Multi-scale feature fusion composite damage mechanics behavior prediction method and system thereof
By constructing an adaptive graph neural network and dynamically updating node features and topology using multi-source data, the prediction error problem caused by fixed topology in composite material damage prediction is solved, achieving real-time and accurate damage prediction and task decision support.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HOHAI UNIV
- Filing Date
- 2026-04-29
- Publication Date
- 2026-07-07
Smart Images
Figure CN122113689B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of material damage prediction technology, specifically a method and system for predicting the damage mechanical behavior of composite materials by fusing multi-scale features. Background Technology
[0002] Currently, composite materials are widely used in high-end equipment due to their excellent specific strength, specific stiffness and designability. However, under long-term cyclic loading and complex service environment, various internal damages of composite materials will couple and evolve, leading to the degradation of macroscopic mechanical properties.
[0003] The computational cost of traditional finite element analysis for performance degradation increases exponentially with the model's degrees of freedom, making it difficult to meet the needs of online monitoring and real-time prediction in engineering sites. Therefore, a prediction framework based on microscopic images and graph neural networks has emerged. It constructs graph node features by extracting material structural parameters and learns the micro-macro mapping relationship by using a message passing mechanism.
[0004] However, the topology of existing graph neural networks is fixed and cannot adapt to the dynamic switching of damage modes under different load conditions. The prediction error increases significantly under complex load spectra. Moreover, the output of its prediction framework is only an assessment of the current damage state and cannot be combined with specific task load profiles to make task-oriented lifetime predictions. It lacks the ability to provide quantitative risk basis for task planning decisions. To address the above problems, existing technologies urgently need to be improved. Summary of the Invention
[0005] The purpose of this application is to provide a method and system for predicting the damage mechanical behavior of composite materials by fusing multi-scale features. It has the advantages of tracking the micro-damage evolution state, adaptively matching multi-mode damage switching, and quantitative task decision support, which can significantly improve the real-time performance, accuracy and proactive risk avoidance capability of composite material damage prediction.
[0006] The objective of this application can be achieved through the following technical solution: Firstly, a multi-scale feature fusion method for predicting the damage mechanical behavior of composite materials, comprising the following steps:
[0007] Simultaneously acquire initial microscopic images, real-time payload data, real-time acoustic emission data, and profile data to be executed within a preset monitoring period;
[0008] Microscopic feature parameters are extracted from the initial microscopic image to obtain an initial mapping dataset that characterizes the correspondence between microscopic feature parameters and macroscopic mechanical response. An initial graph neural network is then trained based on the initial mapping dataset.
[0009] The real-time acoustic emission data is input into a preset inversion network to output data representing the change in the evolution state of the corresponding microscopic feature parameters, and this data is then updated to the corresponding node in the initial graph neural network.
[0010] Based on the real-time payload data and the updated initial graph neural network, a topology adjustment instruction for the graph structure is obtained, and the initial graph neural network is reconstructed according to the topology adjustment instruction to generate an adaptive graph neural network;
[0011] The real-time load data is input into an adaptive graph neural network to obtain the current damage index and current residual strength of the composite material. The damage accumulation process is iteratively simulated in combination with the profile data to be executed, and residual strength evolution curve and load adjustment suggestions are generated.
[0012] Obtain the measured damage index of the composite material after implementing the load adjustment suggestion, and update the preset weight coefficients in the preset inversion network according to the difference between the measured damage index and the current damage index.
[0013] Secondly, the multi-scale feature fusion composite material damage mechanical behavior prediction system includes the following modules:
[0014] The data acquisition module is used to synchronously acquire initial microscopic images, real-time payload data, real-time acoustic emission data, and profile data to be executed within a preset monitoring period;
[0015] The feature extraction module is used to extract microscopic feature parameters from the initial microscopic image, obtain an initial mapping dataset that characterizes the correspondence between microscopic feature parameters and macroscopic mechanical response, and train an initial graph neural network based on the initial mapping dataset.
[0016] The feature update module is used to input the real-time acoustic emission data into a preset inversion network to output change data representing the evolution state of the corresponding microscopic feature parameters, and update the corresponding node in the initial graph neural network.
[0017] The topology adjustment module is used to obtain topology adjustment instructions for the graph structure based on the real-time load data and the updated initial graph neural network, and to reconstruct the initial graph neural network according to the topology adjustment instructions to generate an adaptive graph neural network.
[0018] The damage simulation process is used to input the real-time load data into an adaptive graph neural network to obtain the current damage index and current residual strength of the composite material, and to iteratively simulate its damage accumulation process in combination with the profile data to be executed, and generate residual strength evolution curves and load adjustment suggestions.
[0019] The feedback optimization module is used to obtain the measured damage index of the composite material after the load adjustment suggestion is executed, and update the preset weight coefficients in the preset inversion network according to the difference between the measured damage index and the current damage index.
[0020] Thirdly, a computer storage medium stores computer-executable instructions, which, when executed, implement the multi-scale feature fusion composite material damage mechanical behavior prediction method and system described in the first aspect.
[0021] Compared with the prior art, the beneficial effects of this application are:
[0022] This application solves the problems of declining prediction accuracy with service time, fixed graph structure that cannot adapt to load switching, and inability to provide quantitative support for mission decision-making by synchronously acquiring multi-source data, constructing an initial graph neural network and dynamically updating node features and adaptively adjusting graph topology, predicting residual strength through recursion and optimizing weight coefficients through closed loop. It can track micro-damage evolution in real time, adaptively match multi-mode damage switching, and provide quantitative support for mission decision-making. Attached Figure Description
[0023] Figure 1 This is a schematic diagram illustrating the steps of the multi-scale feature fusion composite material damage mechanical behavior prediction method of this application;
[0024] Figure 2 This is a schematic diagram of the module of the multi-scale feature fusion composite material damage mechanical behavior prediction system of this application. Detailed Implementation
[0025] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but only to illustrate selected embodiments of this application.
[0026] Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item has been defined in one figure, it does not need to be further defined and explained in subsequent figures. The terms "first", "second", etc. are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.
[0027] In traditional composite material damage prediction technology, the prediction framework based on microscopic images and graph neural networks has three defects: First, the node features are static: the model only collects an initial microscopic image and solidifies the node features once before service. However, under cyclic loading, the porosity of composite materials can increase from 0.8% to more than 2.5% within hundreds of hours, and the fiber breakage rate can increase by tens of times. These key microscopic state changes cannot be reflected by static node features, resulting in a systematic deviation between the model prediction value and the measured value that accumulates over the service time.
