Aircraft defect recognition method and system based on tensor decomposition and attention mechanism

Abstract

The application relates to the technical field of nondestructive testing, and discloses an aircraft defect identification method and system based on tensor decomposition and an attention mechanism. The aircraft defect identification method based on tensor decomposition and the attention mechanism comprises the following steps: acquiring multi-modal data of an aircraft; constructing the multi-modal data into a four-order space-time-modal tensor, and generating a dynamic graph structure based on the modal features of the tensor; applying a mixed constraint when decomposing the four-order tensor to obtain a core tensor and a factor matrix; extracting features through multi-scale pooling, combining topological persistent homology and a gated attention mechanism to distribute weights, and realizing feature fusion; identifying a defect type based on the fused features, and locating a defect area by using a factor matrix gradient amplitude and a dynamic threshold. Through multi-modal data fusion, dynamic graph regularization constraint and mixed tensor decomposition technology, the application improves the detection sensitivity and positioning accuracy of small defects on the surface of the aircraft, and enhances the physical interpretability of features and the robustness of the algorithm to complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, specifically to an aircraft defect identification method and system based on tensor decomposition and attention mechanism. Background Technology

[0002] Accurate early detection of surface defects on aircraft is a core requirement for ensuring flight safety. With the widespread use of composite materials in aircraft, traditional single-modal detection methods (such as optical vision or acoustic emission monitoring) are insufficient to meet the high-sensitivity requirements of detecting minute surface damage and multi-physics coupling effects. Especially in dynamic service environments, single-sensor data is susceptible to noise interference and cannot comprehensively characterize the multi-dimensional features of defects (such as deformation, acoustic response, and strain distribution). Therefore, there is an urgent need for an intelligent detection method that can unify and integrate multi-modal data while considering both physical laws and mathematical constraints.

[0003] In existing technologies, aircraft defect detection mainly employs a multimodal data segmentation strategy. For example, optical images and acoustic emission signals are aligned using time synchronization mechanisms, and independent classifiers are used to extract features from each modality. At the data fusion level, cascaded stitching or weighted averaging methods are used to integrate heterogeneous features. For strain field analysis, low-dimensional features are extracted based on traditional tensor decomposition methods, and threshold segmentation algorithms are combined to locate defect areas. Furthermore, some methods attempt to introduce graph neural networks to model sensor node relationships, or extract multi-resolution features through fixed-scale convolution.

[0004] However, existing technologies still have some shortcomings. In multimodal data fusion, the lack of a unified high-dimensional tensor modeling framework makes it difficult to effectively coordinate cross-modal spatiotemporal correlation features. Furthermore, traditional tensor decomposition methods rely excessively on low-rank mathematical assumptions, neglecting physical laws such as the continuity of material mechanics, causing the decomposition results to deviate from the actual physical field distribution. In addition, static graph models and fixed-scale feature extraction strategies are difficult to adapt to dynamic changes in local anomalies under complex working conditions, while manual threshold settings limit adaptability to different defect types and noisy environments. This invention overcomes these shortcomings by constructing a fourth-order spatiotemporal-modal tensor to uniformly represent multi-source data, designing a hybrid constraint tensor decomposition model to inject physical laws and dynamic graph correlation constraints, and combining a multi-scale attention mechanism and adaptive threshold localization, thereby achieving high-precision and robust aircraft defect detection. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an aircraft defect identification method and system based on tensor decomposition and attention mechanism, which solves the problems of insufficient sensitivity in detecting minute defects on aircraft surfaces, difficulty in modeling cross-temporal correlation of multimodal data, and high false detection rate and poor robustness caused by the separation of physical laws and mathematical constraints in existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an aircraft defect identification method based on tensor decomposition and attention mechanism, comprising the following steps:

[0007] Acquire multimodal data on cracks, fissures, bulges, and dents on the surface of aircraft;

[0008] Multimodal data is constructed as a fourth-order tensor, and a dynamic graph structure is constructed based on the modal features of the tensor, wherein the edge weights of the dynamic graph are adaptively adjusted according to the differences in cross-modal features;

[0009] The fourth-order tensor is decomposed into a tensor, and low-rank constraints, dynamic graph regularization constraints, and physical field continuity constraints are applied during the decomposition process to generate the core tensor and factor matrix.

