an example, upon warehousing, the monofilament diameter is measured to be 7 micrometers using a scanning electron microscope, the surface energy is measured to be 45 millijoules per square meter using a contact angle meter, the sizing agent type is coded as EP01 based on information provided by the supplier, and the surface roughness is measured to be 0.2 micrometers using an atomic force microscope. This data is stored in the database and loaded into the edge computing unit during the production of this batch. Secondly, the second type of data is obtained through an online detection device. A linear industrial camera is installed in front of the impregnation station to acquire images of the prepreg surface at a rate of 500 frames per second, with an image resolution of 2048 pixels × 1 pixel and a coverage width of 50 millimeters. The edge computing unit performs grayscale conversion and filtering denoising on the original images, and then uses a lightweight convolutional neural network to extract the weaving angle, yarn density, and yarn undulation. Simultaneously, it reads the warehousing code of the roll of fabric from the production management system and maps it to a plain weave type code. Finally, the third type of data is acquired in real time through process parameter sensors. A K-type thermocouple is installed at the coating head to measure the resin temperature, currently recorded at 85 degrees Celsius; a Kistler pressure sensor is installed at the hot press roller to measure the impregnation pressure, currently recorded at 0.3 MPa; the linear velocity is calculated to be 2 meters per minute using the traction roller encoder, and the impregnation time is calculated to be 30 seconds based on the 1-meter impregnation zone length. These three types of data are then timestamped to form a complete set of multi-source data samples for the current moment.

[0021] Through the above process, a comprehensive understanding of the multi-dimensional key factors affecting impregnation quality during carbon fiber prepreg production was achieved. Fiber characteristic data provides the microscopic physicochemical basis, braiding structure data provides the macroscopic geometric and topological constraints, and process parameter data provides the external driving conditions at the process scale. The simultaneous acquisition and fusion of these three types of data provides comprehensive input features for the subsequent dual-channel coupled prediction model.

[0022] In some embodiments, multi-source data is acquired during the production process of carbon fiber prepreg, wherein the multi-source data includes a first type of data reflecting fiber characteristics, a second type of data reflecting the weaving structure, and process parameter data, including: Microscopic characteristic data is obtained from a fiber characteristic database as the first type of data; original images are obtained through an online detection device, and the original images are preprocessed and feature extracted to obtain weaving characteristic data as the second type of data; process parameter data, including at least immersion temperature, immersion pressure and immersion time, are obtained through a process parameter sensor.

[0023] Specifically, the fiber property database is a data space in the enterprise-level material information management system used to store the inherent property data of carbon fibers. It can specifically cover microscopic characteristic parameters such as monofilament diameter, surface energy, sizing agent type code and roughness. The parameters in this database are obtained by the fiber manufacturer through standardized testing before leaving the factory and maintain batch consistency.

[0024] Specifically, online inspection devices refer to machine vision systems deployed on prepreg production lines, which may include high-resolution industrial cameras, linear light sources, and image acquisition cards to capture optical images of the fabric surface; preprocessing refers to noise reduction, enhancement, and normalization operations performed on the original images; and feature extraction is the process of parsing weaving geometric parameters from the images using computer vision algorithms.

[0025] Specifically, the process parameter sensors include temperature sensors (K-type thermocouples), pressure sensors (piezoresistive pressure transmitters), and time measuring devices (high-speed counters), which are used to quantify the thermodynamic and kinetic conditions in autoclave or roll forming processes. These process sensors are existing sensors embedded in the production equipment and can be accessed through corresponding data interfaces and protocols.

[0026] Specifically, structured acquisition of multi-source information can be achieved through the aforementioned three-level data channels: The first channel connects to the fiber characteristic database in the enterprise's ERP system, querying microscopic characteristic data using the fiber batch number as an index. For example, for carbon fiber with batch number T800-2024-03-15-B, the data such as single filament diameter of 5.4 micrometers, surface energy of 45.2 mJ / m², sizing agent type code EP-02 (corresponding to epoxy resin sizing agent), and surface roughness Ra of 0.8 micrometers can be retrieved. This data is already fixed before production and does not require real-time measurement. The second channel activates an online detection device, using a 5-megapixel CCD industrial camera to acquire fabric surface images at 30 frames per second. After removing salt-and-pepper noise by median filtering (kernel size 5×5), the raw images are further processed using the Canny edge detection algorithm. The yarn edge contour is captured, and the interlacing angle is identified using Hough transform. Texture features are calculated using gray-level co-occurrence matrix to analyze yarn density (198 warp threads / 10cm, 202 weft threads / 10cm). Yarn undulation is extracted based on Fourier transform spectrum analysis (the ratio of peak-to-trough height difference to wavelength is 0.18), and the final output is a weaving characteristic data vector. The third channel collects process parameters through a distributed sensor network. The immersion temperature is measured in real time by a thermocouple attached to the hot press plate (sampling frequency 100Hz, typical value 120℃), the immersion pressure is fed back by the hydraulic system pressure sensor (range 0-1MPa, accuracy ±0.5%FS, typical value 0.3MPa), and the immersion time is accumulated by a PLC high-speed counter based on encoder pulses (resolution 10ms, typical value 60 seconds). Taking a prepreg production line as an example, the system synchronously outputs the first type of data vector [5.4, 45.2, EP-02, 0.8], the second type of data vector [PL-01, 90, 200, 0.18] (PL-01 is plain weave code) and the process parameter vector [120, 0.3, 60]. The three sets of data are aligned with timestamps to form a complete multi-source data frame.

