XY-axis fiber multilayer weaving density control system
By combining acoustic-optical co-sensing and graph neural network prediction with adaptive impedance control, the problems of fiber tension control lag and yarn breakage in existing weaving systems have been solved, and the uniformity of composite material layer density and impact resistance have been improved.
Patent Information
- Application Number
- CN202611204856.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-10
- Publication Date
- 2026-11-10
AI Technical Summary
Existing weaving systems suffer from response lag in controlling layer density and fiber tension, are unable to detect fiber slippage or micro-damage in real time, and traditional control methods are prone to yarn breakage and stress concentration inside the fabric, affecting the uniformity of layer density and impact resistance of composite materials.
By combining acoustic-optical collaborative sensing, topological isomorphic twin modeling, and adaptive impedance execution, the system monitors strain and micro-damage in real time during the weaving process through an embedded acoustic-optical collaborative sensing network, predicts the dynamic stiffness of the yarn interlacing points using a graph neural network, and regulates the mechanical impedance of the yarn feeding through adaptive impedance control, thereby achieving flexible closed-loop control.
It enables quantitative calculation and real-time control of the local dynamic stiffness at the yarn interlacing point, improving the calculation speed and stability of the weaving process, avoiding yarn breakage and stress concentration, and ensuring the uniformity of the composite material's layer density and impact resistance.
Smart Images

Figure CN123067115A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent equipment technology, specifically relating to an XY-axis fiber multilayer weaving layer density control system. Background Technology
[0002] In the fields of aerospace, defense, and new energy, three-dimensional (3D) woven composite preforms using high-performance fibers such as carbon fiber and aramid as substrates play an irreplaceable role. During the multi-layer weaving process of ultra-thick, irregularly shaped preforms, the uniformity of interlayer density and the interlacing state of XY-axis fibers and normal fibers (Z-axis) directly determine the key mechanical properties of the final composite component, such as its resistance to delamination and impact damage.
[0003] However, existing weaving systems face significant technical bottlenecks in controlling layer density and fiber tension. Traditional tension control relies primarily on tension sensors and PID closed-loop systems, which not only suffer from response lag but also fail to detect fiber slippage or micro-damage that has already occurred within the fabric. Finite element numerical simulation methods, which have emerged in recent years, are computationally intensive and simply cannot meet the real-time online control requirements of the weaving process. Meanwhile, purely data-driven neural network black-box models lack physical constraints, exhibiting extremely poor generalization ability and readily failing when faced with abrupt changes in complex weaving paths.
[0004] Even more critically, existing servo yarn feeding mechanisms rely on rigid control logic. In high-speed weaving, when fibers pass through dense interlacing points or complex contour boundaries, intense high-frequency friction and sudden tension changes occur. Traditional control methods attempt to counteract these sudden changes by rigidly adjusting the yarn feeding speed, which can easily lead to yarn breakage or stress concentration defects within the fabric, severely compromising the uniformity of the preform's layer density. Summary of the Invention
[0005] The present invention aims to at least partially solve the aforementioned technical problems. Therefore, the present invention aims to provide an XY-axis fiber multilayer weaving layer density control system, which achieves intelligent control of the layer density and internal stress of multilayer prefabricated structures through a combination of acousto-optic co-sensing, topological isomorphic twin modeling, and adaptive impedance execution.
[0006] The technical solution adopted in this invention is as follows:
[0007] An XY-axis fiber multilayer weaving layer density control system includes: a yarn supply unit for supplying fiber yarns required for weaving; a weaving execution mechanism for executing fiber yarn laying and interlayer weaving actions; a sensing module for real-time acquisition of multidimensional fabric state parameters during the weaving process; and a control module connecting the sensing module, the yarn supply unit, and the weaving execution mechanism. The sensing module includes an embedded acoustic-optical co-sensing network, which consists of fiber Bragg grating sensing yarns woven into the prefabricated body and an array of acoustic emission sensors deployed at the weaving boundary, used to synchronously acquire low-frequency strain distribution and high-frequency micro-damage acoustic wave energy within the prefabricated body. The control module... The block is configured to: run a graph neural network physical twin model based on topological isomorphic mapping, which maps the three-dimensional interwoven structure of the fabric to nodes and edges of the graph network; calculate the local dynamic stiffness at the yarn interlacing points in real time based on the low-frequency strain and high-frequency sound wave energy collected by the embedded acousto-optic co-sensing network, and update the edge weights of the graph neural network accordingly; predict the global transient stress transmission topology of the weaving network through the message passing mechanism of the graph neural network; generate an adaptive impedance matching command based on the predicted transient stress transmission topology and output it to the yarn supply unit; and achieve flexible closed-loop control of the weaving layer density by dynamically changing the equivalent mechanical impedance parameters of the yarn feeding.