[0010] Multi-scale features are extracted from the core tensor, and weights are assigned to the multi-scale features by combining topological persistent homology analysis to obtain weighted fused multi-scale features.

[0011] The defect type is identified based on the weighted fusion of multi-scale features, and the defect location is located based on the gradient change of the factor matrix.

[0012] This invention provides a method and system for aircraft defect identification based on tensor decomposition and attention mechanism.

[0013] It has the following beneficial effects:

[0014] 1. This invention effectively overcomes the limitations of single sensor data in traditional methods by constructing a fourth-order tensor structure to unify and fuse multimodal data such as optical images, acoustic emission signals, and strain field distribution. At the same time, it utilizes dynamic graph regularization constraints and hybrid constraint tensor decomposition techniques to achieve synergistic optimization of cross-modal features and noise suppression, thereby improving the detection sensitivity and accuracy of minute defects (such as cracks and corrosion).

[0015] 2. This invention introduces a joint optimization model that combines low-rank constraints, dynamic graph Laplace constraints, and physical field continuity constraints. This model forces the preservation of the inherent spatiotemporal correlation and physical rationality of the data during the mathematical decomposition process, avoiding feature distortion caused by the deviation of mathematical assumptions from physical laws in traditional tensor decomposition, thereby enhancing the physical interpretability of the missing features.

[0016] 3. This invention dynamically generates a graph structure adjacency matrix based on multimodal feature differences, adaptively captures abnormal correlation patterns in local regions (such as abrupt changes in acoustic emission energy and strain distribution mismatch), and combines graph Laplace regularization constraints to effectively solve the shortcomings of traditional static graph models that cannot adapt to complex working conditions and improve robustness to non-uniform background interference.

[0017] 4. This invention extracts coarse-grained, medium-grained, and fine-grained features through multi-scale pooling operations, and combines them with the significance index map generated by topological persistent homology analysis. It also uses a gated attention mechanism to adaptively allocate weights, achieving complementary advantages between global distribution patterns and local detailed features, thus avoiding the problem of missed or false detections of minor defects at a single scale.

[0018] 5. This invention is based on the calculation of the gradient magnitude of the spatial factor matrix and the dynamic generation of thresholds using local statistical properties. Combined with morphological post-processing to eliminate noise interference, it does not rely on predefined thresholds or human experience. It can adapt to the needs of defect localization under different materials, lighting and noise environments, and improves the engineering applicability and generalization ability of the algorithm. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention;

[0020] Figure 2 This is a diagram of the module architecture of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see the appendix Figure 1 This invention provides a method and system for aircraft defect identification based on tensor decomposition and attention mechanism, comprising the following steps:

[0023] S1. Acquire multimodal data of cracks, fissures, bulges and dents on the surface of the aircraft;

[0024] Multimodal data on cracks, fissures, bulges, and dents on aircraft surfaces are obtained through the following methods:

[0025] An industrial-grade area array camera is used to acquire optical images of the aircraft surface. The camera has a built-in synchronization trigger module to eliminate ambient light interference and ensure time synchronization. The optical images cover the entire area to be inspected, and surface texture and morphology features are obtained through multi-angle shooting. Preferably, a multi-band combined illumination is used, with the wavelength range covering the visible to near-infrared spectrum to enhance the differences in reflectivity of different material surfaces. The optical images generate an RGB three-channel pixel matrix at each time point and are aligned with the remaining modal data through timestamps.

[0026] Acoustic emission signals from the aircraft structure are acquired using a multi-channel acoustic emission sensor array, with sensors positioned at key load-bearing components and connections. The acoustic emission sensors possess wide frequency response characteristics, enabling them to capture elastic wave signals generated by defects such as material crack propagation and debonding. Preferably, the sensor array achieves millisecond-level time synchronization with the optical camera and strain acquisition system via a synchronization trigger module, ensuring the spatiotemporal consistency of multimodal data. After preprocessing, the acoustic emission signals are used to extract time-domain energy characteristics and frequency-domain resonance peak parameters, calculated using the following formula:

[0027]

[0028] Wherein, STFT is the Short-Time Fourier Transform, used to convert a time-domain signal into a time-frequency domain representation; s(t,h,w) is the time-domain waveform of the acoustic emission signal at time t and spatial location (h,w); f min ,f max The effective frequency range of the acoustic emission signal is defined by the lower and upper limits, determined based on the sensor's frequency response characteristics; f is the frequency index variable, iterating through [f... min ,f max Discrete frequency points within the interval.

