A quality assessment method and system for composite panels

Abstract

This invention discloses a quality assessment method and system for composite plates, relating to the field of quality inspection technology. The method includes: constructing a latent variable feature space characterizing interface binding energy anomalies based on an initial response data field and the physical constraint relationship between local interface response features and binding energy distribution; extracting the topological correlation structure between local latent variable features within the latent variable feature space using a graph convolutional network to obtain initial candidate regions for local binding energy anomalies; performing active response mode enhancement excitation on the local regions of the initial candidate regions to obtain dynamic enhancement responses for each region and determining the binding energy anomaly region; and determining the specific location and range of hidden weld bands based on the difference in the correlation topological structure between local features within the binding energy anomaly region and the normal region. This invention achieves accurate identification and high-precision positioning of hidden weld bands at the interface of titanium-steel composite plates.

Description

Technical Field

[0001] This invention relates to the field of quality testing technology, specifically to a method and system for quality assessment of composite panels. Background Technology

[0002] Titanium-steel composite plates are widely used in aerospace, chemical equipment, shipbuilding, and other fields. Their interfacial bonding performance has a decisive impact on the safety and durability of the overall structure. However, in actual production, titanium-steel composite plates often have bonding defects at the interface, such as incomplete welds and weak bonding areas. The presence of these defects not only reduces the mechanical properties of the material but may also lead to premature failure or even catastrophic damage during service.

[0003] Traditional interface inspection methods typically rely on mechanical shear tests and ultrasonic testing, but these methods suffer from limitations such as limited detection accuracy, difficulty in accurately locating hidden microscopic defects, time and labor consumption, and lack of real-time monitoring capabilities. Furthermore, most existing numerical simulation and non-destructive testing techniques fail to effectively integrate the deep-seated correlation between interface mechanical response and physical properties, making it difficult to accurately reveal the true distribution and anomalous regions of interface bonding energy at the microscopic scale.

[0004] Therefore, there is an urgent need to propose a new method that can accurately identify and locate abnormal bonding areas at the interface of titanium-steel composite plates, in order to overcome the shortcomings of existing detection methods and improve the accuracy and reliability of interface bonding status assessment. Summary of the Invention

[0005] The purpose of this invention is to provide a quality assessment system and method for composite panels to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for quality assessment of composite panels, comprising:

[0008] S101: Apply a preset excitation field to the interface of the titanium-steel composite plate to obtain the initial response data field of the interface.

[0009] S102: Based on the initial response data field, a latent variable feature space characterizing interface binding energy anomalies is constructed through the physical constraint relationship between the interface local response characteristics and the binding energy distribution.

[0010] S103: Extract the topological correlation structure between local latent variable features in the latent variable feature space through graph convolutional networks to obtain the initial candidate region of local binding energy anomaly.

[0011] S104: Active response mode enhancement excitation is performed on the local area of ​​the initial candidate region to obtain the dynamic enhancement response of each local area, and the binding energy anomalous region is determined according to the rate of change of the dynamic enhancement response.

[0012] S105: Based on the differences in the correlation topology between the local characteristics of the binding energy anomaly region and the normal region, determine the specific location and range of the hidden weld band in the interface of the titanium-steel composite plate.

[0013] Secondly, the present invention provides a quality assessment system for composite panels, implemented based on the aforementioned quality assessment method for composite panels, comprising:

[0014] The excitation acquisition module is used to apply a preset form of excitation field to the interface of the titanium-steel composite plate and acquire the initial response data field of the interface.

[0015] The feature construction module is used to construct a latent variable feature space characterizing interface binding energy anomalies based on the initial response data field and through the physical constraint relationship between the interface local response features and the binding energy distribution.

[0016] The topology extraction module is used to extract the topological relationship structure between local latent variable features in the latent variable feature space through graph convolutional networks, so as to obtain the initial candidate region of local binding energy anomaly.

[0017] The dynamic determination module is used to perform active response mode enhancement excitation on the local area of ​​the initial candidate region to obtain the dynamic enhancement response of each local area, and to determine the binding energy anomalous region based on the rate of change of the dynamic enhancement response.

[0018] The cold weld location module is used to determine the specific location and range of the hidden cold weld band in the interface of the titanium-steel composite plate based on the difference in the correlation topology between the local characteristics of the abnormal bonding energy region and the normal region.

[0019] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0020] This invention employs finite element numerical simulation combined with multiphysics excitation methods, and utilizes a genetic algorithm to optimize and select the most sensitive excitation type in the local interface area. This constructs a high-precision initial data field that simultaneously includes displacement and stress responses, effectively overcoming the shortcomings of traditional detection methods in terms of single response information and insufficient accuracy. It significantly improves the quality and reliability of initial interface defect feature extraction, laying a solid data foundation for subsequent accurate identification of hidden interface defects.

[0021] This invention combines a variational autoencoder deep learning model with an explicit physical energy function to construct a latent variable feature space that can accurately reflect the abnormal characteristics of interface binding energy. It utilizes physical constraint regularization to achieve iterative optimization and precise correction of latent variable features, effectively solving the problems of insufficient feature representation ability and poor physical interpretability in traditional machine learning, making the identification and representation of interface micro-binding defects more accurate and reliable.

[0022] This invention uses latent variable feature topological correlation structure analysis based on graph convolutional networks, combined with dynamic enhancement excitation of real-time active response mode and topological structure difference cross-validation method, to accurately locate the specific location and range of the interface bonding energy anomaly region. In particular, it realizes the fine identification and spatial continuity definition of continuous hidden weld band region, which significantly improves the accuracy, real-time performance and engineering application value of titanium steel composite plate interface quality defect detection and monitoring. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0024] Figure 1 This is a flowchart of a quality assessment method for composite panels according to the present invention;

[0025] Figure 2 This is a framework diagram of a composite board quality assessment system according to the present invention. Detailed Implementation

[0026] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make the description of this application more complete and comprehensive, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The drawings are merely illustrative illustrations of this application and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0027] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more exemplary embodiments. Numerous specific details are provided in the following description to give a full understanding of the exemplary embodiments disclosed in this application. However, those skilled in the art will recognize that the technical solutions disclosed in this application can be practiced with one or more specific details omitted, or other methods, components, steps, etc., can be employed. In other instances, well-known structures, methods, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of the disclosure of this application.