[0008] Furthermore, the control module is further configured to process the data of the embedded acousto-optic co-sensing network through an acousto-optic co-sensing algorithm, wherein the center wavelength drift of the fiber Bragg grating sensing yarn is extracted as a low-frequency strain reference, and the acoustic event count and absolute energy of the acoustic emission sensor are extracted as high-frequency transient features. The two are fused to distinguish between elastic strain caused by macroscopic tension and inelastic micro-deformation caused by fiber slippage and compression.
[0009] Furthermore, in the graph neural network physical twin model based on topological isomorphic mapping, the nodes of the graph network correspond to the spatial interlacing points of XY-direction fibers and normal-direction bonded yarns, and the edges of the graph network correspond to the yarn segments between adjacent interlacing points.
[0010] Furthermore, the forward propagation process of each layer of the graph neural network is mathematically equivalent to the discrete transmission equation of stress load along the yarn axis and across node friction in the fiber interwoven topology.
[0011] Furthermore, the yarn supply unit includes multiple independently configured servo drivers, each of which has a built-in mechanical impedance control algorithm that can respectively exhibit programmable equivalent stiffness coefficient, equivalent damping coefficient, and equivalent mass parameter for the controlled yarn channel.
[0012] Furthermore, the adaptive impedance matching instruction refers to the following: when the graph neural network physical twin model predicts that a stress surge will occur in front of a certain yarn channel, the control module instructs the corresponding servo driver to reduce the equivalent stiffness coefficient and increase the equivalent damping coefficient, so that the yarn feeding mechanism exhibits high compliance to absorb frictional impact energy; when the prediction requires maintaining a high-precision dense geometric arrangement, it instructs to increase the equivalent stiffness coefficient to resist external disturbances.
[0013] Furthermore, the sensing module also includes an external non-contact visual sensing unit for capturing the macroscopic geometry, surface layer density, and yarn gap of the woven surface in real time.
[0014] Furthermore, the control module also uses the surface macroscopic geometric parameters obtained by the external non-contact visual sensing unit as dynamic boundary conditions for the graph neural network physical twin model to periodically calibrate the global node spatial coordinate positions of the graph network.
[0015] Furthermore, the control module adopts an event-triggered high-frequency impedance reconstruction strategy: when the local dynamic stiffness degradation calculated by the graph neural network is lower than a preset safety threshold, the system refreshes the impedance parameters periodically at a low frequency; when the local dynamic stiffness degradation exceeds the preset safety threshold, the system triggers high-frequency calculation, reconstructs in real time, and sends out the transient impedance parameter sequence.
[0016] Furthermore, the control module is further configured to use the deviation between the prediction results of the graph neural network and the actual measurement values of the subsequent sensing modules as an error signal to fine-tune the parameters of the graph neural network model online, thereby achieving adaptive learning for material batch differences or environmental changes.
[0017] The beneficial effects of this invention are as follows:
[0018] 1. This invention utilizes the low-frequency continuity of fiber optic sensing and the high-frequency transients of acoustic emission sensing, and through an acousto-optic co-processing algorithm, successfully decouples these two completely different mechanical behaviors. This enables the system to quantitatively calculate the local dynamic stiffness degradation coefficient of the yarn at the interlacing point for the first time, achieving a leap from perceiving phenomena to inverting the physical essence.
[0019] 2. This invention directly maps the physical structure of the fabric to the computational structure of a Graph Neural Network (GNN) and embeds the mechanical transfer equations into the message passing function of the GNN. This ensures that stress prediction follows the physical conservation laws and utilizes the parallel computing advantages of GNNs, increasing the deduction speed to the millisecond level, thus resolving the contradiction between computational speed and physical accuracy in the field of real-time control.
[0020] 3. This invention combines adaptive mechanical impedance control with the predictive capabilities of a GNN. The system no longer simply controls position but actively manages interactive forces. It can proactively adjust the mechanical characteristics of the yarn feeding mechanism from rigid links to flexible springs to mitigate rather than resist stress impacts.