[0029] The strain field distribution on the aircraft surface is obtained using digital image correlation (DIC) technology. A high-contrast target pattern is attached to the surface of the structure under test, and the deformation process is continuously captured by a high frame rate area array camera. A sub-pixel displacement tracking algorithm is then used to calculate the strain tensor components. The strain field distribution data includes strain components in three orthogonal directions, namely:

[0030] E(h,w)=[∈ xx (h,w),∈ yy (h,w),∈ xy (h,w)];

[0031] Where, ∈ xx (h,w) represents the normal strain component along the horizontal direction at location (h,w), calculated using the DIC algorithm; ∈ yy (h,w) represents the normal strain component along the vertical direction at position (h,w); ∈ xy (h,w) represents the shear strain component at position (h,w), which characterizes the angular distortion in the xy plane.

[0032] Optical images, acoustic emission signals, and strain field distribution data are aligned in the time dimension and mapped to a unified spatial grid coordinate system using a timestamp synchronization mechanism. Preferably, the spatial grid resolution is determined by the density of the DIC target pattern, with each grid cell corresponding to the physical coordinates of the aircraft surface. The multimodal data is ultimately filled into a fourth-order tensor structure.

[0033]

[0034] Where t is the time index variable, representing the time point number of synchronous acquisition; h is the spatial height index variable, representing the row number in the two-dimensional gridded spatial coordinate system; w is the spatial width index variable, representing the column number in the two-dimensional gridded spatial coordinate system; m is the modality index variable; R(h,w), G(h,w), and B(h,w) are the pixel intensity values ​​of the red, green, and blue channels of the optical image at position (h,w).

[0035] Among them, multi-band illumination of optical images is used to enhance the contrast of minor surface defects (such as scratches and corrosion) and avoid feature loss caused by a single light source.

[0036] Frequency domain analysis of acoustic emission signals: Identifying the specific frequency band characteristics of different types of defects (such as cracks and debonding) through formant parameters;

[0037] Tensor-quantized representation of strain field: mapping strain components to physical quantity dimension vectors to provide structured input for subsequent tensor decomposition;

[0038] Spatiotemporal synchronization mechanism: ensures that multimodal data are strictly aligned in time and space, avoiding feature fusion errors caused by temporal deviations.

[0039] S2. Construct the multimodal data into a fourth-order tensor, and build a dynamic graph structure based on the modal features of the tensor, wherein the edge weights of the dynamic graph are adaptively adjusted according to the differences in cross-modal features;

[0040] The multimodal data acquired in step S1 is mapped to a fourth-order tensor structure, which includes time, space, modality, and physical quantity dimensions. The time dimension represents the synchronous acquisition time series of the multimodal data, with each time point corresponding to a complete multimodal data acquisition cycle. The space dimension characterizes the physical location of the aircraft surface using a two-dimensional gridded coordinate system. The grid resolution is determined by the distribution density of the digital image correlation (DIC) target pattern, ensuring that each grid cell corresponds to a unique coordinate on the actual structural surface. The modality dimension is divided into three categories: optical images, acoustic emission signals, and strain field distribution, each carrying data with different physical properties. The physical quantity dimension stores the quantization indicators of each mode, such as the RGB channel intensity of the optical image, the energy value of the acoustic emission signal, and the orthogonal components of the strain tensor.

[0041] The filling rule for fourth-order tensors is as follows: for each time point t, spatial location (h, w), and mode m, the corresponding physical quantity vector is filled into the tensor slice.

[0042] A dynamic graph structure is constructed based on the multimodal features of a fourth-order tensor. The dynamic graph uses spatial locations as nodes and cross-modal feature differences as edge weights to achieve adaptive modeling of data relationships. A node is defined as the multimodal feature vector of each spatial location at time t, calculated using the following formula:

[0043]

[0044] in, : represents the physical quantity vector of the fourth-order tensor at time t, spatial position (h, w), and mode M, where M = 1, 2, and 3 correspond to the optical image, acoustic emission signal, and strain field distribution, respectively; : represents a full slice of the physical quantity dimension.