[0028] Example 1

[0029] like Figure 1 As shown in the figure, this embodiment discloses a method for quality assessment of composite panels, including:

[0030] S101: Apply a preset excitation field to the interface of the titanium-steel composite plate to obtain the initial response data field of the interface.

[0031] In specific implementation, obtaining the initial response data field of the interface local area includes:

[0032] Based on the elastic modulus, density, and acoustic impedance distribution data of the titanium-steel composite plate, an initial finite element simulation model of the titanium-steel composite plate is established.

[0033] Specifically, the titanium-steel composite plate was first sampled and tested to obtain the elastic modulus distribution. Density distribution Acoustic impedance distribution The specific values ​​form a set of material parameters; among them, the elastic modulus Density can be determined by tensile testing or ultrasonic measurement. Acoustic impedance measured by weighing or density meter. It is obtained by direct measurement using an ultrasonic detector;

[0034] Subsequently, using finite element analysis software (such as ABAQUS and ANSYS), based on the actual geometric dimensions of the titanium-steel composite plate, an initial finite element simulation model of the titanium-steel composite plate was constructed according to the spatial distribution data obtained above. Specifically, taking an element size of 1 mm × 1 mm × 0.5 mm as an example, the geometric model of the composite plate was meshed, and the obtained material parameters were assigned to each corresponding mesh element to accurately reflect the mechanical properties of the real titanium-steel interface.

[0035] For example, define any unit The set of material properties is as follows:

[0036]

[0037] In the formula, i, j, and k represent the spatial location indices of the unit.

[0038] In the initial finite element simulation model, thermal field, ultrasonic stress wave and electromagnetic field excitations are applied respectively. The most sensitive excitation type is determined by the difference in displacement response distribution after the three excitations.

[0039] Specifically, determining the most sensitive stimulus type includes:

[0040] Using the interface bonding energy threshold of the normal bonding region and the virgin weld region as constraints, the displacement sensitivity matrix of each locality under different excitation types is calculated respectively.

[0041] Specifically, the interface binding energy threshold of the normal binding region is first defined. The interfacial bonding energy threshold of the poor solder joint region is Thresholds can be measured using standard interfacial peeling or shearing tests.

[0042] Subsequently, in the simulation model, unit thermal field excitation (temperature rise T=1℃) and unit ultrasonic stress wave excitation (stress amplitude) were applied to the interface respectively. Unit electromagnetic field excitation (magnetic field strength) ), calculate and record the displacement response values ​​of each element under each excitation;

[0043] Furthermore, the displacement sensitivity matrix S is constructed, with its specific elements defined as follows:

[0044]

[0045] in, This represents the displacement response value of the i-th element. This represents the unit applied intensity of the j-th type of excitation;

[0046] It should be noted that in the formula, index i∈[1,N] represents the number of the i-th spatial grid element in the composite plate, and N is the total number of elements; where index j∈{1,2,3} represents the correspondence between the three different excitation types applied:

[0047] .

[0048] With the goal of maximizing the entropy of the local binding energy difference of titanium-steel composite plates, a genetic algorithm is used to optimize the selection of the displacement sensitivity matrix and determine the most sensitive excitation type.

[0049] Specifically, the entropy value H of the interface binding energy difference is defined as:

[0050]

[0051] in, The proportion of the displacement sensitivity difference of element k to the sum of the sensitivity differences of all elements is calculated as follows:

[0052]

[0053] in, Let k be the displacement sensitivity value of the k-th element. The average value of the displacement sensitivity of all units;

[0054] Furthermore, a genetic algorithm is used to determine the combination of incentive types that maximizes entropy:

[0055] (1) Represent each type of stimulus as a chromosome using binary encoding;

[0056] (2) Randomly initialize the chromosome population;

[0057] (3) Calculate the entropy value corresponding to each chromosome as the fitness;

[0058] (4) Select chromosomes with fitness greater than the preset value for crossover and mutation;

[0059] (5) Repeat the iteration until the entropy value converges to determine the optimal excitation type.

[0060] The sensitivity matrix corresponding to the most sensitive excitation type is cross-feeded to the initial finite element simulation model. By re-correcting the boundary conditions of the simulation model, the most sensitive excitation type is updated and determined.

[0061] Specifically, the displacement sensitivity matrix corresponding to the optimal excitation type determined in the previous step is used as the feedback basis. The boundary load conditions (including position, direction of action and intensity) of the simulation model are adjusted according to the local area with high displacement sensitivity, and the displacement response is simulated and calculated again.

[0062] It should be noted that: through several feedback iterations, the excitation boundary conditions of the simulation model are optimized and stabilized, and the most sensitive excitation type is finally determined, resulting in a high-quality simulation response;

[0063] Based on the determined most sensitive excitation type, the initial displacement response of the interface locality under this excitation is extracted, and the initial displacement response field is established.

[0064] Specifically, the most sensitive excitation type (e.g., ultrasonic stress wave excitation) is determined with a practically feasible excitation intensity. (For example The excitation is applied to the finite element simulation model, and the initial displacement response values ​​of each interface local element under this excitation are obtained through numerical simulation.

[0065] Suppose any local element in the simulation model The initial displacement response is After integrating the displacement response data of all elements, the initial displacement response field is formed. ):

[0066]

[0067] It should be noted that the initial displacement response field is indexed by the coordinates of the element center, which ensures the spatial consistency of subsequent data processing and analysis.

[0068] Based on the initial displacement response field, the initial stress response field is inverted using the material mechanics equilibrium equations.

[0069] Specifically, using the elastic equilibrium differential equations in mechanics of materials:

[0070]

[0071] Under static or quasi-static conditions (ignoring the inertia term), the equation simplifies to:

[0072]

[0073] In the formula: For stress tensor; As physical exertion, it is usually negligible; Density; This represents the displacement field.

[0074] displacement field Substitute into the material constitutive equation (e.g., the constitutive equation for isotropic materials):

[0075]

[0076] in: , The Lamé constant of the material can be determined by the elastic modulus E and Poisson's ratio v, specifically:

[0077]

[0078] Where: I is the unit tensor; Let be the displacement gradient tensor.

[0079] The initial displacement response field and the initial stress response field are fused to form an initial response data field characterizing the local interface.