[0027] Through the above process, multi-scale information integration of fiber micro-properties, fabric macro-structure, and process environment parameters can be achieved, solving the problems of data lag and error accumulation caused by traditional manual input. Offline access to the fiber property database avoids repeated detection, automated processing of online visual inspection ensures the accuracy of weaving feature extraction, and real-time acquisition of process parameter sensors ensures the timeliness of state perception. The parallel architecture of the three-level data channels provides a standardized input interface for the dual-channel coupled prediction model, enabling the wetting dynamics model to calculate the theoretical wetting depth based on accurate fiber-process coupling conditions, while enabling the anisotropic mode model to analyze the flow resistance distribution based on accurate weaving geometry features, providing a high-fidelity state space representation for subsequent reinforcement learning control.

[0028] In some embodiments, the acquisition of the second type of data further includes: The three-dimensional topographic point cloud of the prepreg surface is acquired in real time by an online detection device, and the weaving angle, yarn undulation and weaving density are extracted and output as online weaving characteristic data; the prepreg database is acquired, and offline batch-specific weaving characteristic data is acquired; the online weaving characteristic data and the batch-specific weaving characteristic data are fused to obtain the second type of data.

[0029] Specifically, the three-dimensional topography point cloud refers to the set of discrete spatial coordinate points on the surface of the prepreg obtained through laser triangulation or structured light scanning technology; the weaving angle refers to the angle between the yarn direction and the length direction of the prepreg, reflecting the anisotropic orientation of the fabric; the yarn undulation refers to the amplitude of the yarn fluctuation in the thickness direction, characterizing the degree of surface unevenness caused by the weaving structure; and the weaving density refers to the number of warp or weft yarns per unit length, determining the density of the resin flow channels.

[0030] Specifically, online weaving characteristic data refers to the structural characteristics of the current roll material calculated based on real-time scanning data; the warehousing database refers to the enterprise resource planning system or production execution system database that stores the inherent parameters of each batch of fiber fabrics; batch inherent weaving characteristic data refers to static information such as fabric type, theoretical density, and specifications that have been measured and archived during the raw material warehousing stage; fusion is the process of aligning, verifying, and combining real-time online data with offline static data to form a second type of data with more complete information.

[0031] Through the above process, a deep fusion of online real-time perception of the second type of data and offline prior knowledge can be achieved. Specifically, the online 3D topographic point cloud provides the true geometric state of the current roll material, capturing dynamic structural changes caused by tension fluctuations, uneven roller pressure, etc., while the batch-specific characteristics provided by the database ensure the accurate introduction of static information such as fabric type and theoretical specifications. This fusion compensates for potential noise interference and local anomalies from single online measurements, while avoiding the disconnect from reality caused by relying solely on offline data. It provides anisotropic pattern models with complete and reliable input features (the second type of data), helping to improve the adaptability and prediction accuracy of the dual-channel coupled prediction model to changes in weaving structure.

[0032] The multi-source data is transmitted to a pre-constructed dual-channel coupled prediction model to obtain the predicted infiltration distribution of the coupled output, and the predicted infiltration degree is calculated based on the predicted infiltration distribution. The dual-channel coupled prediction model includes a machine learning-based infiltration dynamics model and an anisotropic mode model set in parallel.

[0033] Specifically, the dual-channel coupled prediction model is a hybrid neural network architecture that integrates data-driven and physical constraints. The main body of the model consists of two complementary sub-models operating in parallel, and information fusion is achieved through an attention mechanism.

[0034] Preferably, the impregnation kinetics model refers to the first machine learning model constructed based on the physical information neural network. This model takes fiber characteristics and process parameters as input, and outputs the one-dimensional theoretical spatiotemporal distribution of impregnation depth along the fiber direction by embedding prior mechanism constraints such as Darcy's law and capillary flow equation. That is, the impregnation depth in the direction where fiber impregnation is easiest.