[0045] The feature vector integrates optical, acoustic emission, and strain data at the same location, and retains cross-modal interaction information through splicing operations.

[0046] Edge weights are dynamically calculated based on the cross-modal feature differences between nodes, using adjacency matrix elements. The calculation formula is:

[0047]

[0048] in, Let be the edge weight between node i and node j at time t; and Let i and j represent the feature vectors of nodes i and j, respectively. This represents the mean difference in neighborhood features between nodes i and j.

[0049] Generate the graph Laplacian matrix L based on the dynamic graph structure. G This is used for regularization constraints in subsequent tensor decomposition. The formula for calculating the graph Laplacian matrix is:

[0050] L G =DA t ;

[0051] Among them, L G A is the Laplacian matrix of the dynamic graph, used to constrain the tensor decomposition process and reflect the topological relationships between nodes; D is the degree matrix of the dynamic graph; A t Let be the adjacency matrix at time t, whose elements are generated by the edge weight calculation formula.

[0052] Among them, the structured representation of the fourth-order tensor integrates multimodal data through a unified spatiotemporal framework, solves the feature alignment problem caused by data heterogeneity in traditional methods, and provides normalized input for subsequent tensor decomposition;

[0053] Adaptive edge weights in dynamic graphs: Based on cross-modal feature differences, the node connection strength is dynamically adjusted, which can effectively capture abnormal association patterns in local regions;

[0054] Generation of the Graph Laplacian Matrix: Transforms the dynamic graph structure into mathematical constraints, ensuring that the tensor decomposition process and the intrinsic relationship between the data are optimized in synergy, thereby improving the physical consistency of the decomposition results.

[0055] S3. Perform tensor decomposition on the fourth-order tensor, applying low-rank constraints, dynamic graph regularization constraints, and physical field continuity constraints during the decomposition process to generate the core tensor and factor matrix.

[0056] Hybrid constrained tensor decomposition is achieved in the following way:

[0057] TuckEr decomposition model construction:

[0058] The fourth-order tensor constructed in step S2 Decomposed into core tensor sum factor matrix And satisfy: Among them, R T R H R W R M For the rank parameter of the decomposition; factor matrix U (1) U (2) U (3) These respectively represent the temporal evolution pattern, spatial distribution pattern, and modal interaction pattern.

[0059] During the decomposition process, low-rank constraints, dynamic graph regularization constraints, and physical field continuity constraints are applied jointly. The objective function is defined as follows:

[0060]

[0061] Among them, ||·|| F The Frobenius norm of a matrix or tensor is used to measure reconstruction error. tr(·) is the nuclear norm of the core tensor, which is the sum of all its singular values; tr(·) is the trace operation of the matrix, representing the dynamic graph regularization term; λ1, λ2, and λ3 are the gradient Frobenius norm of the spatial factor matrix, calculated using the finite difference method; λ1, λ2, and λ3 are regularization coefficients used to balance the constraint strengths of low rank, graph structure consistency, and physical field continuity. For low-rank constraints; λ2tr(U (3)T L G U (3) ) represents the dynamic graph regularization constraint; This is a constraint on the continuity of the physical field.

[0062] The following variables are updated alternately using an improved alternating direction multiplier method:

[0063] Core tensor update:

[0064]

[0065] in, U is a tensor; k is the ADMM iteration number; λ1 is the regularization coefficient of the low-rank constraint; U (1) U (2) U (3) This is the factor matrix for the current iteration;

[0066] Factor matrix update:

[0067]

[0068] Similarly, update U (2) U (3) , among which, T (1) For tensor The modulo-1 expansion matrix of G; (1) For the core tensor The matrix after modulo-1 expansion; Represents the Kronecker product;

[0069] Dynamic graph constraint synchronous update: based on the current factor matrix U (3) Adjusting the Laplacian matrix L of the dynamic graph G This is then injected into the regularization term of the next iteration, achieving collaborative optimization between the decomposition process and the dynamic graph structure.

[0070] The physical meaning of low-rank constraints is that aircraft defects are usually manifested as local anomalies, and the low-rank property of the core tensor can effectively separate the normal background pattern from the defect features.