[0080] Specifically, the obtained initial displacement response field The initial stress response field obtained in step S101.4 By performing one-to-one spatial mapping and fusion, the initial response data of each interface local unit can be represented as:

[0081]

[0082] Furthermore, the set of response data from all local units is defined as the initial response data field. :

[0083]

[0084] It should be noted that this initial response data field contains both displacement and stress mechanical characteristics, providing a more comprehensive, objective, and reliable data foundation for subsequent anomaly detection.

[0085] S102: Based on the initial response data field, a latent variable feature space characterizing interface binding energy anomalies is constructed through the physical constraint relationship between the interface local response characteristics and the binding energy distribution.

[0086] In specific implementation, the construction of the latent variable feature space representing the interface binding energy anomaly includes:

[0087] Using the initial response data field of the interface locality as input, latent space features are extracted based on the variational autoencoder (VAE) model;

[0088] Specifically, this embodiment employs a variational autoencoder (VAE) model to extract low-dimensional latent space feature vectors that effectively represent the local response characteristics of the interface from the obtained initial response data field.

[0089] (1) Input data definition:

[0090] The obtained initial response data field The displacement-stress vector of each local element As the input feature vector of the VAE model;

[0091] (2) VAE model structure:

[0092] Encoder network structure:

[0093] Input layer: Input initial response feature vector ;

[0094] First hidden layer: 128 neurons, activation function ReLU;

[0095] Second hidden layer: 64 neurons, activation function ReLU;

[0096] Output layer: Generates the mean vector of latent variables with standard deviation vector All dimensions are 20.

[0097] The specific calculation formula is as follows:

[0098]

[0099]

[0100]

[0101] Where W and b represent the weight and bias coefficient matrices of each layer.

[0102] Decoder network structure:

[0103] Input layer: Input latent variable vector z;

[0104] First hidden layer: 64 neurons, activation function ReLU;

[0105] Second hidden layer: 128 neurons, activation function ReLU;

[0106] Output layer: Reconstruct the initial response data vector ;

[0107] The specific calculation formula is as follows:

[0108]

[0109]

[0110]

[0111] The meanings of the parameters in each layer are the same as those in the encoder network;

[0112] (3) Definition of loss function:

[0113] The loss function of the VAE in this embodiment includes two parts: reconstruction error and KL divergence.

[0114]

[0115] KL divergence is used to constrain the latent variable distribution to approximate a standard normal distribution. This is to ensure that the generated latent variable feature space has good continuity and expressiveness.

[0116] By utilizing the physical constraint relationship of the interface binding energy anomaly, the latent space features are physically constrained and regularized to generate latent variable features constrained by the interface binding energy distribution.

[0117] Specifically, the physical constraint regularization of the latent space features includes:

[0118] Based on the intrinsic physical relationship between interfacial bonding energy and strain energy density, a physical energy function for bonding energy is established; specific implementation method:

[0119] According to fracture mechanics theory, the interfacial bonding energy With local strain energy density The following explicit linear relationship exists between them:

[0120]

[0121] in, and The material interface property constants are determined through interfacial shear tests and numerical fitting.

[0122] Strain energy density Defined as the volume integral of the product of stress-strain tensors within a local element:

[0123]

[0124] in, For stress tensor, For the strain tensor, and “:” denotes the tensor inner product;

[0125] Specifically, strain tensor From displacement field The spatial gradient is derived as follows:

[0126]

[0127] Among them, displacement gradient Defined as:

[0128] ;

[0129] Using latent space characteristics as input, the strain energy density distribution of each locality is calculated through the physical energy function;

[0130] It should be noted that the multilayer perceptron network (MLP) used in this step is used to map the latent space features extracted by VAE to the strain energy density distribution, so as to achieve an effective connection between the latent space and the actual physical features.

[0131] The specific implementation method is as follows:

[0132] (1) Constructing a multilayer perceptron network structure for calculating strain energy density, specifically including:

[0133] Input layer: The dimension is the same as the latent space feature vector z (e.g., 20 dimensions);

[0134] Hidden layer: 128 neurons, ReLU activation function;

[0135] Output layer: Output strain energy density value A single neuron with no activation function (linear activation);

[0136] The specific calculation formula is as follows:

[0137]

[0138]

[0139] 2) The MLP network described above is trained under supervision using a large amount of experimentally tested sample data, with each sample inputting latent space features. The monitoring label is the true strain energy density value obtained through finite element calculation or experimental measurement. ;

[0140] The specific training loss function is defined as follows:

[0141]

[0142] In the formula, N represents the total number of samples used for training. The optimization process uses conventional gradient descent (such as the Adam algorithm) until convergence.

[0143] Based on the binding energy threshold condition, a physical relationship mapping is performed between the strain energy density distribution and the initial response field of the interface.

[0144] It should be understood that the core purpose of this step is to determine, by combining energy threshold conditions, whether there may be anomalies (such as cold solder joints or weak bonds) in the local interface area:

[0145] (1) Using the local strain energy density distribution obtained in the above steps And substitute it into the established physical energy function:

[0146]

[0147] (2) The predicted local binding energy value of the interface With a pre-set binding energy threshold A direct comparison is performed to determine the preliminary candidate criteria for anomaly regions as follows:

[0148] If the predicted binding energy value of a certain local cell satisfies:

[0149]

[0150] The cell is then initially marked as a candidate for anomaly region;

[0151] Based on the physical relationship mapping error, the latent space features are iteratively corrected by optimizing the regularization loss function to obtain latent variable features that meet the interface binding energy anomaly constraint.

[0152] The specific optimization process for this step includes:

[0153] (1) Define the overall optimization loss function This function integrates the VAE reconstruction error ( ) and physical relationship mapping error ( ):

[0154]

[0155] in: The combined loss of VAE network reconstruction error and KL divergence; To combine the error definition between the predicted value and the threshold, for example:

[0156]

[0157] Where M represents the number of candidate units in the abnormal region; λ is the regularization weight factor, with a value ranging from 0.1 to 1.0;

[0158] It should be noted that, in order to ensure a balanced accuracy in predicting abnormal and normal regions, a two-way penalty is chosen to improve the ability of latent space features to distinguish between the two types of regions.

[0159] (2) Using the latent space feature vector z as the optimization variable, the gradient descent method (such as the Adam optimizer) is used for optimization. The optimization update rule is as follows:

[0160]

[0161] in: The learning rate is 0.001 to 0.005; the iteration termination condition is that the loss function converges stably or the maximum number of iterations is reached (e.g., 50 to 100 times).