[0035] Preferably, the anisotropic mode model refers to a second machine learning model built on a graph neural network. This model takes the braided structure data as input and learns the directional resistance effect of the yarn interlacing topology on resin flow through graph convolution. The goal is to output the corresponding permeability anisotropy coefficient matrix. This matrix can characterize the predicted wetting distribution mentioned above, which is the saturation spatial distribution field of the resin in the thickness direction and in-plane direction of the prepreg. Furthermore, the predicted wetting degree is the statistical integral of this distribution field, which characterizes the overall wetting quality. In other words, the coupled output fuses the one-dimensional dynamic prediction with the three-dimensional anisotropy correction through attention weighting to generate a high-dimensional wetting distribution that conforms to physical consistency.

[0036] Specifically, this step achieves multi-scale prediction of the wetting process through dual-channel parallel inference and coupling layer fusion: The first channel inputs the first type of data [single filament diameter 5.4 μm, surface energy 45.2 mJ / m², sizing agent code EP-02, roughness 0.8 μm] and process parameter data [temperature 120℃, pressure 0.3 MPa, duration 60 seconds] into the wetting kinetic model. Inside the model, the process conditions are mapped to resin viscosity (constrained by the Arrhenius equation, calculated as η = 0.85 Pa·s) and capillary pressure (constrained by the Washburn equation) through a fully connected layer (hidden layer neurons 256-128-64). The calculated Pc = 0.12 MPa was obtained. Solving the one-dimensional convection-diffusion equation, the theoretical wetting depth is parabolic along the fiber direction, with a maximum wetting depth of 8.5 mm (located in the central layer of the fabric), and a drop to 6.2 mm in the edge area due to rapid heat dissipation. The second channel inputs the second type of data [weaving type PL-01, angle 89.65°, warp density 198.6 threads / 10cm, weft density 200.7 threads / 10cm, undulation 0.25, porosity 18%] into the anisotropic mode model. The model constructs the weaving structure as graph data (400 nodes corresponding to 20×20 interlacing points, 760 edges corresponding to yarn segments), and performs 3 layers of graph convolution. The network (GCN, 128 hidden dimensions) learns message passing between nodes and outputs a 3×3 penetration anisotropy coefficient matrix K=diag(0.85, 0.72, 0.45), where Kxx=0.85 represents the warp flow resistance coefficient (normalized to the baseline value), Kyy=0.72 represents the weft resistance coefficient, and Kzz=0.45 represents the thickness resistance coefficient, reflecting the physical fact that plain weave yarns are most densely interwoven and have the greatest flow resistance in the thickness direction. In the coupling layer, a multi-head attention mechanism (8 heads, 64 dimensions) is used to use the anisotropy pattern matrix as a spatial correction factor to weightedly expand the one-dimensional theoretical penetration depth. After Softmax normalization, the force weights are assigned dynamic correction coefficients for different spatial locations to generate a three-dimensional predicted wetting distribution field (resolution 50×50×20 voxels). This distribution shows that the central region is fully wetting due to low meridional resistance (saturation 0.92), while the edge latitudinal intersection points form local high-resistivity areas due to lower Kyy (saturation 0.78). The thickness direction shows a gradient decay (surface 0.88 → bottom layer 0.65). Finally, the volume integral of the distribution field is used to calculate the predicted wetting degree. The overall wetting degree is 0.82 using a saturation-weighted average, or a conservative wetting degree of 0.65 is obtained using the minimum value of the thinnest wetting area. The output is selected according to the product standard.

[0037] Through the above process, a unified modeling of multi-scale wetting behavior, from microscopic fiber-resin interaction to macroscopic fabric structural effects, was achieved, solving the dual bottlenecks of poor physical consistency of single data models and long computation time of pure mechanism models. This provides high-fidelity, low-latency state feedback signals for subsequent control, enabling control decisions to be based on accurate prediction of future wetting quality rather than delayed detection.

[0038] In some embodiments, the construction of the dual-channel coupled prediction model includes: Acquire sample wetting data, wherein the sample wetting data includes sample first-class data, sample second-class data, sample process parameter data, and corresponding label data; initialize and construct a first machine learning model, and train the first machine learning model using the sample first-class data, the sample process parameter data, and the corresponding label data, constrained by a prior wetting mechanism model, and output the training result as the wetting dynamics model; initialize and construct a second machine learning model, and train the second machine learning model using the sample second-class data and the corresponding label data to obtain the anisotropic pattern model; connect the wetting dynamics model and the anisotropic pattern model in parallel to the wetting distribution coupling layer based on the attention mechanism to obtain the dual-channel coupled prediction model and perform joint training.