[0071] The role of dynamic graph regularization: to constrain the modal interaction factor U through the dynamic graph Laplacian matrix. (3) This enhances the interpretability of cross-modal correlations, such as the synergistic characterization of acoustic emission energy abrupt changes and strain anomalies;

[0072] The introduction of physical field continuity constraints: The gradient constraint of the spatial factor matrix ensures that the strain distribution after decomposition conforms to the material continuity medium assumption, avoiding physical discontinuities caused by mathematical decomposition.

[0073] Improvement of ADMM algorithm: This invention achieves multi-objective joint optimization by alternately updating the core tensor, factor matrix and dynamic graph constraints, thereby improving convergence efficiency.

[0074] S4. Extract multi-scale features from the core tensor and combine topological persistent homology analysis to assign weights to the multi-scale features to obtain weighted fused multi-scale features.

[0075] Multi-scale feature extraction and weight allocation are achieved through the following methods:

[0076] The core tensor obtained from step S3 Multi-scale features are extracted, including coarse-grained, medium-grained, and fine-grained features. Coarse-grained features are generated through max pooling operations on each modal slice of the core tensor. Perform max pooling with a kernel size of 2×2 and a stride of 2 to output the feature map. Medium-grained features are generated through average pooling. The core tensor is subjected to average pooling with a kernel size of 3×3 and a stride of 1, and the resulting feature map is output. Fine-grained features directly preserve the original core tensor As a feature map F3, multi-scale pooling operations capture both local details and global distribution patterns of defects through features of different granularities.

[0077] The multi-scale feature maps F1, F2, and F3 are concatenated along the channel dimension to form a multi-scale feature pyramid. Preferably, bilinear interpolation upsampling is performed on the coarse-grained feature map F1 to restore its spatial dimension to R1×R2, ensuring spatial alignment of feature maps at all scales.

[0078] Based on the strain field distribution data acquired in step S1 The topological persistent cohomology is calculated to generate a saliency index map. Persistent cohomology is achieved by constructing a Vietoris-Rips complex, extracting topological features (such as connected components and annular holes) from the strain field, and calculating the persistence (i.e., the lifetime) of each feature. Saliency index map. The calculation formula is:

[0079]

[0080] in, Centered on (h,w); c i The topological features that fall within this neighborhood; persistence(c i ) = death i -birth i Representing topological feature c i The persistence, birth i and death i These are the thresholds for the appearance and disappearance of the feature during the filtering process, respectively.

[0081] The multi-scale feature pyramid Fpyramid is fused with the saliency index map P, and attention weights are generated through a learnable parameter matrix. For the k-th scale feature map F... k Its attention weight α k The calculation formula is:

[0082]

[0083] Among them, F k For the k-th scale feature map, This indicates a channel splicing operation.

[0084] Attention weights are normalized and limited to the [0,1] interval to represent the contribution of features at each scale to defect detection.

[0085]

[0086] Where, α k F represents the attention weights at the k-th scale; k This is the feature map at the k-th scale; These are the fused multi-scale features, used as input to the subsequent classifier.

[0087] Among them, the multi-scale pooling design includes: max pooling (coarse-grained) to suppress noise and retain significant defect features; average pooling (medium-grained) to capture local average response and balance details and smoothness; and original scale (fine-grained) to retain high-resolution spatial information and ensure that small defects are not missed.

[0088] Topological persistent cohomology analysis: Identifies discontinuous regions in materials through the topological features of the strain field (such as pores and connected branches), which is consistent with the physical fracture mechanism; the saliency map P quantifies the local topological stability, providing prior guidance for the attention mechanism.

[0089] Gated attention mechanism: Dynamic weight allocation avoids the limitations of fixed weighting strategies and adapts to the characteristic differences of different defect types (such as cracks and corrosion); splicing saliency map P introduces physical prior knowledge and enhances the model's sensitivity to topological anomalies.

[0090] Feature fusion rules: The weighted summation operation combines the advantages of multi-scale features, taking into account both global distribution and local details; channel-dimensional splicing preserves the independent information of each scale and avoids feature confusion.