[0162] The interface local response data field is reconstructed in reverse using latent variable features, the reconstruction error distribution is calculated, and the latent variable features are corrected by cross-validation with the difference from the initial response data field.

[0163] Specifically, the detailed process for this step is as follows:

[0164] (1) The optimized latent variable features In the decoder network input to the VAE, the response data field is reconstructed in reverse:

[0165]

[0166] (2) Calculate the reconstruction error distribution, and define the reconstruction error as the element-wise Euclidean distance between the original response data field and the reconstructed data field:

[0167]

[0168] (3) Check whether the reconstruction error is within the allowable range by cross-validation (such as 5-fold cross-validation) (typical error threshold is set to within 0.05). If the error exceeds the threshold, return to the above optimization steps for further optimization.

[0169] The corrected latent variable features are used to form the final interface binding energy anomaly latent variable feature space; the specific implementation method is as follows:

[0170] Latent variable characteristics after optimization and cross-validation Integrating into a complete latent variable feature space :

[0171]

[0172] Each of them This represents the final latent variable features of the interface local unit, which are used for topological correlation analysis and accurate identification of abnormal regions in the next step.

[0173] S103: Extract the topological correlation structure between local latent variable features in the latent variable feature space through graph convolutional networks to obtain the initial candidate region of local binding energy anomaly.

[0174] In specific implementation, the step of extracting the topological association structure between local latent variable features within the latent variable feature space using a graph convolutional network includes:

[0175] Based on the latent variable characteristics of each interface in the latent variable feature space, the Euclidean distance and cosine similarity between nodes are determined, and an initial association graph is constructed.

[0176] The specific implementation method is as follows:

[0177] First, the latent variable feature space of the local interface is defined as follows: Among them, the latent variable characteristics of each interface local unit (node) The latent variable vector obtained in step S102 has a dimension of d=20;

[0178] Latent variable eigenvectors for any two nodes and Calculate their Euclidean distances respectively. Similarity to cosine :

[0179] The Euclidean distance calculation formula is as follows:

[0180]

[0181] Where: the latent variable eigenvector of node i A vector of dimension d=20 is represented as: ;in, This represents the component of the k-th dimension of the latent variable eigenvector of node i;

[0182] The formula for calculating cosine similarity is as follows:

[0183]

[0184] The calculation result ranges from [-1, 1]. The closer the value is to 1, the more similar the direction of the feature vectors between nodes are.

[0185] To construct the initial association graph, the criteria for determining the association edges are defined as simultaneously satisfying the following two threshold conditions:

[0186] Euclidean distance Less than the preset threshold (For example );

[0187] Cosine similarity Greater than the preset threshold (For example );

[0188] When the features of two nodes simultaneously satisfy the above conditions, an edge is established between the nodes; otherwise, no edge is established; thus, an initial association graph is obtained. , where the set of nodes For all local nodes of the interface, the edge set This is the set of associated edges determined based on the above criteria.

[0189] The initial association graph is input into a graph convolutional network to extract node feature topology.

[0190] It should be noted that the Graph Convolutional Network (GCN) used in this embodiment has a defined network structure to fully extract the topological information of the association graph. The specific network structure is defined as follows:

[0191] (1) Graph convolutional network structure:

[0192] The graph convolutional network used in this embodiment is a three-layer graph convolutional structure, and the propagation formula for each layer is as follows:

[0193]

[0194] in: Indicates the first Layer node feature matrix, layer 0 This is the initial node feature matrix, composed of the latent variable features of the nodes. constitute; This is the network weight matrix for the l-th layer, which needs to be obtained through training. , is the adjacency matrix of the graph. Add identity matrix The results take into account the characteristics of the nodes themselves; For a degree matrix, the diagonal elements ; For the activation function, this embodiment uses the conventional ReLU function.

[0195] The specific three-layer GCN network structure is as follows:

[0196] Input layer: Input initial feature matrix The dimension is the number of nodes N×d, where d=20;

[0197] First hidden layer: Output feature dimension is 64;

[0198] Second hidden layer: Output feature dimension is 32;

[0199] Output layer: The output feature dimension is 16, and the final topological feature representation matrix of the nodes is extracted. The dimension is N×16;

[0200] The input to the trained graph convolutional network is the adjacency matrix A of the initial association graph and the node features. The output is the node features enhanced with topological features. .

[0201] Specifically, the method for obtaining the graph convolutional network includes:

[0202] A large amount of titanium-steel composite plate sample data with known interface bonding states were collected, and the latent variable characteristics of the sample interface and the corresponding true distribution of bonding energy were obtained respectively.

[0203] Specifically, multiple titanium-steel composite plate samples with known interface bonding states were collected. Using the methods disclosed in steps S101-S102, the interface latent variable features and the true bonding energy state distribution of each sample were extracted, and the interface regions were explicitly labeled to obtain the training dataset.

[0204] Sample latent variable feature matrix: ;

[0205] Abnormal label for true binding energy of sample: ;

[0206] The label is defined as follows:

[0207] ;

[0208] A supervised loss function for determining binding energy anomalies is constructed, using the true binding energy anomaly distribution of the samples as the supervised label;

[0209] Specifically, the training objective is to enable the network output node features to accurately determine whether the local binding energy of the interface is abnormal, and the loss function adopts the conventional binary classification cross-entropy loss:

[0210]

[0211] in: For GCN-based output features The result of binary classification prediction (after Sigmoid activation).

[0212] Train the graph convolutional network until the decision error is less than a predetermined loss threshold, and alternately use training data and simulation data for bidirectional cross-validation;

[0213] In implementation, training data and simulation data are used alternately for bidirectional cross-validation, including:

[0214] The training dataset is randomly divided into k disjoint subsets (k=5 in this embodiment). In each cross-validation iteration, one subset is selected as the validation set, and the remaining k-1 subsets are used to train the model. The above process is repeated k times to ensure that each subset is used as the validation set once.

[0215] The goal of the cross-validation process is to achieve a prediction accuracy higher than a preset threshold (e.g., accuracy > 95%) and a decision error (loss function) lower than a predetermined loss threshold (e.g., loss < 0.05), ultimately obtaining a stable graph convolutional network model.