[0021] 4. The technical points of this invention are not simply a combination of independent improvements, but constitute a tightly coupled, mutually causal closed-loop system. The local stiffness of the acoustic-optical co-sensing is the core driver of the dynamic evolution of the GNN model; the extremely fast prediction of the GNN is the only prerequisite for the adaptive impedance control to achieve forward-looking compliance; and the flexible execution of the impedance control, in turn, ensures the stability of the weaving process, providing a more reliable monitoring environment for acoustic-optical sensing. Attached Figure Description
[0022] Figure 1 This is a block diagram illustrating the topology and collaborative control principle of the XY-axis fiber multilayer weaving density control system of the present invention. Detailed Implementation
[0023] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of 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.
[0024] It should be understood that, and also noted, in the embodiments, the functions / actions may appear in a different order than those shown in the figures. For example, depending on the functions / actions involved, they may actually be performed substantially concurrently, or sometimes the two figures shown consecutively may be performed in reverse order.
[0025] like Figure 1 As shown, the XY-axis fiber multilayer weaving density control system of this invention is an intelligent equipment integrating advanced sensing, artificial intelligence computing, and compliant robotic execution technology. It is designed for high-quality, high-efficiency automated molding of large-thickness, complex-geometric composite material preforms required in aerospace, high-end equipment, and other fields. The system's working mechanism can be divided into a sensing layer, a twin computing layer, and a flexible execution layer, which work collaboratively through deep coupling. The following will elaborate on the implementation details of each physical component and core algorithm of the system.
[0026] The yarn supply unit in this embodiment provides multi-channel, independent, and highly dynamic response yarn tension and speed control for the system. It mainly consists of a yarn frame, a tension pretreatment device, and a core adaptive impedance servo drive array.
[0027] The yarn carrier employs a modular design, capable of accommodating hundreds of individual fiber bobbins, including XY-axis main reinforcing fibers (such as carbon fiber), Z-axis splicing fibers (such as aramid fiber), and special sensing fibers containing fiber Bragg gratings. Each bobbin is equipped with an independent rotational damper to provide initial, passive tension control.
[0028] After being drawn from the yarn registry, the yarn passes through an array of multiple independent drive units before entering the weaving actuator. Each independent drive unit consists of a high-bandwidth, low-inertia direct-drive frameless torque motor and a high-resolution absolute encoder (>24 bits) coaxially. The yarn is wound onto precision guide rollers directly connected to the torque motor.
[0029] Each torque motor is independently controlled by a high-performance digital servo driver. This driver integrates a field-programmable gate array (FPGA) and a multi-core ARM processor, providing powerful local computing capabilities. The solution to the second-order dynamic equations of the adaptive mechanical impedance control algorithm is directly embedded and fixed in the firmware of these servo drivers. This ensures that the equivalent stiffness (K0) of each yarn feeding channel is... d ), damping (B) d ) and inertia (M d The parameters can be dynamically written by the main control module at a frequency in the kilohertz range via a high-speed real-time industrial bus and take effect immediately. This distributed underlying control architecture is the hardware foundation for realizing the dynamic reconfiguration of yarn tension to be either compliant or rigid.
[0030] The weaving actuator is responsible for precisely laying the yarn to the designated location and completing the interlayer weaving. In this embodiment, the actuator consists of a six-degree-of-freedom industrial robot and a lightweight, multi-functional weaving head mounted at its end effector.
[0031] The six-degree-of-freedom industrial robot utilizes a long arm span and high repeatability (better than ±0.05mm). The robot body is mounted on high-precision linear guides to extend its working range and adapt to the manufacturing of large prefabricated structures. The robot's motion trajectory is planned and commands are issued in real-time by the main control module.
[0032] The multi-functional weaving head features a lightweight design and is made of high-strength aluminum alloy or composite materials, integrating multiple functional components:
[0033] Multi-channel yarn guide: The multi-channel yarn guide integrates dozens of independent ceramic yarn guides, each corresponding to different warp, weft, and Z-direction knotting channels, ensuring precise positioning of the yarn in space.
[0034] The micro beat-up mechanism uses a micro beat-up comb driven by a piezoelectric ceramic or a linear motor. It can quickly compact the weft yarn with extremely high frequency and precise force after each weft yarn is laid, ensuring dense layering.
[0035] Z-axis knotting needles, controlled by a miniature servo motor, are used to perform the piercing and locking actions of Z-axis yarns between multiple layers of fabric, forming an interlocking or orthogonal structure.