[0028] Secondly, the graph topology is fixed: the spatial resolution requirements of the graph structure are completely opposite in the damage diffusion stage and the damage localization stage. Fixed graph topology cannot meet the characterization requirements of both stages at the same time, and the prediction error increases under complex load spectra. Thirdly, the prediction results are disconnected from the task decision: the existing system can only output the current damage state assessment value and cannot answer the core task-oriented decision questions such as when the remaining strength first falls below the safety threshold when executing subsequent task profiles.
[0029] For example, in a health monitoring scenario of a carbon fiber reinforced composite vertical tail panel of a certain type of aircraft, after the aircraft performed multiple high-intensity maneuver flights, obvious matrix cracking and interface debonding occurred in local areas of the panel. The porosity increased from the initial 0.8% to 2.1%, the average pore morphology factor decreased from 0.72 to 0.43, and the equivalent stiffness decreased by about 7%. However, the existing graph neural network prediction model based on the initial image still uses the microstructure parameters extracted before service, and the output current residual strength is about 11% higher than the measured value, which seriously underestimates the degree of damage.
[0030] In this scenario, the graph neural network with fixed graph topology cannot provide sufficient spatial resolution near the critical section during the damage localization stage, further amplifying the prediction error. At the same time, the system cannot predict the remaining intensity evolution trend based on subsequent mission profiles, and mission planners cannot know whether the structural safety margin is sufficient in the next 6g peak overload mission, posing a risk of structural failure.
[0031] If the above problems are not solved, the accuracy of composite material damage prediction models will continue to decline over time, mission planning systems will be unable to obtain accurate structural health information, proactive risk intervention measures will be unable to respond in a timely manner, and ultimately service safety margins will be overestimated, high-risk missions will be wrongly approved, seriously threatening the service safety of equipment.
[0032] Therefore, such as Figure 1 As shown, this application provides a method for predicting the damage mechanical behavior of composite materials by fusing multi-scale features, including the following steps:
[0033] Simultaneously acquire initial microscopic images, real-time payload data, real-time acoustic emission data, and profile data to be executed within a preset monitoring period;
[0034] Microscopic feature parameters are extracted from the initial microscopic image to obtain an initial mapping dataset that characterizes the correspondence between microscopic feature parameters and macroscopic mechanical response. An initial graph neural network is then trained based on the initial mapping dataset.
[0035] The real-time acoustic emission data is input into a preset inversion network to output data representing the change in the evolution state of the corresponding microscopic feature parameters, and this data is then updated to the corresponding node in the initial graph neural network.
[0036] Based on the real-time payload data and the updated initial graph neural network, a topology adjustment instruction for the graph structure is obtained, and the initial graph neural network is reconstructed according to the topology adjustment instruction to generate an adaptive graph neural network;
[0037] The real-time load data is input into an adaptive graph neural network to obtain the current damage index and current residual strength of the composite material. The damage accumulation process is iteratively simulated in combination with the profile data to be executed, and residual strength evolution curve and load adjustment suggestions are generated.
[0038] Obtain the measured damage index of the composite material after implementing the load adjustment suggestion, and update the preset weight coefficients in the preset inversion network according to the difference between the measured damage index and the current damage index.
[0039] During the data acquisition step, the system simultaneously acquires initial microscopic images, real-time payload data, real-time acoustic emission data, and profile data to be executed. This step can be configured to be manually triggered by operators at the start of each preset monitoring cycle to initiate data acquisition from each sensor channel. For example, during a downtime maintenance window, maintenance personnel use a portable endoscopic micro-CT device to acquire cross-sectional images of composite materials, simultaneously reading the payload and acoustic emission data from the previous sortie stored in the data logger, and inputting the mission profile data for the next sortie.
[0040] As another implementation approach, the data acquisition step can be configured to synchronously collect multi-sensor data through periodic automatic triggering. For example, the system uses each flight sortie as a monitoring cycle, automatically triggering the data acquisition process after each flight to read payload data and acoustic emission data in real time, and acquiring the profile data to be executed. Furthermore, the data acquisition step can be configured to continuously record payload data and acoustic emission data at a high frequency during flight using real-time streaming acquisition, segmenting and archiving the data by flight sortie, and interfacing with the profile data from the mission planning system in real time.
[0041] In the initial graph neural network construction step, the system extracts microscopic feature parameters from the initial microscopic image, constructs a mapping dataset, and trains the graph neural network. This step can be configured to allow materials engineers to manually measure the fiber density and pore distribution in the initial microscopic image and estimate the fiber volume fraction and porosity as initial parameter inputs by manually extracting the microscopic feature parameters.
[0042] As an alternative implementation, this step can be configured to extract microscopic feature parameters using image processing algorithms (such as Otsu thresholding and connected component labeling), automatically distinguishing between fiber regions, matrix regions, and pore regions, and statistically analyzing the area ratio and shape parameters of each region. Further, this step can be configured to perform pixel-by-pixel precise segmentation of the scanning electron microscope image using a deep learning semantic segmentation model (such as U-Net or DeepLab), automatically extracting the fiber volume fraction and average pore morphology factor, and combining this with high-throughput simulation results from a parametric finite element model library to construct a high-quality initial mapping dataset, thereby training and generating an initial graph neural network.
[0043] In the node feature dynamic update step, the system inputs real-time acoustic emission data into a preset inversion network, outputs the evolution state changes of micro-feature parameters, and updates the graph neural network nodes. This step can be configured to allow acoustic emission experts to qualitatively determine the damage type and degree based on the changing trends of parameters such as ring count and peak frequency, and manually update the feature vectors of some key nodes by manually interpreting the acoustic emission data.
[0044] As another implementation, this step can be configured to parse acoustic emission data using preset rules. For example, when the peak frequency is below 150kHz, it can be identified as a matrix cracking event, and a fixed increment can be added to the porosity feature value of the corresponding node. Further, this step can be configured to map the time-frequency domain features (average energy, peak frequency, ring count) of the acoustic emission signal into evolution coefficients using a pre-trained inversion neural network, and to precisely update the feature vectors of the corresponding nodes in each stress concentration region of the graph neural network using preset attenuation factors and preset learning factors.
[0045] In the graph topology adaptive adjustment step, the system calculates the topology sensitivity coefficient of each node based on real-time load data and the updated graph neural network, and performs node splitting or merging operations. This step can be configured to refresh the graph topology structure at fixed-period reconstruction intervals, re-planning the graph structure based on the current node feature distribution.