[0039] Specifically, the sample wetting data is a historical dataset used for model training, containing complete input features and corresponding true output labels. Among them, the first type of sample data refers to the fiber microstructure data collected in the historical batches, the second type of sample data refers to the weaving structure data collected in the historical batches, the sample process parameter data refers to the time series data of real-time process parameters recorded in the historical batches, and the label data refers to the true wetting degree and wetting distribution of the sample obtained through offline detection methods.

[0040] Specifically, the first machine learning model is used to simulate the wetting dynamics process; the prior wetting mechanism model refers to a mathematical model based on physicochemical principles that describes the resin flow behavior, such as Darcy's law and capillary flow equations; the second machine learning model is used to extract the anisotropic features of the braided structure, and in the preferred embodiment, a graph neural network is used; the wetting distribution coupling layer refers to a functional layer that uses an attention mechanism to weight and fuse the outputs of the two channels, which can adaptively learn the dependence weights of different spatial locations on the wetting dynamics features and anisotropic features.

[0041] For example, the above samples are divided into training set, validation set and test set in a ratio of 8:1:1.

[0042] Specifically, the trained infiltration dynamics model and the anisotropic mode model are connected in parallel, and their outputs are connected to an attention-based infiltration distribution coupling layer. The coupling layer first expands the one-dimensional distribution of the theoretical infiltration depth along the fiber direction output by the infiltration dynamics model to a two-dimensional space, serving as the query matrix; it then converts the permeability components output by the anisotropic mode model into a two-dimensional spatial distribution map, serving as the key and value matrices. The dependency weights of each spatial location on the dynamic features are calculated using scaled dot product attention, and the weighted summation yields the two-dimensional predicted infiltration distribution. A fully connected output layer follows the coupling layer, converting the infiltration distribution map into the final output. Then, using training set samples as input and corresponding labeled data as supervision, end-to-end training is performed using a joint loss function. For example, the joint loss function includes a mean squared error term for the predicted infiltration distribution and a regularization term for the two-channel output. The Adam optimizer is used for 50 training epochs with a learning rate of 0.0001 to obtain the final dual-channel coupled prediction model.

[0043] Through the above process, a complete dual-channel coupled prediction model was constructed. The prior wetting mechanism model provides physical constraints for the first machine learning model, ensuring its predictions conform to the fundamental laws of resin flow and avoiding the extrapolation risk at the edge of the process window inherent in purely data-driven models. The graph neural network architecture enables the second machine learning model to adapt to different weaving structures; when the fabric type changes, only the graph structure needs adjustment without retraining. The attention mechanism coupling layer achieves adaptive fusion of the two channel outputs, allowing the model to automatically weigh the contributions of kinetic and anisotropic factors to wetting based on the characteristics of different spatial locations. Joint training further optimizes the overall model's collaborative performance, providing a high-precision wetting prediction foundation for subsequent real-time control.

[0044] In some embodiments, a first machine learning model is initialized and constructed, and constrained by a prior infiltration mechanism model. The first machine learning model is trained using the first type of sample data, the sample process parameter data, and the corresponding label data, and the training result is output as the infiltration dynamics model, including: A first machine learning model comprising an input layer, a hidden layer, and an output layer is constructed. The input of the first machine learning model is defined as the first type of data and the process parameter data, and the output is a one-dimensional distribution of the theoretical wetting depth along the fiber direction. A prior wetting mechanism model is obtained and a mechanism loss factor is defined. The residual between the output of the first machine learning model and the sample wetting data is defined as a data loss factor, and a first loss function is constructed. Combining the first loss function, the first type of sample data and the sample process parameter data are used as inputs, and the corresponding label data is used as supervision to construct and train the first machine learning model, and the trained first machine learning model is output as the wetting dynamics model.

[0045] Specifically, the one-dimensional distribution refers to the sequence of wetting depth values ​​at different locations along the fiber length; the prior wetting mechanism model refers to the mathematical equations describing the resin flow behavior based on physicochemical principles, preferably equations based on Darcy's law and the law of conservation of mass; the mechanism loss factor refers to the degree of deviation between the model's predicted output and the constraints of the physical equations, calculated using automatic differentiation techniques; the data loss factor refers to the error between the model's predicted output and the true label, measured using mean square error; the first loss function is a weighted sum of the mechanism loss factor and the data loss factor, used to guide the model to simultaneously fit the data and satisfy physical laws.