[0216] The training data ratio is dynamically adjusted based on the cross-validation results, and the network parameters are iteratively optimized to ultimately form a graph convolutional network with generalization capabilities.

[0217] To further improve the generalization performance of the network model, the ratio of positive to negative samples used for training is dynamically adjusted based on the results of each cross-validation. The specific method is as follows:

[0218] (1) If the cross-validation results show that the model’s accuracy in judging abnormal regions (positive samples) is lower than that in judging normal regions (negative samples), it indicates that there is insufficient training of positive samples. In the next round of training, the proportion of abnormal region samples should be increased (e.g., the ratio of positive to negative samples should be adjusted from the initial 1:1 to 1.5:1 or 2:1), and retraining should be performed.

[0219] (2) If the accuracy of the judgment is relatively balanced but the overall accuracy does not reach the preset target, the total number of training samples should be increased or the learning rate should be adjusted (e.g., from 0.005 to 0.001) until the performance meets the preset requirements.

[0220] The node feature topology is mapped inversely to the latent variable feature space, and the constraint strength of the topology features on the latent variable space is calculated.

[0221] In implementation, the constraint strength of topological features on the latent variable space is calculated, including:

[0222] The feature matrix of the output node from the trained graph convolutional network. In the process, the topological feature vector (16-dimensional) of each node is extracted.

[0223] The constraint strength of each node on other nodes in the latent variable feature space is calculated based on the topological feature vector. Specifically, the constraint strength is defined as the cosine similarity between topological features:

[0224]

[0225] In the formula: This represents the strength of the topological feature constraint of node i on node j. By calculating the constraint strength, the degree of topological association between nodes in the latent variable feature space can be quantitatively expressed.

[0226] The initial correlation graph is corrected based on the feedback of constraint strength differences to form a final correlation graph with topological feedback structure.

[0227] It should be noted that the goal of this step is to calculate the topological constraint strength matrix. The initial correlation graph is iteratively revised and its structure further optimized to more accurately identify regions with abnormal interface bonding energy. The specific implementation method is as follows:

[0228] 1) Define the topology feedback correction rules:

[0229] First, set a threshold for the strength of topological constraints. (Determined through experiments) for the initial association graph Check each connecting edge in the process:

[0230] If the topological constraint strength between any two nodes i and j is... Less than the threshold If the topological constraint strength is greater than or equal to the threshold, then the edge connecting nodes i and j is removed from the association graph; otherwise, if the topological constraint strength is greater than or equal to the threshold, then the edge connection is retained.

[0231] (2) Using the above correction rules, the initial association graph is modified. Perform a complete edge correction to obtain the intermediate association graph. ;

[0232] (3) Further feedback iteration:

[0233] The revised intermediate relationship diagram The graph convolutional network is re-inputted, and the topological feature extraction and constraint strength calculation process is executed again to generate a new constraint strength matrix. ;

[0234] (4) Compare the degree of difference between the old and new constraint strength matrices (e.g., calculate the degree of difference by the norm of the matrix difference). If the difference between the old and new constraint strength matrices is less than a preset threshold (e.g., the difference norm is less than 0.01), the graph structure is considered to be stable; otherwise, repeat the above steps with the new constraint strength matrix until the association graph is stable, and obtain a stable topological feedback association graph. .

[0235] Based on the topological feedback characteristics of the final correlation graph, the initial candidate regions for local binding energy anomalies are determined.

[0236] It should be noted that the goal of this step is based on the final topological feedback correlation graph formed. Accurately identify areas where abnormal bonding energy may exist (such as cold solder joints or weak bonds). The specific implementation method is as follows:

[0237] (1) Based on the final association graph Extract connected subgraphs or connected regions from the graph;

[0238] (2) Calculate the mean topological constraint strength of nodes within each connected region. :

[0239]

[0240] in, This represents the total number of connecting edges between nodes within the region;

[0241] (3) Set the threshold for the mean strength of topological constraints (For example, a value of 0.75) is used to further screen initial candidate regions with potential binding energy anomalies:

[0242] If the average topological constraint strength of a certain region is Less than the threshold This indicates that the constraint relationship between nodes in this region is weak and the topology is abnormal. It is preliminarily determined that there may be an abnormal binding energy problem in this region.

[0243] Conversely, if the mean strength of the topological constraints in the region is greater than or equal to the threshold... If the topological structure of the region is stable and there is no obvious binding energy anomaly, then it is considered that the region has a stable topological structure and there is no obvious binding energy anomaly.

[0244] (4) Define all connected regions that meet the preliminary judgment criteria for binding energy anomalies as the initial candidate regions for local binding energy anomalies. , represented as: .

[0245] S104: Active response mode enhancement excitation is performed on the local area of ​​the initial candidate region to obtain the dynamic enhancement response of each local area, and the binding energy anomalous region is determined according to the rate of change of the dynamic enhancement response.

[0246] In specific implementation, the determination of the binding energy anomaly region based on the dynamic enhancement response change rate includes:

[0247] Based on the differences in local latent variable characteristics of candidate regions, determine the spatial application strategy for dynamic enhancement incentives;

[0248] Specifically, firstly, from the initial candidate region of binding energy anomalies obtained in step S103... In the process, the latent variable feature vector corresponding to each local node is extracted. (20-dimensional), calculate the variance or dispersion of the latent variable feature vector within the local region of each candidate region to measure the degree of difference in latent variable features;

[0249] Define the latent variable feature difference measure for the m-th candidate region as the mean variance of the feature vector set:

[0250]

[0251] In the formula, Let m be the number of local nodes in the m-th candidate region. It represents the average value of the latent variable eigenvectors of the nodes in the m-th candidate region.

[0252] Furthermore, based on the aforementioned feature difference metric... The priority of applying dynamic enhancement stimuli to each candidate region is determined in descending order; the specific strategy is as follows:

[0253] Regions with significant differences in characteristics indicate that there is a significant non-uniformity in the distribution of local characteristics, and there is a greater possibility of binding energy anomalies. Therefore, dynamic enhancement excitation should be applied preferentially.

[0254] The incentive application strategy is as follows: apply the highest intensity incentive to the region with the greatest difference, apply a relatively low intensity incentive to the next region, and so on down.

[0255] The incentive intensity is allocated according to a preset threshold ratio. For example, the incentive intensity in the largest difference zone is 100% of the baseline value, the second largest difference zone is 80%, and so on.