[0046] As an alternative implementation, this step can be set to be threshold-triggered. When the feature vector change of any node is detected to exceed a preset threshold, a topology adjustment evaluation for that node's region is triggered to determine whether a split or merge operation needs to be performed. Further, this step can be configured to use a reinforcement learning-based intelligent decision-making mechanism to calculate the topology sensitivity coefficient of each node. When the load adjustment coefficient exceeds a preset splitting threshold, a node splitting operation is performed; when the load adjustment coefficient is below a preset merging threshold, a node merging operation is performed. The edge weight correction coefficients of the adjacency matrix are dynamically updated to generate an adaptive graph neural network that accurately matches the current damage evolution scale.
[0047] In the damage state prediction and task risk assessment step, the system calculates the current damage index and remaining strength based on an adaptive graph neural network, and recursively predicts the profile to be executed. This step can be configured to estimate the current damage index using a simple interpolation method, and estimate the current remaining strength of the composite material by interpolating the statistical mean of the node feature vectors with a preset strength degradation curve.
[0048] One implementation approach is to calculate the current damage index using forward inference via an adaptive graph neural network, approximating the profile data to be executed as several equivalent load blocks with representative load values, and estimating the decrease in residual intensity block by block. Further, this step can be configured to completely discretize the profile data to be executed into a set of load sequences with a preset sampling step size, using the current damage index as the initial state, and an adaptive graph neural network as the state transition equation, iteratively calculating the damage increment at each load step, accumulating the predicted damage index and predicted residual intensity at each time step, generating a complete residual intensity evolution curve, and automatically outputting a quantified load adjustment suggestion containing a load adjustment index.
[0049] In the feedback optimization step, the system obtains the measured damage index after implementing the load adjustment suggestion and updates the preset weight coefficients in the preset inversion network accordingly. This step can be configured to be verified manually by engineers who measure the actual damage area using ultrasound C-scan or X-CT after implementing the load adjustment suggestion and compare it with the predicted value manually, adjusting the weight parameters of the inversion network if necessary.
[0050] One implementation approach is to use batch offline updates to retrain and optimize the weight coefficients of the inversion network after accumulating measured data over several monitoring periods. Alternatively, this step can be configured to use online adaptive learning to calculate the prediction residuals in real-time within each monitoring period, and to incrementally update the weight coefficients of the inversion network's time-frequency domain features online using gradient update rules with pruning constraints, thus forming a closed-loop self-optimization mechanism for the entire prediction system.
[0051] It should be further explained that, in the specific implementation process, the process of simultaneously acquiring the initial microscopic image, real-time payload data, real-time acoustic emission data, and profile data to be executed within the preset monitoring period includes:
[0052] The initial microscopic image refers to two-dimensional or three-dimensional image data reflecting the internal microstructure of the composite material structure, acquired through high-resolution imaging technology at the beginning of a preset monitoring cycle (such as before a new component is put into use, after each maintenance, or at the start of each monitoring cycle). Its core function is to provide initial state information of the material's microstructure, serving as a benchmark for subsequent damage evolution and the basis for constructing the initial graph neural network.
[0053] The initial microscopic image typically contains the following information: 1) Fiber distribution characteristics: the cross-sectional shape, size, position coordinates, and spacing between fibers. 2) Matrix characteristics: the uniformity of the matrix material and the presence of initial micropores or microcracks. 3) Interface characteristics: the continuity of the fiber-matrix interface and the presence of initial debonding. 4) Pore characteristics: the location, size, shape, and distribution pattern of pores.
[0054] The real-time load data refers to time-series data continuously collected at a certain sampling frequency by sensors deployed at key parts of the composite material structure within a preset monitoring period, reflecting the external forces, deformations, and environmental conditions acting on the structure. Its core function is to provide the mechanical boundary conditions during the structure's service life, serving as the driving input for damage evolution analysis.
[0055] The real-time load data typically includes the following parameters: 1) Strain data: linear strain and shear strain along different directions, reflecting the degree of local deformation. 2) Stress data: principal stress and equivalent stress obtained through strain-stress conversion or direct measurement. 3) Temperature data: temperature of the structural surface or interior, as the mechanical properties of composite materials are temperature-dependent. 4) Acceleration / vibration data: for rotating or vibrating structures, reflecting dynamic load characteristics. 5) Displacement / deformation data: absolute or relative displacement of key points.
[0056] The real-time acoustic emission data refers to transient elastic wave signals generated by internal material damage (such as matrix cracking, fiber breakage, and interface debonding) that are continuously collected at high frequency through a piezoelectric sensor array arranged on the surface of the composite material structure within a preset monitoring period. Its core function is to act as a stethoscope for the evolution of microscopic damage, capturing the timing, location, and intensity information of damage events in real time.
[0057] The real-time acoustic emission data typically includes: 1) Waveform data: the raw voltage-time series recorded by each sensor channel, reflecting the time-domain characteristics of the acoustic emission event. 2) Feature parameters: simplified features extracted from the waveform, including amplitude, energy, ring count, rise time, duration, peak frequency, etc. 3) Location information: the spatial coordinates of the acoustic emission source calculated using the time difference of arrival from multiple sensors.
[0058] The profile data to be executed refers to the set of data obtained from the mission planning system at the end of or just before the start of the current preset monitoring cycle, describing the load-time history that the structure will experience over a future period. Its core function is to serve as input for mission remaining lifetime prediction, combining static remaining strength assessment with dynamic mission loads.
[0059] The profile data to be executed typically includes: 1) Load sequence: load steps arranged in chronological order, each load step containing stress state, temperature, and duration. 2) Environmental conditions: environmental parameters affecting material properties, such as humidity and concentration of corrosive media. 3) Mission identifier: mission number, mission type, and estimated takeoff / landing time. 4) Safety constraints: maximum allowable load and minimum residual strength requirements for the mission.
[0060] This application further proposes a specific process for generating the initial graph neural network: performing semantic segmentation on the initial microscopic image at the same time to identify the fibrous region and the pore region, and obtaining the pixel ratio of the fibrous region to the initial microscopic image as the fibrous volume fraction. The area of each individual pore in the pore region is obtained. and perimeter And obtain the average pore morphology factor. ;
[0061] Where n is the number of identified independent pores, k = 1, 2, ..., n, and the microscopic characteristic parameters include fiber volume fraction and average pore morphology factor. A representative volume element model is constructed based on these microscopic characteristic parameters and high-throughput solution is performed to obtain its macroscopic stiffness matrix parameters. and equivalent yield strength ;
[0062] Obtain the microscopic feature parameter vector With macroscopic mechanical response vector The mapping tuples are combined into mapping tuples to obtain multiple consecutive mapping tuples to form an initial mapping dataset. The mapping tuples are mapped to the feature vectors of the corresponding nodes in the preset graph neural network. The initial mapping dataset is used as the training set, and the preset graph neural network is iteratively optimized through a preset loss function until the preset loss function is less than or equal to a preset convergence threshold. The preset graph neural network at this point is used as the initial graph neural network.