[0046] Specifically, firstly, the network structure of the first machine learning model is constructed, defining a sequential model containing an input layer, three hidden layers, and an output layer. The number of nodes in the input layer is set to 28, corresponding to the concatenated vector of the first type of sample data (8 dimensions) and the sample process parameter data (20 dimensions). The first hidden layer has 64 nodes, using the tanh activation function; the second hidden layer has 64 nodes, using the tanh activation function; and the third hidden layer has 32 nodes, using the tanh activation function. The number of nodes in the output layer is set to 50, corresponding to the theoretical wetting depth value when the fiber length direction is discretized into 50 equidistant positions. The total number of parameters in the model is approximately 12,000 (lightweight design) to suit edge deployment.

[0047] Then, the input-output mapping relationship of the model is defined. The first type of sample data and the sample process parameter data are concatenated into an input vector x, which is input into the first machine learning model. After forward propagation, an output vector with dimension 50 is obtained, representing the theoretical impregnation depth at 50 discrete points along the fiber direction from the inlet to the outlet, in millimeters. For example, the predicted output of a sample may be [2.1, 2.3, 2.5, ..., 8.2, 8.1, 7.9], indicating that the resin impregnation depth gradually increases from 2.1 mm at the inlet to a peak of 8.2 mm in the middle and then decreases slightly.

[0048] Next, a priori infiltration mechanism model is obtained and a mechanism loss factor is defined. Based on Darcy's law and the law of mass conservation, a priori infiltration mechanism model is constructed. During model training, for each input sample, the residuals are calculated using automatic differentiation techniques. The mechanism loss factor is defined as the mean of the squared residuals of all samples.

[0049] Then, the first loss function is constructed as a weighted sum of the mechanism loss factor and the data loss factor, so that the model, during training, both strives to fit the real data and is constrained by the physical equations, ensuring that the prediction results do not violate the basic infiltration laws.

[0050] Finally, using the first type of data from the training set and the sample process parameters as input, and the corresponding label data as supervision, the Adam optimizer was used with an initial learning rate of 0.001, a batch size of 128, and 100 training epochs. During training, the first loss function value was calculated in each epoch, and the network weights were updated via backpropagation. After training, the model performance was evaluated on the validation set, and the trained model was output as an immersion dynamics model for subsequent dual-channel coupled prediction.

[0051] In some embodiments, initializing and constructing a second machine learning model, and training the second machine learning model using the sample second-class data and the corresponding label data to obtain the anisotropic pattern model, includes: A graph neural network is constructed as the second machine learning model, and the weaving structure in the second type of sample data is converted into graph data, where nodes correspond to yarn intersections and edges correspond to yarn segments. Node features and edge features are extracted from the second type of sample data. The output of the second machine learning model is defined as an anisotropic pattern characterizing the permeability components of resin in several directions, where the anisotropic pattern is characterized as a permeability anisotropic coefficient matrix. Using the graph data as input, the corresponding sample anisotropic pattern in the label data as supervision, and minimizing the anisotropic pattern residual as the objective, the second machine learning model is trained to obtain the anisotropic pattern model.

[0052] Specifically, yarn intersections refer to the physical contact points where warp and weft yarns intertwine in a weaving structure, which are key nodes in the resin flow path; yarn segments refer to the continuous yarn portions between adjacent intersections, forming the channel boundaries for resin flow in the fabric gaps; node characteristics include geometric descriptors such as local weaving angle, yarn density, and undulation at the intersection; edge characteristics include connection attributes such as the length, orientation, and cross-sectional shape of the yarn segments.

[0053] Specifically, the permeability anisotropy coefficient matrix is ​​a second-order tensor describing the difference in the permeability of fluid in different directions in porous media. It can be represented as a 3×3 symmetric matrix, with diagonal elements representing the permeability in the principal direction and off-diagonal elements representing the directional coupling effect. The anisotropic model residual refers to the difference between the permeability matrix predicted by the model and the experimental measurement or high-fidelity simulation value based on the Frobenius norm.