[0091] S5. Identify the defect type based on the multi-scale features after weighted fusion, and locate the defect location based on the gradient change of the factor matrix;

[0092] Defect type identification and location are achieved through the following methods:

[0093] The fusion features obtained in step S4 Flattened into a one-dimensional feature vector The input is a fully connected neural network for classification, where d = R1 × R2 × 3R3. The fully connected network contains one hidden layer and a softmax output layer, and its computation process is as follows:

[0094] y = Softmax(W2·ReLU(W1F));

[0095] in, is the learnable weight matrix; ReLU is the modified linear unit activation function; Softmax is the normalized exponential function, which outputs the class probability.

[0096] Based on the spatial factor matrix obtained from step S3 The spatial gradient magnitude is calculated to locate the defect region. The gradient magnitude is calculated in the horizontal direction using the finite difference method. and vertical direction The partial derivatives are calculated, and a gradient magnitude map is synthesized:

[0097]

[0098] in, The calculations were performed using the finite difference method in both the horizontal and vertical directions.

[0099] Preferably, the dynamic threshold τ(h,w) of the defective region is adaptively determined through local statistical characteristics, and the calculation formula is as follows:

[0100] τ(h,w)=μ P +β·σ P ;

[0101] Where, μ P (h,w), σ P (h,w) represent the mean and standard deviation of the significance index P, respectively; β is the adjustment coefficient.

[0102] Based on the comparison between the gradient magnitude and the dynamic threshold, a binary mask D(h,w)∈{0,1} for the defective region is generated. The calculation formula is as follows:

[0103]

[0104] Where τ(h,w)=μ P (h, w) + β·σ P (h,w) represents the dynamic threshold, μ P (h,w), σ P (h,w) represent the mean and standard deviation of the significance index P, respectively; β is the adjustment coefficient; D(h,w) indicates that there is a defect at position (h,w).

[0105] Pixels with a value of 1 in the mask are marked as potential defect areas. Isolated noise points are removed and broken areas are connected by morphological closing operations (such as dilation-erosion).

[0106] Among them, the fully connected network design is as follows: the hidden layer dimension h is optimized through the validation set to avoid overfitting or underfitting; the ReLU activation function introduces sparsity and enhances the non-linear expressive ability of the model; the Softmax output layer normalizes the classification results into probabilities to support multi-class decision-making.

[0107] Gradient magnitude localization mechanism: spatial factor matrix U (2) It contains the material deformation distribution pattern, and its gradient abrupt change reflects the physical field discontinuity caused by the defect; the finite difference method approximates the spatial derivative, which conforms to the discrete data characteristics in engineering practice.

[0108] Dynamic threshold adaptation: Local statistical properties (mean and standard deviation) avoid the sensitivity of the global threshold to illumination or noise; morphological post-processing eliminates isolated noise points and improves the connectivity and robustness of the positioning mask.

[0109] Classification and localization are coordinated: the classifier identifies the defect type based on fused features, and the localization module relies on the physical interpretability of the spatial factor matrix. The two complement each other to improve the reliability of detection; the binarized mask and classification probability are jointly output to support visualization and subsequent maintenance decisions.

[0110] The aircraft defect identification system based on tensor decomposition and attention mechanism described below can be referred to in correspondence with the aircraft defect identification method based on tensor decomposition and attention mechanism described above.

[0111] Please see the appendix Figure 2 The present invention also provides an aircraft defect identification system based on tensor decomposition and attention mechanism, comprising:

[0112] A multimodal data acquisition module is used to acquire optical images, acoustic emission signals, and strain field distribution.

[0113] The tensor dynamic graph construction module is used to model multimodal data into high-dimensional tensor and dynamic graph structures;

[0114] The hybrid constraint decomposition module performs tensor decomposition optimization that includes low-rank properties, dynamic graph regularization, and physical field continuity.

[0115] A multi-scale decoupling module, combined with topological persistent homology analysis, identifies decoupling defect characteristics.

[0116] The defect identification module outputs the defect type and location.

[0117] The system in this embodiment can be used to execute the above method embodiments, and its principle and technical effect are similar, so they will not be described again here.