[0063] The semantic segmentation module can employ the Otsu thresholding method based on traditional image processing. It directly distinguishes three types of regions—fibers (high grayscale), matrix (medium grayscale), and pores (low grayscale)—through multi-threshold segmentation based on the grayscale distribution characteristics of fibers (high grayscale), matrix (medium grayscale), and pores (low grayscale) in the scanning electron microscope (SEM) image. As another implementation approach, the semantic segmentation module can use fully convolutional neural networks such as U-Net, performing pixel-by-pixel classification of the image through an encoder-decoder structure. Transfer learning fine-tuning can be performed on 50-200 manually labeled samples, achieving a segmentation accuracy of over 95%. Furthermore, the semantic segmentation module can utilize advanced models such as SegFormer based on the Transformer architecture, combined with a multi-scale feature fusion strategy, to achieve sub-pixel-level accuracy in fiber-matrix-pore three-class classification for SEM images with a resolution of at least 1 μm / pixel, exhibiting significant advantages, particularly in the fine recognition of pore boundary contours.
[0064] Fiber volume fraction is the most critical microscopic parameter determining the longitudinal elastic modulus and axial strength of composite materials, and is also known as the average pore morphology factor. Statistical analysis was performed using the normalized values of the area ratios of the isoperimeter inequalities for each independent pore. A value close to 1 indicates that the pores are circular, the stress concentration factor is approximately 3, and the effect on strength is relatively mild. A value close to 0 indicates that the pores are elongated cracks, with an end stress concentration factor exceeding 5, significantly increasing the risk of crack initiation. The significance of introducing the average pore morphology factor lies in the fact that, at the same porosity, the impact of pores dominated by spherical pores on material strength can differ by several times compared to pores dominated by flat cracks, and a single porosity parameter cannot distinguish this difference.
[0065] The representative volume element (RVE) parametric finite element model library is constructed as follows: in the microscopic feature parameter space ( ∈[0.45,0.65], Within the range [0.20, 0.95], several sets of parameter combinations (usually 200-500 sets) are generated using the Latin hypercube sampling method. For each set of parameters, a two-dimensional or three-dimensional RVE geometric model is automatically generated based on the statistical equivalence principle using finite element software (such as ABAQUS, ANSYS). Periodic boundary conditions are applied to the RVE boundary, and the macroscopic stiffness matrix parameters (9 independent components, reflecting the anisotropic stiffness of the material) and the equivalent yield strength (reflecting the plastic initiation point of the material) are obtained by solving through the homogenization method.
[0066] High-throughput simulations run several finite element method (FEM) tasks in parallel on a multi-core server, completing batch simulations of all parameter combinations within several hours and constructing a micro-macro mapping dataset covering the parameter space. Each record in this dataset is a mapping tuple, representing the microscopic feature parameter vectors. With macroscopic mechanical response vector The combination serves as the training sample for the feature vectors of the corresponding nodes in the graph neural network.
[0067] During the training of the graph neural network, the cross-section of the composite material is discretized into several representative sub-regions, each sub-region corresponding to a node in the graph. The initial feature vector of the node is extracted from that sub-region. The graph neural network consists of several components; adjacent sub-regions are connected by edges, and the edge weights are initialized based on the stress transfer relationship between the two nodes. The preset loss function uses mean squared error (MSE) or Huber loss, and the Adam optimizer iteratively updates the graph neural network parameters. The loss function converges to a preset threshold (typically 10) on the validation set. -3 After reaching a certain magnitude, training is stopped, and the current parameters are fixed as the weights of the initial graph neural network.
[0068] Through the above technical solutions, this application can automatically extract two complementary microstructure quantification parameters, fiber volume fraction and average pore morphology factor, from the initial microscopic image, overcoming the limitation that simply using porosity cannot distinguish the influence of different pore morphologies on strength. Through high-throughput simulation of a representative volume element parameterized finite element model library, a high-quality training dataset covering the parameter space is constructed, making the macroscopic mechanical response prediction error of the initial graph neural network significantly lower than that of the traditional analytical model. By mapping microscopic feature parameters to graph neural network node feature vectors, a dynamically updatable micro-macroscopic correlation carrier is established, laying a solid data foundation for subsequent acoustic emission inversion node updates and graph topology adaptive adjustments.
[0069] This application further proposes a specific process for updating the initial graph neural network: extracting time-frequency domain features from the real-time acoustic emission data, including average energy. Peak frequency Ringing count The time-frequency domain features are combined into an acoustic emission feature vector, which is then input into a pre-defined inversion network to obtain the evolution coefficients. ;
[0070] in, These are preset weighting coefficients corresponding to the time-frequency domain features, and the change data are... , The preset unit evolution vector is used to identify various stress concentration regions in the corresponding composite material structure based on the real-time load data.
[0071] Obtain the feature vector of a single stress concentration region in the initial graph neural network before the corresponding node is updated. And obtain the feature update increment of the corresponding node. , To obtain the updated feature vector of the corresponding node by setting a preset decay factor. , This is a preset learning factor.
[0072] The average energy reflects the total elastic strain energy released by the acoustic emission event. It is calculated by integrating the mean square value of the acoustic emission signal envelope over the duration and then normalizing it. Higher energy indicates a larger damage event and a more severe degree of microstructural degradation. The peak frequency is the frequency component with the largest amplitude in the acoustic emission signal spectrum and is highly correlated with the damage mechanism. Matrix cracking typically produces low-frequency components of 90–200 kHz, interface debonding typically produces mid-frequency components of 200–350 kHz, and fiber breakage typically produces high-frequency components of 350–500 kHz. The peak frequency can be used to preliminarily determine the dominant damage mode. The ringing count is the number of times the acoustic emission signal crosses a preset threshold voltage, reflecting the duration and cumulative activity of the acoustic emission event, and is positively correlated with the amount of accumulated damage.
[0073] The pre-defined inversion network can be based on three-dimensional acoustic emission feature vectors. Given a simple linearly weighted network as input, an activation function of logistic regression maps the input to evolution coefficients in the range (0,1). Parameters These are preset weighting coefficients corresponding to average energy, peak frequency, and ring count, respectively. The initial values can be determined by supervised learning fitting of acoustic emission data under several known damage evolution states.
[0074] As one implementation approach, the pre-defined inversion network can employ a multi-layer fully connected neural network, introducing a nonlinear activation function (such as ReLU) to nonlinearly map the three-dimensional time-frequency features and output evolution coefficients, thereby better capturing the complex nonlinear relationship between acoustic emission features and microstructural changes. Furthermore, the pre-defined inversion network can incorporate the time-difference localization results of the acoustic emission sensor array, including the spatial location information of acoustic emission events in the input features, to achieve spatially resolved updates of the damage evolution state in different regions, enabling precise spatial targeting of node feature updates.