[0054] Specifically, this embodiment achieves end-to-end mapping from weave structure to flow resistance through graph structure representation and supervised learning: In the graph data construction stage, for plain weave samples (warp density 200 threads / 10cm, weft density 200 threads / 10cm), the fabric is discretized into a 20×20 interlacing point array, constructing a graph structure containing 400 nodes. Each node corresponds to a warp and weft interlacing point, and the node feature vector is 5-dimensional [local weave angle, warp yarn density, weft yarn density, local undulation, porosity]. Edges are divided into two categories: warp edges and weft edges. Warp edges connect adjacent warp yarn intersections in the same weft direction (380 in total), and edge features include an edge length of 0.5mm (corresponding to warp yarn spacing), an edge orientation of 0° (horizontal), and a yarn cross-section aspect ratio of 1.2; weft edges connect adjacent weft yarn intersections in the same warp direction (380 in total), and the edge orientation is 90° (vertical). In the model architecture stage, the following is adopted: Using GraphSAGE or GAT (Graph Attention Network) as the basic architecture, three graph convolutional layers are set (128 hidden dimensions, mean pooling as the aggregation function). Each layer is followed by ReLU activation and BatchNorm. The output layer embeds and aggregates nodes into a graph-level representation through a fully connected network (128→64→9), which is reshaped into a 3×3 permeability anisotropy coefficient matrix K, where Kxx represents the meridional permeability, Kyy represents the latitudinal permeability, Kzz represents the thickness direction permeability, and off-diagonal elements Kxy represent the shear coupling effect. During the training phase, the label data is obtained through experimental measurement or CFD simulation. The loss function is defined as the squared error of the Frobenius norm between the prediction matrix and the label matrix. The Adam optimizer (learning rate 0.001, batch size 32) is used to train for 200 epochs. The validation set error converges to within 5%, and the anisotropic pattern model is finally output.

[0055] Through the above process, an automated and differentiable mapping from weaving geometry parameters to flow resistance tensors is realized, which solves the contradiction between generalization and real-time performance in traditional anisotropic characterization methods based on empirical formulas or CFD simulations; and provides a spatial correction benchmark for wetting dynamics models.

[0056] In some embodiments, the multi-source data is transmitted to a pre-built dual-channel coupled prediction model to obtain the predicted infiltration distribution of the coupled output, and the predicted infiltration degree is calculated based on the predicted infiltration distribution, including: The first type of data and the process parameter data are input into the wetting kinetics model to obtain the predicted one-dimensional theoretical wetting depth; the second type of data is input into the anisotropic mode model to obtain the predicted anisotropic mode; in the coupling layer of the dual-channel coupled prediction model, the predicted anisotropic mode is used as a correction factor to perform spatial weighted fusion of the predicted one-dimensional theoretical wetting depth to generate the predicted wetting distribution of the prepreg; the predicted wetting distribution is statistically calculated to obtain the predicted wetting degree at the current time.

[0057] Specifically, spatial weighted fusion refers to using the permeability anisotropy coefficient as a weight to perform non-uniform correction on the theoretical wetting depth in the in-plane direction, reflecting the local flow differences caused by the weaving structure; predicted wetting distribution refers to the resin saturation field S at each position in the three-dimensional space of the prepreg, with a value range of 0 (completely dry) to 1 (completely saturated); statistical calculation refers to performing mathematical operations such as volume integration, mean extraction, or extreme value search on the distribution field to compress the field data into a scalar index.

[0058] Through the above process, multi-scale infiltration behavior prediction from macroscopic process parameters to microscopic spatial distribution was achieved, solving the problem that a single homogenization model cannot capture local defects induced by the braided structure.

[0059] The control objective is to minimize the deviation between the predicted wettability and the target wettability. The process parameter data is used as the state, and the process parameter adjustment amount is used as the action for iterative optimization. Control commands are output, and the process parameters are adjusted according to the control commands.

[0060] Specifically, process parameter adjustment amount refers to the numerical change in adjusting the current production process parameters, including the increase or decrease in wetting temperature, the increase or decrease in wetting pressure, etc.; state refers to the environmental information input to the decision, including the process parameter data at the current moment and the predicted wetting degree; action refers to the output decision variable, i.e. the aforementioned process parameter adjustment amount; iterative optimization refers to gradually approaching the optimal combination of process parameters by repeatedly executing the "prediction-decision-adjustment-reprediction" cycle.

[0061] In some embodiments, the control objective is to minimize the deviation between the predicted wettability and the target wettability. Iterative optimization is performed using process parameter data as the state and process parameter adjustment amounts as actions. Control commands are output, and process parameters are adjusted according to the control commands, including: Adjust the process parameter data according to the process parameter adjustment amount; update the adjusted process parameter data to the dual-channel coupled prediction model to obtain the updated predicted wettability; calculate the deviation between the updated predicted wettability and the target wettability, and iteratively optimize the process parameter adjustment amount according to the deviation and the preset deviation target until the deviation target is met; output the control command according to the process parameter adjustment amount.