[0118] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An aircraft defect identification method based on tensor decomposition and attention mechanism, characterized in that, Includes the following steps: Acquire multimodal data on cracks, fissures, bulges, and dents on the surface of aircraft; Multimodal data is constructed as a fourth-order tensor, and a dynamic graph structure is built based on the modal features of the tensor. The edge weights of the dynamic graph are adaptively adjusted according to the differences in cross-modal features. The fourth-order tensor includes a time dimension, a spatial dimension, a modal dimension, and a physical quantity dimension. The time dimension represents the synchronous acquisition time series of multimodal data, the spatial dimension represents the two-dimensional gridded spatial coordinates of the aircraft surface, the modal dimension represents the distribution of optical images, acoustic emission signals, and strain fields, and the physical quantity dimension represents the optical pixel intensity, acoustic emission signal energy, and strain tensor components. The fourth-order tensor is decomposed into a tensor, and low-rank constraints, dynamic graph regularization constraints, and physical field continuity constraints are applied during the decomposition process to generate the core tensor and factor matrix. Multi-scale features are extracted from the core tensor, and weights are assigned to the multi-scale features by combining topological persistent homology analysis to obtain weighted fused multi-scale features. The defect type is identified based on the weighted fusion of multi-scale features, and the defect location is located based on the gradient change of the factor matrix.

2. The aircraft defect identification method based on tensor decomposition and attention mechanism according to claim 1, characterized in that, The steps for acquiring multimodal data on cracks, fissures, bulges, and dents on the aircraft surface include: Acquire optical images using industrial cameras; Acoustic emission signals are acquired using a multi-channel acoustic emission sensor array equipped with a synchronous triggering module; The DIC system is combined with a high frame rate camera to perform area array measurements, and the strain field distribution is collected in conjunction with the target pattern attached to the structural surface.

3. The aircraft defect identification method based on tensor decomposition and attention mechanism according to claim 1, characterized in that, The step of constructing a fourth-order tensor from multimodal data includes: Define the fourth-order tensor dimension: The time dimension T represents the sequence of synchronous acquisition time points of multimodal data, with N time points. t ; The spatial dimension H×W represents the two-dimensional gridded spatial coordinates of the aircraft surface, with height and width resolutions of H and W, respectively; The modal dimension M represents the data type of multimodal data, including optical images, acoustic emission signals, and strain field distributions; The physical quantity dimension Q represents the physical quantification index of each mode, including optical pixel intensity, acoustic emission signal energy, and strain tensor components; Tensor filling rule: Fill each time point t∈[1,N] with the tensor filling rule. t The three modal data, namely [1,H], [h,w], and [1,W], are mapped to a fourth-order tensor. The corresponding position.

4. The aircraft defect identification method based on tensor decomposition and attention mechanism according to claim 3, characterized in that, The construction of the dynamic graph structure based on the modal features of the tensor includes: Node definition: The multimodal feature vector of each spatial location (h, w) at time t. As a node in a dynamic graph, use the following formula: in, : represents the physical quantity vector of the fourth-order tensor at time t, spatial position (h, w), and mode M, where M = 1, 2, and 3 correspond to the optical image, acoustic emission signal, and strain field distribution, respectively; : represents a full slice of the physical quantity dimension; Edge weight calculation: Dynamically update the adjacency matrix based on cross-modal feature differences. The formula for calculating the edge weight is: in, Let be the edge weight between node i and node j at time t; and Let i and j represent the feature vectors of nodes i and j, respectively. This represents the mean difference in neighborhood features between nodes i and j.

5. The aircraft defect identification method based on tensor decomposition and attention mechanism according to claim 4, characterized in that, The step of performing tensor decomposition on the fourth-order tensor includes: Decomposition Model Construction: The Tucker decomposition model is used to decompose the fourth-order tensor. Decomposed into core tensors sum factor matrix And satisfy: Among them, R T R H R W R M To decompose the rank parameter; Hybrid constraint introduction: The following constraints are applied during the decomposition process: Low-rank constraint: Minimize the core tensor nuclear norm To suppress noise interference and implement dynamic graph regularization constraints: through the dynamic graph Laplacian matrix L G Constraint factor matrix U (3) ,Right now Enforcing the consistency of the factor matrix with the dynamic graph structure and physical field continuity constraints: Minimize the spatial factor matrix U (2) gradient norm Ensure that the decomposition results are consistent with the mechanical properties of materials; The following variables are updated alternately using an improved alternating direction multiplier method: Core tensor update: in, U is a tensor; k is the ADMM iteration number; λ1 is the regularization coefficient of the low-rank constraint; U (1) U (2) U (3) This is the factor matrix for the current iteration; Factor matrix update: Similarly, update U (2) U (3) , among which, T (1) For tensor The modulo-1 expansion matrix of G; (1) For the core tensor The matrix after modulo-1 expansion; Represents the Kronecker product; Dynamic graph constraint synchronous update: based on the current factor matrix U (3) Adjusting the Laplacian matrix L of the dynamic graph G .