[0075] Preset attenuation factor The introduction of this approach takes into account the propagation attenuation effect of acoustic emission signals in elastic media: the greater the distance between the acoustic emission source and the sensor, the greater the signal energy loss, the lower the effective signal quality, and the corresponding feature update amount of the node should be reduced accordingly. In practical applications, Calibration can be performed based on the acoustic emission signal attenuation rate of the composite material (typically 0.5~2dB / cm) and the distance relationship between the sensor deployment location and the node location.
[0076] Preset learning factors The step size of the state update is indicated by the feature vector of the control node in the current acoustic emission signal, ∈(0,1]. (e.g., 0.1~0.2) enables nodes to maintain a high historical state memory ability and has a strong resistance to single acoustic emission noise interference; a larger (e.g., 0.5~0.8) makes the node features more sensitive to sudden damage events, and can be adaptively configured according to the criticality of the structure and the reliability of the sensor.
[0077] Through the above technical solutions, this application establishes a quantitative inversion path from real-time acoustic emission signals to the evolution of microscopic feature parameters, enabling graph neural network node features to be dynamically updated based on real physical signals in each monitoring cycle. This overcomes the fundamental defect of existing methods where static node features lead to a continuous decline in prediction accuracy over service time. The introduction of preset attenuation factors and learning factors takes into account both the physical laws of signal propagation and the stability of feature updates, enabling node feature updates to respond to real damage events while suppressing occasional noise interference. The multidimensional design of the change data allows node features to track the independent evolution of fiber volume fraction and pore morphology factor, providing reliable node state information for subsequent accurate calculation of damage index.
[0078] This application further proposes a specific process for generating an adaptive graph neural network: obtaining the topological sensitivity coefficient of a single stress concentration region. ,in, These are the corresponding preset allocation coefficients. This represents the Euclidean distance between the feature vectors of the corresponding nodes before and after the update. For the preset distance threshold, This refers to the real-time load data at the corresponding node. The preset load threshold;
[0079] If the topology sensitivity coefficient of a single stress concentration region is greater than a preset splitting threshold, a node splitting command is output for that region. If the topology sensitivity coefficient of a single stress concentration region is less than a preset merging threshold, a node merging command is output for that region. No other adjustments are made. The preset splitting threshold is greater than the preset merging threshold. The topology adjustment command includes a node splitting command and a node merging command. For a node splitting command, the number of nodes in the corresponding stress concentration region is increased. For a node merging command, the number of nodes in the corresponding stress concentration region is decreased.
[0080] After executing the topology adjustment instruction, the stress gradient components between adjacent nodes are recalculated based on the added or reduced node positions. The stress gradient components are then converted into corresponding edge weight correction coefficients using a preset mapping function, and their adjacency matrix is updated. The initial graph neural network at this point is then used as an adaptive graph neural network.
[0081] Among them, the topology sensitivity coefficient is a key indicator that comprehensively characterizes the demand for local graph structure adjustment, and it is determined by two complementary components: the feature change component. The significance of state changes caused by acoustic emission inversion updates to node feature vectors within the current monitoring period was quantified. To update the Euclidean distance between the feature vectors of the preceding and following nodes, This component, representing a preset distance threshold, reflects whether the damage evolution rate exceeds the descriptive precision of the current map resolution; the load component... Incorporate the current local load level into the topology decision. This represents the current real-time load at this node location. A preset load threshold is used to ensure that image resolution is prioritized in high-stress areas. Preset allocation coefficient. The relative weights of the two components in the topology decision can be configured according to the relative importance of the damage evolution rate and load level in the application scenario.
[0082] The principle for setting the preset splitting threshold and preset merging threshold is as follows: the splitting threshold should be higher than the merging threshold, forming a dead zone between them. This ensures that the adjustment of the topology structure has a lag, avoiding graph structure oscillations caused by frequent fluctuations in the topology sensitivity coefficient near the threshold. The specific implementation of the node splitting operation is to divide the sub-region corresponding to the original node into two equal sub-regions according to the direction of the maximum current load gradient. Each sub-region inherits the feature vector of the original node and adds the local correction amount corresponding to its own acoustic emission data.
[0083] The edge connections between the two newly generated child nodes and between them and their original neighboring nodes are re-established according to the adjacency relationship of the sub-region space. The initial edge weights are obtained by transforming the stress gradient components between the corresponding sub-regions of the two nodes through a preset mapping function (such as a piecewise linear function or a Gaussian kernel function). The specific implementation of the node merging operation is to merge the two adjacent nodes with the lowest topological sensitivity coefficient and the most similar eigenvectors (maximum cosine similarity). After merging, the eigenvector of the node is the area-weighted average of the two nodes, and the edge connection relationship is the union of the edge sets of the two nodes, with duplicate edges removed.
[0084] Through the above technical solutions, this application establishes a dynamic graph structure adjustment mechanism based on topology sensitivity coefficient, enabling the spatial resolution of the graph neural network to adaptively optimize with changes in the damage evolution stage: In the early stage of damage diffusion, the ξ values of each node are generally low, the graph structure remains sparse, and the computational efficiency is high; after entering the damage localization stage, the ξ values of nodes near the critical section exceed the splitting threshold, automatically triggering node splitting to improve local resolution and accurately capture intensity gradients; in the stage where the damage tends to stabilize and the load is low, the ξ values of nodes in some areas are lower than the merging threshold, automatically performing node merging to reduce computational redundancy; the reinforcement learning-driven threshold adaptive optimization mechanism ensures the continuous optimization of the topology adjustment strategy throughout the entire life cycle, significantly outperforming the prediction performance of existing fixed graph structure methods.
[0085] This application further proposes a specific process for generating the residual intensity evolution curve: obtaining the real-time load data of each node in the adaptive graph neural network. eigenvectors under the action Let i = 1, 2, ..., N, where N is the number of nodes in the adaptive graph neural network, to obtain the current damage index of the composite material. ;in, The preset fusion coefficients of the i-th node in the adaptive graph neural network are used to obtain the current residual strength of the composite material based on the current loss exponent. , This represents the initial strength value of the corresponding composite material;
[0086] The profile data to be executed is discretized into a set of load sequences composed of multiple load steps according to a preset sampling step size, and the current damage index is... Using the adaptive graph neural network as the initial evolutionary state and the state transition equation, the damage increment at each subsequent time step is calculated iteratively. ;
[0087] in, This represents the mapping function of an adaptive graph neural network. Let t be the load step at time t. The graph topology at time t-1. Using preset model parameters, the predicted damage index at time t is obtained through cumulative calculation. and predicting residual intensity The predicted residual intensity at each subsequent time point is summarized to generate the residual intensity evolution curve.