[0062] Specifically, firstly, the process parameter data is adjusted according to the process parameter adjustment amount. Taking the real-time operation of a carbon fiber prepreg production line as an example, the process parameter status at the current time t is: impregnation temperature T = 82.5 degrees Celsius, impregnation pressure P = 0.28 MPa, and traction speed v = 2.2 meters per minute. The dual-channel coupled prediction model calculates a predicted impregnation degree of 92.3% based on the current multi-source data, which is lower than the target impregnation degree of 95%, with a deviation of -2.7%. The reinforcement learning decision model (preferably the Deep Deterministic Policy Gradient (DDPG) algorithm) receives the current state s. t =[T, P, v, 92.3%], after forward propagation, the output action a is calculated. t =[ΔT, ΔP, Δv].

[0063] Then, the updated process parameter data, along with the currently unchanged fiber characteristic data and weaving characteristic data, are re-input into the dual-channel coupled prediction model, and forward inference yields the updated predicted wettability I. pred_new Assume that the result of this calculation is I. pred_new =94.2%, an improvement of 1.9 percentage points compared to the previous 92.3%, but still below the target value of 95%. At this point, the process parameter adjustment is iteratively optimized based on the deviation and the preset deviation target. The current deviation is -0.8%, and the preset deviation target is within ±1.0%, so the requirement has not yet been met. Based on the new state, the action is output again, and the model prediction is updated again after adjustment. The deviation is obtained as +0.1%, which is within the preset target range, and the iterative optimization stops.

[0064] Finally, the process parameter adjustments are converted into control commands in real time and executed, while the changes in wettability are continuously monitored and predicted before entering the next control cycle.

[0065] In summary, the real-time control method for resin wettability of carbon fiber prepreg based on machine learning provided by this invention has the following technical effects: By collecting multi-source information during the carbon fiber prepreg production process, including first-type data characterizing fiber properties, second-type data reflecting braided structure characteristics, and corresponding process parameter data, this multi-source information is input into a pre-constructed dual-channel coupled prediction model. The coupled output predicts the wetting distribution, and the predicted wetting degree is calculated based on this distribution. The dual-channel coupled prediction model includes a parallel-running machine learning wetting dynamics model and an anisotropic mode analysis model. With minimizing the deviation between the predicted and target wetting degrees as the optimization objective, process parameter data is used as system state variables, and process parameter adjustments are used as control actions for iterative optimization to generate control commands. These commands are then used to dynamically adjust the production process parameters. This achieves accurate prediction and proactive intervention before defects occur, improving product qualification rate and production efficiency.

[0066] It should be understood that the embodiments disclosed in this invention and the above description enable those skilled in the art to implement this invention. However, this invention is not limited to the embodiments mentioned above. It should be understood that those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be included within the protection scope of this invention.

Claims

1. A method for real-time control of resin wettability in carbon fiber prepreg based on machine learning, characterized in that, include: Acquire multi-source data during the production process of carbon fiber prepreg, wherein the multi-source data includes a first type of data reflecting fiber characteristics, a second type of data reflecting the weaving structure, and process parameter data; The multi-source data is transmitted to a pre-constructed dual-channel coupled prediction model to obtain the predicted infiltration distribution of the coupled output, and the predicted infiltration degree is calculated based on the predicted infiltration distribution. The dual-channel coupled prediction model includes a machine learning-based infiltration dynamics model and an anisotropic mode model set in parallel. The control objective is to minimize the deviation between the predicted wettability and the target wettability. The process parameter data is used as the state, and the process parameter adjustment amount is used as the action for iterative optimization. Control commands are output, and the process parameters are adjusted according to the control commands.

2. The method for real-time control of resin wettability of carbon fiber prepreg based on machine learning as described in claim 1, characterized in that, Acquire multi-source data during the carbon fiber prepreg production process, wherein the multi-source data includes a first type of data reflecting fiber characteristics, a second type of data reflecting the weaving structure, and process parameter data, including: Microscopic property data are obtained from a fiber property database and used as the first type of data; The original image is acquired through an online detection device, and the original image is preprocessed and its features are extracted to obtain weaving characteristic data as the second type of data; The process parameter data, including at least the immersion temperature, immersion pressure, and immersion time, are obtained through a process parameter sensor.

3. The method for real-time control of resin wettability of carbon fiber prepreg based on machine learning as described in claim 1, characterized in that, Acquire multi-source data during the carbon fiber prepreg production process, wherein the multi-source data includes a first type of data reflecting fiber characteristics, a second type of data reflecting the weaving structure, and process parameter data, including: The first type of data includes at least the monofilament diameter, surface energy, sizing agent type code, and roughness; The second type of data includes at least the weave type code, weave angle, yarn density, and yarn undulation.