[0256] Real-time acquisition of local dynamic displacement response of the interface under active enhancement incentive;

[0257] In specific implementation: Based on the spatial application strategy determined above, the candidate region of the titanium-steel composite plate interface is subjected to enhanced excitation in an active response mode, such as by using ultrasonic stress wave dynamic excitation. The excitation intensity and frequency can be selected according to the actual engineering implementation, such as an intensity range of 5MPa-10MPa and a frequency range of 50kHz-200kHz.

[0258] The dynamic displacement response signals of each node in the local area of ​​the interface are acquired in real time using laser Doppler vibration meter (LDV) or digital image correlation (DIC) technology. ;

[0259] The specific operations include:

[0260] The measuring device is precisely aligned with the identified candidate area;

[0261] During the duration of the excitation, the dynamic displacement response waveform of each local node is recorded;

[0262] The sampling frequency is set to more than 5 times the dynamic excitation frequency to ensure the accuracy of signal capture;

[0263] The data obtained through the above process constitutes the dynamic augmented response dataset for each local node: .

[0264] The dynamic stress distribution field is derived from the dynamic displacement response of the interface, and the dynamic stress change rate is calculated.

[0265] In practical implementation, this step aims to further calculate the dynamic stress distribution field of the interface region using the obtained dynamic displacement response data, specifically including:

[0266] (1) First, the dynamic displacement response is collected in real time from the interface using the equilibrium relationship of material mechanics and the constitutive relationship of material. The dynamic strain field is obtained through inversion:

[0267] It should be understood that, considering the potential noise effects in actual measurements, the acquired dynamic displacement response needs to be... Digital filtering is performed, such as using a Butterworth low-pass filter or a Savitzky-Golay smoothing filter, to obtain smoother and more accurate displacement gradient data, thereby ensuring the reliability of dynamic stress inversion calculation.

[0268] According to the linear elastic constitutive relation, the relationship between strain and displacement gradient is as follows:

[0269] strain tensor This is the symmetric part of the displacement gradient tensor, i.e.:

[0270]

[0271] (2) Subsequently, the dynamic stress field was obtained based on Hooke's law and the material constitutive relation:

[0272] According to the constitutive relation of conventional linear elastic materials, the stress tensor With strain tensor The relationship is:

[0273]

[0274] (3) Thus, the dynamic stress response field corresponding to the local nodes of each interface is obtained. Specifically, it is expressed as: ;

[0275] Further calculation of the dynamic stress change rate is defined as the relative change in stress at each local node over time. Specifically, for any node i, the dynamic stress change rate is defined as: ;

[0276] The dynamic stress rate of change field of all nodes is obtained: .

[0277] Based on the degree of deviation between the dynamic stress change rate and the latent variable characteristics, and combined with the local topological characteristics of the candidate region, the actual binding energy anomaly region is determined.

[0278] In practical implementation, the goal of this step is to accurately identify regions with abnormal bonding energy (such as incomplete welds, weak bonds, etc.) at the titanium-steel interface by integrating dynamic stress change rate characteristics, latent variable characteristics, and topological structure characteristics. The specific details are as follows:

[0279] (1) First, clarify the relationship between the dynamic stress change rate and the characteristics of latent variables:

[0280] Specifically, through multiple linear regression or canonical correlation analysis, a statistical relationship is established between the dynamic stress change rate characteristics of each node within the candidate region and the latent variable characteristics of the corresponding nodes, as specifically expressed as:

[0281] Define a predictive mapping function between the rate of change of dynamic stress and the characteristics of latent variables. Its input is the nodal latent variable feature vector. The output is the predicted dynamic stress change. ;

[0282] To implement the above mapping function, this embodiment specifically adopts a conventional multilayer perceptron (MLP) model:

[0283] Input layer: Latent variable feature vectors of input nodes (20 dimensions).

[0284] Hidden layers: 2 hidden layers, each containing 64 neurons, with ReLU as the activation function.

[0285] Output layer: Outputs the predicted dynamic stress change rate. .

[0286] The training process of this model is as follows:

[0287] Training data: The actual values ​​of the dynamic stress change rate of the candidate region nodes are used. As a supervisory label, the latent variable feature vector As input;

[0288] Training objective: Define the mean squared error (MSE) loss function, specifically:

[0289]

[0290] Where N is the total number of training samples. The training process uses the existing Adam optimization algorithm and continues until the loss function converges.

[0291] (2) Calculate the degree of correlation deviation of the dynamic stress change rate:

[0292] Based on the trained prediction model, the predicted dynamic stress change rate is calculated for each node. Then compare with the actual measured value Compare and define the degree of correlation deviation (prediction residuals). ;

[0293] The degree of deviation of this correlation quantifies the bias that the latent variable characteristics cannot accurately characterize the dynamic stress response, that is, it characterizes the degree to which the local node may have anomalies.

[0294] (3) Combine the topological characteristics of the candidate regions to perform fine screening of abnormal regions:

[0295] In this step, the obtained topological association structure is further utilized, specifically through the following method:

[0296] First, calculate the average association deviation of all nodes within each candidate region (connected region):

[0297]

[0298] Furthermore, considering the average topological constraint strength within the region Define the comprehensive index for anomaly detection in the region ,in, This is a weighting factor (e.g., a value of 0.7).

[0299] Identify the actual binding energy anomaly region:

[0300] Set comprehensive indicators Abnormal region judgment threshold (For example, a value of 0.3, to be determined experimentally); when the comprehensive index of the candidate region exceeds the threshold, that is:

[0301] like If so, the region is clearly identified as an area with an abnormal actual binding energy.

[0302] like If so, the region is considered a normal region or a region with insignificant abnormality;

[0303] The final determined actual binding energy anomaly region is explicitly represented as: .

[0304] S105: Based on the difference in the correlation topology between the local characteristics in the region of abnormal binding energy and the normal region, determine the specific location and range of the hidden weld band in the interface of the titanium-steel composite plate.