[0088] The current damage index is calculated using a weighted fusion method based on the local damage level of each node: for each node, its feature vector and the current load data are input into the Sigmoid activation function, and the estimated local damage level at that node is output (range (0,1)); the preset fusion coefficient reflects the importance weight of the sub-region represented by the i-th node in the overall load-bearing capacity assessment of the composite material.
[0089] Nodes on the critical path bearing the main load transfer should be assigned higher weights, which can be pre-calibrated based on the stress distribution results given by finite element analysis. The current residual strength is based on the theory of continuous damage mechanics, defining the damage variable as the reduction ratio of the effective load-bearing area. =0 corresponds to a perfectly intact state ( ),when →1 corresponds to a near-complete failure state ( →0), The original tensile strength of the material can be determined by a material mechanical properties handbook or a standard tensile test.
[0090] The choice of discretization step size for the profile data to be executed requires a trade-off between prediction accuracy and computational efficiency: too large a step size can lead to the loss of details of load changes, especially the decrease in peak load, which may affect the accurate estimation of damage increment; too small a step size will significantly increase the number of iterations and prolong the computation time. For flight load profiles, it is recommended to use one discretization step size per flight, which is sufficient to capture the differences in peak overload across flights while maintaining high computational efficiency. In each iteration step, the graph topology state... It will automatically adjust as the damage state evolves (based on the aforementioned topology sensitivity coefficient mechanism) to ensure the continuous adaptation of the graph structure to changes in damage patterns during the damage accumulation process.
[0091] The adaptive graph neural network adopts a graph convolutional network architecture, and its mapping function... This is achieved through multi-level message passing: For the l-th level, the feature update of node i is:
[0092] ;
[0093] in, Let i be the set of neighbors of node i. and For learnable weight matrix, It is the ReLU activation function. Let i be the feature vector of node i in the l-th layer. Let i be the feature vector of the q-th neighbor node at layer l. After multiple layers of message passing, the node features are mapped to damage increments through the output layer. .
[0094] Through the above technical solutions, this application establishes a damage recursive prediction framework, which deeply integrates the current damage state with the specific task load profile, and realizes accurate prediction of the dynamic evolution of material bearing capacity throughout the entire task execution process. The weighted design of the preset fusion coefficients ensures that the damage index calculation fully considers the relative importance of each region to the overall structural safety, avoiding the underestimation of the impact of simple arithmetic average on local high-damage areas. The iterative update of the graph topology ensures the model's continuous adaptability to the dynamic switching of damage modes in long-term prediction, thereby generating a highly reliable prediction curve that can accurately reflect the evolution trend of residual strength, providing a reliable basis for the generation of subsequent load adjustment suggestions.
[0095] This application further proposes a specific process for generating load adjustment suggestions: taking the minimum value in the residual strength evolution curve... With preset safety threshold A comparison is performed, and when the minimum value is less than a preset safety threshold, an output including the load adjustment index is generated. The load adjustment recommendations include, , These are preset decision sensitivity factors.
[0096] The preset safety threshold is typically determined based on a comprehensive consideration of structural safety factor requirements, material strength dispersion coefficient, and service environment factors to ensure safety even when... At the corresponding residual strength level, the structure still has sufficient safety margin to withstand unexpected overloads. For composite material main load-bearing structures in civil aircraft, Typically, it is taken as 60% to 70% of the ultimate bearing capacity; load adjustment index The remaining intensity deficiency in (0,1) is reduced by an exponential decay function. Mapped to the load reduction ratio;
[0097] when When the intensity is low (the remaining intensity is only slightly below the safety threshold), A value close to 1 corresponds to a smaller load reduction, preserving more mission capabilities; when When the strength is relatively high (the remaining strength is significantly lower than the safety threshold). Approaching zero corresponds to a larger reduction in load, requiring more stringent protection of the structure. Decision-sensitive factors. Controlled Sensitivity to insufficient intensity: relatively high The value allows the system to provide significant load reduction recommendations even for small strength deficiencies, making it suitable for critical structures with high safety margin requirements; smaller The value allows for maintaining a high mission load with a certain degree of intensity margin consumption, making it suitable for mission scenarios where combat effectiveness is prioritized.
[0098] Through the above technical solutions, this application transforms the prediction results of the residual strength evolution curve into quantitative load adjustment suggestions with clear physical meaning and operability, achieving a key leap from knowing the danger to knowing how to avoid it; the exponential decay function design of the load adjustment index ensures that the adjustment suggestions are reasonably proportional to the degree of risk, avoiding excessive conservatism or radicalism in extreme cases of linear mapping; the configurability of the decision sensitivity factor κ enables the system to flexibly adapt to the safety margin requirements of different equipment platforms and mission types, balancing mission flexibility and structural safety; the entire load adjustment suggestion generation process forms a direct data interface with the mission planning system, realizing the deep integration of structural health management and mission planning.
[0099] This application further proposes a specific process for obtaining the measured damage index: taking the maximum value in the load sequence set as the peak load. According to the load adjustment index in the load adjustment recommendation The peak load is adjusted to obtain the adjusted peak load. The adjusted peak load is replaced at the corresponding load step in the load sequence set, and real-time load data is obtained when the updated load sequence set is executed. And the real-time payload data of each node in the adaptive graph neural network. eigenvectors under the action The measured damage index of the composite material was obtained. .
[0100] The measured damage index is an estimated damage state value obtained by re-collecting real-time load data under the current service state and inputting it into an adaptive graph neural network for forward inference calculation after the actual load adjustment recommendation has been implemented. Compared with the current damage index, the measured damage index reflects the true damage response of the material after actual load adjustment and is a key reference for evaluating the accuracy of the inversion network prediction. Real-time load data can be acquired in real time during the execution of the adjusted mission load or extracted from the actual load history of the corresponding sortie from the flight data recorder after the mission. The specific implementation of the peak load replacement operation is as follows: in the complete load sequence set, find the time t corresponding to the maximum load step, and replace the load step value at that time with... Keeping other load steps unchanged, an updated load sequence set is formed.