4. The method for real-time control of resin wettability of carbon fiber prepreg based on machine learning as described in claim 1, characterized in that, The construction of the dual-channel coupled prediction model includes: Obtain sample immersion data, wherein the sample immersion data includes sample first type data, sample second type data, sample process parameter data and corresponding label data; A first machine learning model is initialized and constructed, and constrained by a prior infiltration mechanism model. The first machine learning model is trained using the first type of sample data, the sample process parameter data, and the corresponding label data, and the training result is the infiltration dynamics model. Initialize and construct a second machine learning model, and train the second machine learning model using the second type of sample data and the corresponding label data to obtain the anisotropic pattern model; The infiltration dynamics model and the anisotropic mode model are connected in parallel to the infiltration distribution coupling layer based on the attention mechanism to obtain the dual-channel coupled prediction model and perform joint training.

5. The method for real-time control of resin wettability of carbon fiber prepreg based on machine learning as described in claim 4, characterized in that, A first machine learning model is initialized and constructed, constrained by a prior infiltration mechanism model. The first machine learning model is trained using the first type of sample data, the sample process parameter data, and the corresponding label data. The output training result is the infiltration dynamics model, including: Construct the first machine learning model, which includes an input layer, a hidden layer, and an output layer; The input to the first machine learning model is defined as the first type of data and the process parameter data, and the output is a one-dimensional distribution of the theoretical wetting depth along the fiber direction. Obtain a priori infiltration mechanism model and define a mechanism loss factor. Define the residual between the output of the first machine learning model and the sample infiltration data as a data loss factor and construct a first loss function. Combining the first loss function, the first type of sample data and the sample process parameter data are used as inputs, and the corresponding label data is used as supervision to construct and train the first machine learning model, and the trained first machine learning model is output as the immersion dynamics model.

6. The method for real-time control of resin wettability of carbon fiber prepreg based on machine learning as described in claim 4, characterized in that, Initialize and construct a second machine learning model, and train the second machine learning model using the second type of sample data and the corresponding label data to obtain the anisotropic pattern model, including: A graph neural network is constructed as the second machine learning model, and the weaving structure in the second type of sample data is converted into graph data, wherein nodes correspond to yarn intersections, edges correspond to yarn segments, and node features and edge features are extracted from the second type of sample data. The output of the second machine learning model is defined as an anisotropic pattern characterizing the permeability components of the resin in several directions, wherein the anisotropic pattern is characterized as a permeability anisotropic coefficient matrix. Using the graph data as input, the corresponding anisotropic patterns of the samples in the label data as supervision, and minimizing the residuals of the anisotropic patterns as the objective, the second machine learning model is trained to obtain the anisotropic pattern model.

7. The method for real-time control of resin wettability of carbon fiber prepreg based on machine learning as described in claim 1, characterized in that, The multi-source data is transmitted to a pre-constructed dual-channel coupled prediction model to obtain the predicted infiltration distribution of the coupled output, and the predicted infiltration degree is calculated based on the predicted infiltration distribution, including: The first type of data and the process parameter data are input into the wetting kinetics model to obtain the predicted one-dimensional theoretical wetting depth. The second type of data is input into the anisotropic pattern model to obtain the predicted anisotropic pattern; In the coupling layer of the dual-channel coupled prediction model, the predicted anisotropic mode is used as a correction factor to perform spatial weighted fusion on the predicted one-dimensional theoretical wetting depth to generate the predicted wetting distribution of the prepreg. The predicted infiltration distribution is statistically calculated to obtain the predicted infiltration level at the current time.

8. The method for real-time control of resin wettability of carbon fiber prepreg based on machine learning as described in claim 1, characterized in that, The control objective is to minimize the deviation between the predicted wettability and the target wettability. Iterative optimization is performed using process parameter data as the state and process parameter adjustment amounts as actions. Control commands are output, and process parameters are adjusted according to these commands, including: Adjust the process parameter data according to the adjustment amount of the process parameter; The adjusted process parameter data is updated to the dual-channel coupled prediction model to obtain the updated predicted wettability; Calculate the deviation between the updated predicted wettability and the target wettability, and iteratively optimize the process parameter adjustment amount based on the deviation and the preset deviation target until the deviation target is met; The control command is output according to the adjustment amount of the process parameters.

9. The method for real-time control of resin wettability of carbon fiber prepreg based on machine learning as described in claim 2, characterized in that, The acquisition of the second type of data also includes: The three-dimensional topography point cloud of the prepreg surface is acquired in real time by an online detection device, and the weaving angle, yarn undulation and weaving density are extracted and output as online weaving characteristic data. Obtain the prepreg inventory database and retrieve offline batch-specific weaving characteristic data; The second type of data is obtained by combining the online knitting characteristic data with the inherent knitting characteristic data of the batch.

Images