[0036] The weaving head also integrates an external non-contact vision sensing unit, including a high frame rate industrial camera and a line laser generator, whose optical axis is precisely calibrated to align with the formed area behind the weaving head.
[0037] The perception layer of this system adopts an integrated internal and external deployment approach, realizing full-dimensional monitoring from macroscopic morphology to microscopic stress.
[0038] Embedded acousto-optic co-sensing networks consist of fiber Bragg grating (FBG) sensing yarns and acoustic emission (AE) sensor arrays.
[0039] Fiber Bragg grating (FBG) sensing yarn is a special composite yarn with a single-mode optical fiber of extremely fine diameter (e.g., 50 micrometers) coated with a polyimide protective layer embedded at its center. Multiple FBGs with different reflection wavelengths are pre-etched at specific locations on the fiber using an ultraviolet laser. This sensing yarn is fed from a dedicated bobbin and can be selectively woven into key stress areas of the preform as weft or Z-direction knot yarn by a weaving head. The fiber's tail end is connected to a high-speed FBG demodulator via a fiber optic rotary connector. This demodulator can simultaneously calculate the wavelength shift of all gratings at frequencies in the kHz range, thereby obtaining time-series data of the internal strain.
[0040] The acoustic emission (AE) sensor array uses miniature (diameter less than 5 mm), wide-bandwidth (100 kHz - 2 MHz) piezoelectric ceramic AE sensors. These sensors are glued or piezoelectrically coupled to the inner surface of the weaving mold of the fixed preform, as well as to non-moving parts in the weaving head that have frictional contact with the yarn. The sensors are connected to a multi-channel, high-speed (sampling rate >10 MS / s) AE data acquisition card via a low-noise preamplifier and coaxial cable. By analyzing the differences in arrival times of multiple sensor signals, the AE event source can be located in three dimensions.
[0041] An external non-contact vision sensing unit is integrated into the weaving head. The camera and the line laser form a fixed triangulation geometry to acquire the three-dimensional point cloud of the fabric surface in real time, thereby calculating the macroscopic layer density, thickness and smoothness.
[0042] The control module's hardware adopts a master-slave, heterogeneous real-time computing architecture to balance complex AI computing with stringent real-time control requirements.
[0043] The control module's hardware platform comprises two core computers within the control cabinet. The real-time main controller utilizes a Linux operating system based on an Intel Core i9 processor and the PREEMPT_RT real-time patch, or a dedicated industrial real-time operating system. This controller is responsible for task scheduling, robot trajectory planning, EtherCAT bus master communication, and event triggering logic. All hard real-time tasks (such as servo loops and sensor synchronization) run here. The AI computing coprocessor is a server equipped with a high-performance NVIDIA RTX series or data center-grade A100 / H100 GPU. It connects to the main controller via a high-speed PCIe bus or 10 Gigabit Ethernet. This server is dedicated to running computationally intensive graph neural network (GNN) physical twin model inference (forward propagation) and online learning (backward propagation).
[0044] The control module's software adopts a layered and modular design. The bottom driver layer includes the robot controller interface, EtherCAT master protocol stack, FBG demodulator driver, AE acquisition card driver, and camera SDK. Based on DDS or a similar high-performance real-time publish / subscribe middleware, it is responsible for efficient, low-latency data exchange between the various software modules. The core algorithm layer runs on the AI coprocessor and includes the implementation of the GNN model (based on PyTorch Geometric or DGL libraries), the acousto-optic co-processor stiffness inversion algorithm module, and the online learning optimizer module. The application and decision-making layer runs on the real-time main controller and includes the adaptive impedance parameter mapping module, the event-triggered decision logic module, and the human-machine interface (HMI) software.
[0045] The core algorithm and mathematical model of this invention will be further described below.
[0046] 1. Algorithm for time alignment and dynamic stiffness degradation coefficient inversion of acousto-optic co-sensing data
[0047] In the high-speed weaving process of multi-layer preforms, relying solely on FBG sensors to measure strain cannot distinguish between elastic deformation and material yielding (slip / fracture). This system uses complementary high and low frequency acousto-optic signals for decoupling calculations.