[0101] Through the above technical solutions, this application establishes a mechanism to quantitatively transform the actual execution results of load adjustment suggestions into the basis for inversion network correction, enabling the entire prediction system to form a complete data closed loop from prediction → suggestion → execution → verification; the peak load replacement operation after adjustment accurately restores the actual effect of task intervention, ensuring that the measured damage index and the predicted damage index are compared under the same load adjustment background, so that the residual has accurate physical meaning; the measured damage index is directly calculated by an adaptive graph neural network, ensuring the algorithmic consistency between the measured value and the predicted value, so that the residual truly reflects the interpretation error of the acoustic emission signal by the inversion network, providing a reliable correction signal for accurate weight coefficient updates.
[0102] This application further proposes a specific process for updating the preset weighting coefficients: [The measured damage index is then updated...] With the current damage index The difference is used as the prediction residual. The predicted residuals are used to update the preset weight coefficients of each time-frequency domain feature in the preset inversion network, thereby obtaining the updated preset weight coefficients. Where b = 1, 2, 3, These are the preset weighting coefficients corresponding to the time-frequency domain features before update. As a preset update factor, The minimum and maximum preset weight coefficients are used to apply the updated preset weight coefficients of the corresponding time-frequency domain features to the next preset monitoring cycle.
[0103] The prediction residual quantifies the systematic bias of the inversion network towards the damage state of the current monitoring period: a positive value indicates that the actual damage is higher than the predicted value, meaning the inversion network underestimates the severity of damage indicated by the acoustic emission signal, requiring an increase in the weighting coefficients of high-damage-contribution features (usually energy and ring count); a negative value, conversely, indicates that the inversion network overestimates the damage, requiring a decrease in the corresponding weighting coefficients. The preset update factor controls the update step size of the weighting coefficients within each monitoring period, and its setting requires a trade-off between rapid response to damage mode switching and maintaining the stability of historical data accumulation: a larger... (e.g., 0.05~0.1) makes the weighting coefficient highly sensitive to the current residual, suitable for scenarios with rapid damage evolution and frequent mode switching; smaller values... (e.g., 0.005~0.02) allows for smooth and gradual updates of the weighting coefficients, making it suitable for scenarios with slow damage evolution and high data noise.
[0104] Through the above technical solutions, this application establishes an online adaptive update mechanism for weight coefficients based on prediction residuals, enabling the preset inversion network to continuously track the evolution of composite material damage modes and dynamically adjust the weight allocation of acoustic emission time-frequency domain characteristics, overcoming the problem of decreased prediction accuracy when fixed weight coefficients switch damage modes. The introduction of pruning constraints ensures the stability of weight coefficient updates and prevents parameter divergence under extreme conditions. The configurability of preset update factors allows the system to flexibly adapt to the weight update requirements of different material systems and service conditions. The online update mechanism using the prediction residuals of each monitoring cycle as learning signals enables the entire prediction framework to form a complete data closed-loop self-optimization system, ensuring the continuous maintenance and gradual improvement of composite material damage prediction accuracy throughout the entire service life.
[0105] In another implementation, such as Figure 2 As shown, this application also provides a multi-scale feature fusion composite material damage mechanical behavior prediction system, including the following modules:
[0106] The data acquisition module is used to synchronously acquire initial microscopic images, real-time payload data, real-time acoustic emission data, and profile data to be executed within a preset monitoring period;
[0107] The feature extraction module is used to extract microscopic feature parameters from the initial microscopic image, obtain an initial mapping dataset that characterizes the correspondence between microscopic feature parameters and macroscopic mechanical response, and train an initial graph neural network based on the initial mapping dataset.
[0108] The feature update module is used to input the real-time acoustic emission data into a preset inversion network to output change data representing the evolution state of the corresponding microscopic feature parameters, and update the corresponding node in the initial graph neural network.
[0109] The topology adjustment module is used to obtain topology adjustment instructions for the graph structure based on the real-time load data and the updated initial graph neural network, and to reconstruct the initial graph neural network according to the topology adjustment instructions to generate an adaptive graph neural network.
[0110] The damage simulation process is used to input the real-time load data into an adaptive graph neural network to obtain the current damage index and current residual strength of the composite material, and to iteratively simulate its damage accumulation process in combination with the profile data to be executed, and generate residual strength evolution curves and load adjustment suggestions.
[0111] The feedback optimization module is used to obtain the measured damage index of the composite material after the load adjustment suggestion is executed, and update the preset weight coefficients in the preset inversion network according to the difference between the measured damage index and the current damage index.
[0112] In another embodiment, this application also provides a computer storage medium storing computer-executable instructions, which, when executed, implement the multi-scale feature fusion composite material damage mechanical behavior prediction method.
[0113] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.
Claims
1. A method for predicting the damage mechanical behavior of composite materials by fusing multi-scale features, characterized in that, Includes the following steps: Simultaneously acquire initial microscopic images, real-time payload data, real-time acoustic emission data, and profile data to be executed within a preset monitoring period; Microscopic feature parameters are extracted from the initial microscopic image to obtain an initial mapping dataset that characterizes the correspondence between microscopic feature parameters and macroscopic mechanical response. An initial graph neural network is then trained based on the initial mapping dataset. The real-time acoustic emission data is input into a preset inversion network to output data representing the change in the evolution state of the corresponding microscopic feature parameters, and this data is then updated to the corresponding node in the initial graph neural network. Based on the real-time payload data and the updated initial graph neural network, a topology adjustment instruction for the graph structure is obtained, and the initial graph neural network is reconstructed according to the topology adjustment instruction to generate an adaptive graph neural network; The real-time load data is input into an adaptive graph neural network to obtain the current damage index and current residual strength of the composite material. The damage accumulation process is iteratively simulated by combining the data of the profile to be executed, and residual strength evolution curve and load adjustment suggestions are generated. Obtain the measured damage index of the composite material after implementing the load adjustment suggestion, and update the preset weight coefficients in the preset inversion network according to the difference between the measured damage index and the current damage index.
2. The multi-scale feature fusion composite material damage mechanical behavior prediction method according to claim 1, characterized in that, The process of generating the initial graph neural network includes: Semantic segmentation is performed on the initial microscopic image at the same time point to identify fibrous regions and pore regions. The pixel ratio of the fibrous region to the initial microscopic image is obtained as the fiber volume fraction. The area of each individual pore in the pore region is obtained. and perimeter And obtain the average pore morphology factor. ; Where n is the number of identified independent pores, k = 1, 2, ..., n, and the microscopic characteristic parameters include fiber volume fraction and average pore morphology factor. A representative volume element model is constructed based on these microscopic characteristic parameters and high-throughput solution is performed to obtain its macroscopic stiffness matrix parameters. and equivalent yield strength ; Obtain the microscopic feature parameter vector With macroscopic mechanical response vector They are combined into mapping tuples to obtain multiple consecutive sets of mapping tuples to form an initial mapping dataset, and the mapping tuples are mapped to the feature vectors of the corresponding nodes in the preset graph neural network. The initial mapping dataset is used as the training set, and the preset graph neural network is iteratively optimized using a preset loss function until the preset loss function is less than or equal to a preset convergence threshold. The preset graph neural network at this point is then used as the initial graph neural network.