[0305] In specific implementation, determining the specific location and range of the hidden weld bands at the interface of the titanium-steel composite plate includes:

[0306] Extract the topological adjacency matrix between the binding energy anomalous region and the normal region;

[0307] Specifically, this step aims to clearly distinguish the differences in topological structure between the local characteristics of regions with anomalous binding energy and those of normal regions (the remaining regions):

[0308] (1) Based on the obtained final association diagram The set of node indices belonging to the abnormal region and the normal region in the node set V are defined as follows:

[0309] Set of abnormal region nodes: ;

[0310] Normal region node set: ;

[0311] (2) Based on the above node set, extract the adjacency matrices of nodes in the abnormal region and the normal region respectively:

[0312] Taking an abnormal region as an example, the adjacency matrix of an abnormal region is defined as follows:

[0313]

[0314] Similarly, the adjacency matrix of a normal region is defined as:

[0315]

[0316] The adjacency matrix represents the topological connection between nodes in two regions.

[0317] Calculate the difference tensor of the topological feature matrices of the two regions;

[0318] Specifically, this step utilizes a graph convolutional network (GCN) to extract topological features from the adjacency matrix and clarifies the differences between the topological features of two regions:

[0319] (1) For the adjacency matrix of the abnormal regions respectively Adjacency matrix with normal regions Perform graph convolution operations to extract the topological features of nodes within the corresponding region:

[0320] Specifically, the graph convolutional network structure explicitly disclosed in step S103 is used for feature extraction to obtain:

[0321] Anomaly region topological feature matrix: ;

[0322] Normal region topological feature matrix: ;

[0323] Among them, the output feature dimension ;

[0324] (2) Calculate the difference tensor:

[0325] Define the difference tensor To account for the statistical differences in node feature distribution between abnormal and normal regions, this embodiment uses the difference in the mean topological features between the two regions:

[0326]

[0327] in:

[0328] Mean topological features of anomaly regions: ;

[0329] Mean topological features of normal regions: .

[0330] Based on the cross-validation results of the differential tensor space distribution and the dynamic enhancement response, the boundary between the abnormal and normal regions of the interface binding energy is determined.

[0331] Specifically, this step aims to clarify the precise boundary between abnormal and normal regions, and the specific implementation method is as follows:

[0332] (1) Using the obtained difference tensor Perform topological feature difference mapping on all nodes of the interface, which involves calculating the inner product of the topological feature vector and the difference tensor of each node to obtain the topological difference mapping value of the node:

[0333] For any node features :

[0334]

[0335] Among them, node topology difference mapping value The larger the absolute value, the more the topology of the node deviates from the normal region;

[0336] (2) Degree of deviation between the mapping value of node topology differences and the dynamic stress change rate obtained in step S104. Perform cross-validation:

[0337] The specific implementation includes:

[0338] Define cross-validation metrics for each node. Taking into account both topological difference mapping and dynamic enhancement response differences:

[0339]

[0340] The typical value range of the weight factor 𝛽 is 0.5 to 0.7, and the typical value in this embodiment is 0.6;

[0341] Determine the cross-validation threshold for topological differences in the interface region. For example, a value of 0.2 (the specific value can be determined through experiments);

[0342] Based on the above indicators Clearly define the boundary between abnormal and normal areas:

[0343] If node indicators This node clearly belongs to the abnormal region;

[0344] If node indicators This node belongs to the normal area;

[0345] Define the set of nodes between the abnormal and normal regions as the boundary node set: .

[0346] Based on boundary continuity constraint rules, spatial clustering analysis is used to precisely determine the specific location and range of continuous implicit solder strips;

[0347] Specifically, the goal of this step is to determine the spatial location and extent of continuous cold solder joints. Specific implementation method:

[0348] (1) Define the boundary continuity constraint rules:

[0349] If nodes are spatially adjacent (e.g., adjacent grid cells or a distance less than a set spatial threshold, such as 1 mm), then these nodes are considered to have spatial continuity.

[0350] (2) Use spatial clustering analysis algorithm to analyze the set of boundary nodes. Perform cluster analysis, such as DBSCAN or K-Means clustering algorithms:

[0351] Taking DBSCAN as an example: set the neighborhood radius ϵ to 1 mm; set the minimum number of nodes MinPts to 3-5; perform cluster analysis on the boundary node set to obtain several clustering results;

[0352] (3) Clearly define the specific location and extent of continuous cold solder joints:

[0353] For each cluster subset, specify the set of node coordinates. ;

[0354] Based on the spatial coordinates of the nodes within the cluster subset, the spatial location, length, and width range of the dummy solder strip are clearly determined;

[0355] This ultimately yields a precise description of the hidden solder joint, for example:

[0356]

[0357] This clearly indicates the specific spatial location and continuous range of the hidden solder joint.

[0358] Example 2

[0359] like Figure 2 As shown in the example, the parts not detailed in this embodiment are as shown in Example 1. This embodiment discloses a quality assessment system for composite panels, including:

[0360] The excitation acquisition module 201 is used to apply an excitation field of a preset form to the interface of the titanium-steel composite plate and acquire the initial response data field of the interface.

[0361] The feature construction module 202 is used to construct a latent variable feature space characterizing interface binding energy anomalies based on the initial response data field and through the physical constraint relationship between the interface local response features and the binding energy distribution.

[0362] Topology extraction module 203 is used to extract the topological association structure between local latent variable features in the latent variable feature space through graph convolutional network to obtain the initial candidate region of local binding energy anomaly.

[0363] The dynamic determination module 204 is used to perform active response mode enhancement excitation on the local area of ​​the initial candidate region to obtain the dynamic enhancement response of each local area, and to determine the binding energy anomalous region based on the rate of change of the dynamic enhancement response.

[0364] The cold weld positioning module 205 is used to determine the specific location and range of the hidden cold weld band in the interface of the titanium-steel composite plate based on the difference in the correlation topology between the local characteristics of the abnormal bonding energy region and the normal region.