[0048] First, the IEEE 1588 Precision Time Protocol (PTP) is used to achieve data alignment between the FBG demodulator and the AE acquisition card at the hardware level, with a time synchronization accuracy of less than 10 microseconds. Let the local macroscopic strain of the grating center wavelength drift conversion measured by the FBG at discrete time step t be: FBG(t); During the same period, the high-frequency acoustic emission voltage signal acquired by the AE array is V AE (τ), where τ is the high-frequency sampling time dimension.
[0049] The system performs absolute energy integration on the AE signal within a basic time window [t, t+Δt] to obtain the acoustic energy release characteristics:
[0050]
[0051] When the fiber interlacing points experience normal elastic tension FBG (t) increases proportionally, but E AE (t) Maintains an extremely low background noise level. When microfilament breakage or severe interlayer misalignment and slippage occurs within the fiber bundle, E AE (t) will exhibit a pulse-like surge.
[0052] Accordingly, a dimensionless damage index D(t)∈[0,1) is defined in the control module. It is calculated using a nonlinear mapping function based on historical energy accumulation.
[0053]
[0054] Where α is an empirical acoustic constant for the material. Subsequently, the system inverts the local dynamic stiffness K of the yarn segment where the interlacing node is located in real time. dynamic (t):
[0055] K dynamic (t)=K0·[1-D(t)]·f( FBG (t))
[0056] In the formula, K0 is the initial intrinsic linear stiffness of the yarn material; f( FBG This is a geometric correction function that takes into account the nonlinear tensile hardening effect of the yarn. This step dynamically corrects microscopic degradations that are imperceptible to purely physical mechanical models by introducing acoustic energy, providing accurate physical weights for subsequent graph networks.
[0057] 2. Isomorphic mapping matrix from 3D physical topology of fabric to Graph Neural Network (GNN) space and forward propagation mechanism
[0058] Traditional finite element analysis (FEA) for composite materials cannot achieve online computation when constructing models of thick precast bodies due to the explosion of mesh numbers. This invention changes the computational paradigm by directly and isomorphically mapping the woven structure to a graph structure G=(V,E).
[0059] Node set $V$: Represents the interlacing points of each warp, weft, and Z-direction knotting yarn in 3D space. Each node v i ∈V maintains a dynamic feature vector This includes the three-dimensional coordinates, velocity, and external contact force state of the point.
[0060] Edge set $E$: represents two adjacent intersecting points v i and v j A section of suspended yarn between them. Each edge e ij We maintain a weighted eigenvector, the core of which is the local dynamic stiffness K calculated in the previous step. dynamic And initial static tension.
[0061] Based on this graph structure, this invention constructs a special physical information message passing network. In the forward propagation of the l-th network layer, the mathematical operations for a node to send messages to its neighbors are strictly designed to follow Hooke's law and Coulomb's law of friction in a discrete form. The stress message m applied by node j to node i... ji Defined as:
[0062]
[0063] In the above formula, p represents the node spatial coordinate vector; L0 is the initial relaxation length of the edge; F is the direction vector; μ is the friction coefficient, F N This refers to the cross-node normal compressive force; This represents the relative sliding speed.
[0064] Subsequently, node i collects stress messages from all its neighbors and performs an update operation to obtain a new hidden state of the node (representing the node's stress and displacement trends at the next moment):
[0065]
[0066] The architecture of embedding mechanical constitutive equations into the message passing function of graph neural networks enables GNNs to not only have the ultra-high speed characteristics of parallel matrix computation in deep learning (three orders of magnitude faster than traditional FEA, with inference time of only a few milliseconds), but also ensure that the deduction results will never violate the conservation laws of fundamental physics and mechanics, providing computational power guarantee for ultra-fast dynamic control.
[0067] 3. Real-time fine-tuning mechanism for graph neural network parameters based on online prediction bias
[0068] The micro-friction coefficient and temperature and humidity sensitivity of fiber materials can drift across different batches. To address this, the GNN model in this embodiment retains a learnable compensation module composed of a multilayer perceptron (MLP) to fit unmodeled complex nonlinear inelastic mechanical behavior.
[0069] During system operation, the GNN model predicts the strain at a certain interleaving point at time t+1 at time t. predict (t+1). When the real time reaches t+1, the FBG sensor reads the true strain. real (t+1). The control module calculates the residual loss in real time:
[0070]
[0071] Where V sensor Let W be the set of nodes containing sensors, and R be a regularization term to prevent overfitting. The system runs a lightweight online gradient descent optimization algorithm (such as the online Adam algorithm) in the background to fine-tune the learnable weight matrix W of the GNN in real time. GNN This physical-driven, data-feedback-based online learning mechanism ensures the absolute accuracy of the digital twin model during the long-term weaving process.