3. The method for predicting the damage mechanical behavior of composite materials by multi-scale feature fusion according to claim 1, characterized in that, The process of updating the initial graph neural network includes: Extract the time-frequency domain features from the real-time acoustic emission data, including the average energy. Peak frequency Ringing count The time-frequency domain features are combined into an acoustic emission feature vector, which is then input into a pre-defined inversion network to obtain the evolution coefficients. ; in, These are preset weighting coefficients corresponding to the time-frequency domain features, and the change data are... , The preset unit evolution vector is used to identify various stress concentration regions in the corresponding composite material structure based on the real-time load data. Obtain the feature vector of a single stress concentration region in the initial graph neural network before the corresponding node is updated. And obtain the feature update increment of the corresponding node. , To obtain the updated feature vector of the corresponding node by setting a preset decay factor. , Preset learning factors.
4. The multi-scale feature fusion composite material damage mechanical behavior prediction method according to claim 3, characterized in that, The process of generating an adaptive graphical neural network includes: Obtain the topology sensitivity coefficient of a single stress concentration region ,in, These are the corresponding preset allocation coefficients. This represents the Euclidean distance between the feature vectors of the corresponding nodes before and after the update. For the preset distance threshold, This refers to the real-time load data at the corresponding node. The preset load threshold; If the topology sensitivity coefficient of a single stress concentration region is greater than the preset splitting threshold, a node splitting command is output for that region; if the topology sensitivity coefficient of a single stress concentration region is less than the preset merging threshold, a node merging command is output for that region. No other adjustments are made. The preset splitting threshold is greater than the preset merging threshold. The topology adjustment instructions include node splitting instructions and node merging instructions. For node splitting instructions, the number of nodes in the corresponding stress concentration region is increased; for node merging instructions, the number of nodes in the corresponding stress concentration region is decreased. After executing the topology adjustment instruction, the stress gradient components between adjacent nodes are recalculated based on the added or reduced node positions. The stress gradient components are then converted into corresponding edge weight correction coefficients using a preset mapping function, and their adjacency matrix is updated. The initial graph neural network at this point is then used as an adaptive graph neural network.
5. The multi-scale feature fusion composite material damage mechanical behavior prediction method according to claim 3, characterized in that, The process of generating the residual intensity evolution curve includes: Obtain each node in the adaptive graph neural network in the real-time load data eigenvectors under the action Let i = 1, 2, ..., N, where N is the number of nodes in the adaptive graph neural network, to obtain the current damage index of the composite material. ; in, The preset fusion coefficients of the i-th node in the adaptive graph neural network are used to obtain the current residual strength of the composite material based on the current loss exponent. , This represents the initial strength value of the corresponding composite material; The profile data to be executed is discretized into a set of load sequences composed of multiple load steps according to a preset sampling step size, and the current damage index is... Using the adaptive graph neural network as the initial evolutionary state and the state transition equation, the damage increment at each subsequent time step is calculated iteratively. ; in, This represents the mapping function of an adaptive graph neural network. Let t be the load step at time t. The graph topology at time t-1. Using preset model parameters, the predicted damage index at time t is obtained through cumulative calculation. and predicting residual intensity The predicted residual intensity at each subsequent time point is summarized to generate the residual intensity evolution curve.
6. The multi-scale feature fusion composite material damage mechanical behavior prediction method according to claim 5, characterized in that, The process of generating load conditioning recommendations includes: The minimum value in the residual intensity evolution curve With preset safety threshold A comparison is performed, and when the minimum value is less than a preset safety threshold, an output including the load adjustment index is generated. The load adjustment recommendations include, , These are preset decision sensitivity factors.
7. The multi-scale feature fusion composite material damage mechanical behavior prediction method according to claim 6, characterized in that, The process of obtaining the measured damage index includes: The maximum value in the load sequence set is taken as the peak load. According to the load adjustment index in the load adjustment recommendation The peak load is adjusted to obtain the adjusted peak load. ; The adjusted peak load is replaced at the corresponding load step in the load sequence set, and real-time load data is obtained when the updated load sequence set is executed. And the real-time payload data of each node in the adaptive graph neural network. eigenvectors under the action The measured damage index of the composite material was obtained. .
8. The method for predicting the damage mechanical behavior of composite materials by multi-scale feature fusion according to claim 7, characterized in that, The process of updating the preset weight coefficients includes: The measured damage index With the current damage index The difference is used as the prediction residual. The predicted residuals are used to update the preset weight coefficients of each time-frequency domain feature in the preset inversion network, thereby obtaining the updated preset weight coefficients. ; Where b = 1, 2, 3, These are the preset weighting coefficients corresponding to the time-frequency domain features before update. As a preset update factor, The minimum and maximum preset weight coefficients are used to apply the updated preset weight coefficients of the corresponding time-frequency domain features to the next preset monitoring cycle.
9. A multi-scale feature fusion composite material damage mechanical behavior prediction system, characterized in that, Includes the following modules: The data acquisition module is used to synchronously acquire initial microscopic images, real-time payload data, real-time acoustic emission data, and profile data to be executed within a preset monitoring period; The feature extraction module is used to extract microscopic feature parameters from the initial microscopic image, obtain an initial mapping dataset that characterizes the correspondence between microscopic feature parameters and macroscopic mechanical response, and train an initial graph neural network based on the initial mapping dataset. The feature update module is used to input the real-time acoustic emission data into a preset inversion network to output change data representing the evolution state of the corresponding microscopic feature parameters, and update the corresponding node in the initial graph neural network. The topology adjustment module is used to obtain topology adjustment instructions for the graph structure based on the real-time load data and the updated initial graph neural network, and to reconstruct the initial graph neural network according to the topology adjustment instructions to generate an adaptive graph neural network. The damage simulation process is used to input the real-time load data into an adaptive graph neural network to obtain the current damage index and current residual strength of the composite material, and to iteratively simulate its damage accumulation process in combination with the profile data to be executed, and generate residual strength evolution curves and load adjustment suggestions. The feedback optimization module is used to obtain the measured damage index of the composite material after the load adjustment suggestion is executed, and update the preset weight coefficients in the preset inversion network according to the difference between the measured damage index and the current damage index.
10. A computer storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, they implement the multi-scale feature fusion composite material damage mechanical behavior prediction method as described in any one of claims 1-8.