[0365] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters, weights, and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0366] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless network. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0367] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for quality assessment of composite panels, characterized in that, include: S101: Apply a preset excitation field to the interface of the titanium-steel composite plate to obtain the initial response data field of the interface. S102: Based on the initial response data field, a latent variable feature space characterizing interface binding energy anomalies is constructed through the physical constraint relationship between the interface local response characteristics and the binding energy distribution. The construction of the latent variable feature space representing the interface binding energy anomaly includes: Using the initial response data field of the interface locality as input, latent space features are extracted based on the variational autoencoder model; By utilizing the physical constraint relationship of the interface binding energy anomaly, the latent space features are physically constrained and regularized to generate latent variable features constrained by the interface binding energy distribution. The physical constraint regularization of the latent space features includes: Based on the intrinsic physical relationship between interfacial bonding energy and strain energy density, a physical energy function for bonding energy is established; specific implementation method: Using latent space characteristics as input, the strain energy density distribution of each locality is calculated through the physical energy function; Based on the binding energy threshold condition, a physical relationship mapping is performed between the strain energy density distribution and the initial response field of the interface. Based on the physical relationship mapping error, the latent space features are iteratively corrected by optimizing the regularization loss function to obtain latent variable features that meet the interface binding energy anomaly constraint. The interface local response data field is reconstructed in reverse using latent variable features, the reconstruction error distribution is calculated, and the latent variable features are corrected by cross-validation with the difference from the initial response data field. The modified latent variable features are used to form the final interface-binding energy latent variable feature space; S103: Extract the topological correlation structure between local latent variable features in the latent variable feature space through graph convolutional networks to obtain the initial candidate region of local binding energy anomaly. S104: Active response mode enhancement excitation is performed on the local area of ​​the initial candidate region to obtain the dynamic enhancement response of each local area, and the binding energy anomalous region is determined according to the rate of change of the dynamic enhancement response. S105: Based on the difference in the correlation topology between the local characteristics in the anomaly region and the normal region, determine the specific location and range of the hidden weld band in the interface of the titanium-steel composite plate. The determination of the specific location and extent of the hidden weld bands at the interface of the titanium-steel composite plate includes: Extract the topological adjacency matrix between the binding energy anomalous region and the normal region; Calculate the difference tensor of the topological feature matrices of the two regions; Based on the cross-validation results of the differential tensor space distribution and the dynamic enhancement response, the boundary between the abnormal and normal regions of the interface binding energy is determined. Based on boundary continuity constraint rules, spatial clustering analysis is used to precisely determine the specific location and range of continuous implicit solder bands.

2. The quality assessment method for composite panels according to claim 1, characterized in that, The acquisition of the initial response data field of the local interface includes: Based on the elastic modulus, density, and acoustic impedance distribution data of the titanium-steel composite plate, an initial finite element simulation model of the titanium-steel composite plate is established. In the initial finite element simulation model, thermal field, ultrasonic stress wave and electromagnetic field excitations are applied respectively. The most sensitive excitation type is determined by the difference in displacement response distribution after the three excitations. Based on the determined most sensitive excitation type, the initial displacement response of the interface locality under this excitation is extracted, and the initial displacement response field is established. Based on the initial displacement response field, the initial stress response field is inverted using the material mechanics equilibrium equations. The initial displacement response field and the initial stress response field are fused to form an initial response data field characterizing the local interface.

3. The quality assessment method for composite panels according to claim 2, characterized in that, The determination of the most sensitive stimulus type includes: Using the interface bonding energy threshold of the normal bonding region and the virgin weld region as constraints, the displacement sensitivity matrix of each locality under different excitation types is calculated respectively. With the goal of maximizing the entropy of the local binding energy difference of titanium-steel composite plates, a genetic algorithm is used to optimize the selection of the displacement sensitivity matrix and determine the most sensitive excitation type. The sensitivity matrix corresponding to the most sensitive excitation type is cross-feeded to the initial finite element simulation model. By re-correcting the boundary conditions of the simulation model, the most sensitive excitation type is updated and determined.

4. The quality assessment method for composite panels according to claim 3, characterized in that, The extraction of the topological association structure between local latent variable features in the latent variable feature space using a graph convolutional network includes: Based on the latent variable characteristics of each interface in the latent variable feature space, the Euclidean distance and cosine similarity between nodes are determined, and an initial association graph is constructed. The initial association graph is input into a graph convolutional network to extract node feature topology. The node feature topology is mapped inversely to the latent variable feature space, and the constraint strength of the topology features on the latent variable space is calculated. The initial correlation graph is corrected based on the feedback of constraint strength differences to form a final correlation graph with topological feedback structure. Based on the topological feedback characteristics of the final correlation graph, the initial candidate regions for local binding energy anomalies are determined.

5. The quality assessment method for composite panels according to claim 4, characterized in that, The method for obtaining the graph convolutional network includes: A large amount of titanium-steel composite plate sample data with known interface bonding states were collected to obtain the latent variable characteristics of the sample interface and the corresponding true distribution of bonding energy. A supervised loss function for determining binding energy anomalies is constructed, using the true binding energy anomaly distribution of the samples as the supervised label; Train the graph convolutional network until the decision error is less than a predetermined loss threshold, and alternately use training data and simulation data for bidirectional cross-validation; The training data ratio is dynamically adjusted based on the cross-validation results, and the network parameters are iteratively optimized to ultimately form a graph convolutional network with generalization capabilities.

6. The quality assessment method for composite panels according to claim 5, characterized in that, The binding energy anomaly region is determined based on the dynamic enhancement response change rate, including: Based on the differences in local latent variable characteristics of candidate regions, determine the spatial application strategy for dynamic enhancement incentives; Real-time acquisition of local dynamic displacement response of the interface under active enhancement incentive; The dynamic stress distribution field is derived from the dynamic displacement response of the interface, and the dynamic stress change rate is calculated. Based on the degree of deviation between the dynamic stress change rate and the latent variable characteristics, and combined with the local topological characteristics of the candidate region, the actual binding energy anomaly region is determined.

7. A quality assessment system for composite panels, implemented based on the quality assessment method for composite panels according to any one of claims 1-6, characterized in that, include: The excitation acquisition module is used to apply a preset form of excitation field to the interface of the titanium-steel composite plate and acquire the initial response data field of the interface. The feature construction module is used to construct a latent variable feature space characterizing interface binding energy anomalies based on the initial response data field and through the physical constraint relationship between the interface local response features and the binding energy distribution. The topology extraction module is used to extract the topological relationship structure between local latent variable features in the latent variable feature space through graph convolutional networks, so as to obtain the initial candidate region of local binding energy anomaly. The dynamic determination module is used to perform active response mode enhancement excitation on the local area of ​​the initial candidate region to obtain the dynamic enhancement response of each local area, and to determine the binding energy anomalous region based on the rate of change of the dynamic enhancement response. The cold weld location module is used to determine the specific location and range of the hidden cold weld band in the interface of the titanium-steel composite plate based on the difference in the correlation topology between the local characteristics of the abnormal bonding energy region and the normal region.

Images