[0072] 4. Mathematical Model and Parameter Mapping Logic of Adaptive Impedance Controller
[0073] After obtaining the global transient stress transfer topology predicted by the GNN, the actuator translates it into compliant mechanical motion. Traditional PID controllers aim to eliminate position or velocity errors at any cost, which are the main culprits of fiber tearing in rigid weaving.
[0074] This invention directly incorporates an impedance control law at the lowest level of the yarn supply servo driver, forcing the servo motor system to exhibit preset second-order mass-spring-damping system dynamic characteristics in response to external tension environments.
[0075] M d ·Δ +B d ·Δ +K d ·Δx=F env -F target
[0076] Where Δx is the yarn position deviation; M d B d K d These represent the target equivalent inertia, target equivalent damping, and target equivalent stiffness exhibited by the servo system, respectively; F env F represents the actual pull-out force experienced by the yarn in the fabric. target The expected tension of the system.
[0077] This control module establishes a dynamic parameter mapping logic, which maps the peak stress σ of the leading node predicted by the GNN. predict and stiffness degradation coefficient Dpredict Mapped to underlying impedance parameters:
[0078] When the GNN predicts that the preceding node experiences a surge in stress (σ) due to complex contour puncture, predict >σ threshold And D predict When there is an upward trend, the controller significantly reduces K through function mapping. d And appropriately increase B d At this point, the stiffness term on the left side of the equation is minimal, which means that only a very small external force deviation F is required. env -F target The servo motor will generate a large position yield Δx (i.e., active yielding of yarn release), and high damping B d It absorbs the oscillating impact energy of sudden changes, preventing the yarn from rebounding and breaking.
[0079] Conversely, when the GNN predicts a uniform mechanical distribution in the region, and the area array visual camera detects loose defects in the dense surface layer, the mapping logic significantly increases K. d (e.g., from 50 N / m to 5000 N / m instantaneously). At this point, the servo system becomes like a rigid linkage, forcibly dragging the yarn to a precise geometric coordinate position, resisting lateral disturbances of the weaving needle bed.
[0080] 5. Specific judgment logic and computing resource scheduling of the event triggering mechanism
[0081] In order to ensure high-frequency control performance without excessively consuming the computing resources of industrial control computers, this system abandons fixed time sampling period control and adopts an event-triggered mechanism instead.
[0082] During the stable two-dimensional base fabric weaving stage, the system sets the communication bus and servo refresh rate to a low base frequency (e.g., 20Hz). In this state, the GNN only performs low-frequency synchronous calculations, and the impedance parameter M... d B d K d It remains at its steady-state constant.
[0083] The control module contains a daemon process that monitors the dynamic stiffness degradation rate dD(t) / dt output by the acousto-optic sensing module and the maximum stress gradient predicted n steps in advance by the GNN every millisecond. A trigger function is defined as follows:
[0084] Γ(t) = w1·dD(t) / dt + w2·max v∈V ( σ predict )
[0085] When Γ(t) > Γ critical (where Γ) criticalWhen the set safety threshold is reached, the system determines that the weaving process is entering a dangerous state (e.g., high-density interlacing, variable cross-section edges). A system hardware interrupt is immediately triggered, forcibly increasing the communication bus cycle to a high-frequency mode of 1000Hz. At this time, the GNN model runs at full capacity, reconstructing the transient impedance parameter sequence in milliseconds and sending it to each axis servo to implement precise transient risk avoidance and stress mitigation actions. Once the dangerous region is passed, the trigger function value drops, and the system automatically returns to a low-frequency state to release computing power.
[0086] This invention, through meticulous construction and deployment of the system hardware, and combined with a series of algorithms based on underlying mathematical models, achieves a complete cross-disciplinary integration and reconstruction of composite material fracture mechanics, graph neural network topology calculation, and adaptive robot impedance control. This enables the system to completely solve the long-standing problems of uneven layer density and internal stress damage in traditional multi-layer weaving processes at the microscopic level, opening up a new technical path for the intelligent molding of high-end preforms.
[0087] This invention is not limited to the above-described optional embodiments. Anyone can derive other various forms of products under the guidance of this invention. However, regardless of any changes made in their shape or structure, any technical solution that falls within the scope of the claims of this invention shall be protected by this invention.