6. The aircraft defect identification method based on tensor decomposition and attention mechanism according to claim 5, characterized in that, The step of extracting multi-scale features from the core tensor includes: Multi-scale pooling operations: Scale 1: For the core tensor Each spatial slice Max pooling is performed, where k is the modality dimension index of the core tensor, the pooling kernel is 2×2, the stride is 2, and the output feature map is obtained. Scale 2: For the core tensor Perform average pooling with a 3×3 kernel and a stride of 1, and output the feature map. Scale 3: Directly preserve the core tensor F3 serves as the original scale feature; Feature pyramid construction: Multi-scale feature maps F1, F2, and F3 are concatenated along the channel dimension to form a multi-scale feature pyramid. Where R′1 and R′2 are the spatial dimensions after pooling.

7. The aircraft defect identification method based on tensor decomposition and attention mechanism according to claim 6, characterized in that, The step of weighting the multi-scale features using topological persistent homology analysis includes: Topological persistent cohomology computation: Input: Strain field distribution data Obtained through the collection of multimodal data; Significant persistence bar generation: Calculate the persistent cohomology of the strain field E and extract the persistence index (c) of each topological feature. i ), and generate a significance index plot. Its definition is: in, Centered on (h,w); c i The topological features that fall within this neighborhood; Gated attention mechanism: Input: Fpyrami multi-scale feature pyramid and P significance index plot; Weight allocation: through a learnable parameter matrix Calculate attention weight α k : Among them, F k For the k-th scale feature map, Indicates a channel splicing operation; Feature fusion: combining weighted multi-scale features Calculate using the following formula: Where, α k F represents the attention weights at the k-th scale; k This is the feature map at the k-th scale; These are the fused multi-scale features, used as input to the subsequent classifier.

8. The aircraft defect identification method based on tensor decomposition and attention mechanism according to claim 1, characterized in that, The step of identifying the defect type based on the weighted fused multi-scale features includes: Feature vectorization: fusing features across multiple scales Flattened into a one-dimensional vector, where d = R1 × y2 × 3R3; Classifier design: Outputting the probability distribution of defect types through a fully connected network. Where C represents the number of defect categories, calculated as follows: y = Softmax(W2·ReLU(W1F)); in, is the learnable weight matrix; ReLU is the modified linear unit activation function; Softmax is the normalized exponential function, which outputs the class probability.

9. The aircraft defect identification method based on tensor decomposition and attention mechanism according to claim 8, characterized in that, The steps for locating the defect based on gradient changes in the factor matrix include: Gradient calculation: for the spatial factor matrix Calculate spatial gradient Its definition is: in, Calculated using the finite difference method in the horizontal and vertical directions respectively; Adaptive threshold localization: based on gradient magnitude Generate a binary defect mask D(h,w)∈{0,1} using the topological significance index P(h,w): Where τ(h,w)=μ P (h,w)+β·σ P (h,w) represents the dynamic threshold, μ P (h,w), σ P (h,w) represent the mean and standard deviation of the significance index P, respectively; β is the adjustment coefficient; D(h,w) indicates that there is a defect at position (h,w).

10. An aircraft defect identification system based on tensor decomposition and attention mechanism, characterized in that, The aircraft defect identification method based on tensor decomposition and attention mechanism according to any one of claims 1-9 includes: A multimodal data acquisition module is used to acquire optical images, acoustic emission signals, and strain field distribution. The Tensor Dynamic Graph Construction Module is used to model multimodal data into high-dimensional tensor and dynamic graph structures. The hybrid constraint decomposition module performs tensor decomposition optimization that includes low-rank properties, dynamic graph regularization, and physical field continuity. A multi-scale decoupling module, combined with topological persistent homology analysis, identifies decoupling defect characteristics. The defect identification module outputs the defect type and location.

Images