Claims
1. A multilayer weaving density control system for XY-axis fibers, characterized in that, include: Yarn supply unit, used to transport the fiber yarn required for weaving; Weaving actuator, used to perform the laying of fiber yarns and interlayer weaving operations; The sensing module is used to collect multi-dimensional state parameters of the fabric in real time during the weaving process; And a control module, which connects the sensing module, the yarn supply unit and the weaving actuator; The sensing module includes an embedded acousto-optic co-sensing network, which consists of fiber Bragg grating sensing yarn woven into the preform and an array of acoustic emission sensors deployed at the weaving boundary, used to synchronously acquire the low-frequency strain distribution and high-frequency micro-damage acoustic wave energy inside the preform. The control module is configured as follows: Run a graph neural network physical twin model based on topological isomorphism mapping, which maps the three-dimensional interwoven structure of the fabric to the nodes and edges of the graph network; Based on the low-frequency strain and high-frequency sound wave energy collected by the embedded acousto-optic co-sensing network, the local dynamic stiffness at the yarn interlacing point is calculated in real time, and the edge weights of the graph neural network are updated accordingly. The transient stress transmission topology of the global weaving network is predicted through the message passing mechanism of the graph neural network. Based on the predicted transient stress transmission topology, an adaptive impedance matching command is generated and output to the yarn supply unit. By dynamically changing the equivalent mechanical impedance parameter of the yarn feed, flexible closed-loop control of the weaving layer density is achieved.
2. The system according to claim 1, characterized in that, The control module is further configured to process the data of the embedded acousto-optic co-sensing network through an acousto-optic co-sensing algorithm. In this algorithm, the center wavelength drift of the fiber Bragg grating sensing yarn is extracted as a low-frequency strain reference, and the acoustic event count and absolute energy of the acoustic emission sensor are extracted as high-frequency transient features. The two are fused to distinguish between elastic strain caused by macroscopic tension and inelastic micro-deformation caused by fiber slippage and compression.
3. The system according to claim 1, characterized in that, In the graph neural network physical twin model based on topological isomorphic mapping, the nodes of the graph network correspond to the spatial interweaving points of XY-direction fibers and normal-direction bonded yarns, and the edges of the graph network correspond to the yarn segments between adjacent interweaving points.
4. The system according to claim 3, characterized in that, The forward propagation process of each layer of the graph neural network is mathematically equivalent to the discrete transmission equation of stress load along the yarn axis and across node friction in the interwoven fiber topology.
5. The system according to claim 1, characterized in that, The yarn supply unit includes multiple independently configured servo drivers, each of which has a built-in mechanical impedance control algorithm that can respectively exhibit programmable equivalent stiffness coefficient, equivalent damping coefficient and equivalent mass parameter for the controlled yarn channel.
6. The system according to claim 5, characterized in that, The adaptive impedance matching command refers to the following: when the graph neural network physical twin model predicts that a stress surge will occur in front of a certain yarn channel, the control module commands the corresponding servo driver to reduce the equivalent stiffness coefficient and increase the equivalent damping coefficient, so that the yarn feeding mechanism exhibits high compliance to absorb frictional impact energy; when the prediction requires maintaining a high-precision dense geometric arrangement, it commands the equivalent stiffness coefficient to increase to resist external disturbances.
7. The system according to claim 1, characterized in that, The sensing module also includes an external non-contact vision sensing unit for real-time capture of the macroscopic geometry of the woven surface, surface layer density, and yarn spacing.
8. The system according to claim 7, characterized in that, The control module also uses the surface macroscopic geometric parameters obtained by the external non-contact visual sensing unit as dynamic boundary conditions for the graph neural network physical twin model to periodically calibrate the global node spatial coordinate positions of the graph network.
9. The system according to claim 1, characterized in that, The control module adopts an event-triggered high-frequency impedance reconstruction strategy: when the local dynamic stiffness degradation calculated by the graph neural network is lower than the preset safety threshold, the system refreshes the impedance parameters periodically at a low frequency; when the local dynamic stiffness degradation exceeds the preset safety threshold, the system triggers high-frequency calculation, reconstructs and sends out the transient impedance parameter sequence in real time.
10. The system according to claim 1, characterized in that, The control module is further configured to use the deviation between the prediction results of the graph neural network and the actual measurement values of the subsequent sensing modules as an error signal to fine-tune the parameters of the graph neural network model online, thereby achieving adaptive learning for material batch differences or environmental changes.
