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

Abstract

The invention 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, and the method comprises the following steps: obtaining multi-modal data of an aircraft; constructing the multi-modal data into a fourth-order space-time-modal tensor, and generating a dynamic graph structure based on the modal characteristics of the tensor; applying mixed constraint to obtain a core tensor and a factor matrix during four-order tensor decomposition; features are extracted through multi-scale pooling, weights are distributed in combination with topological persistent coherence and a gating attention mechanism, and feature fusion is achieved; and identifying defect types based on fusion features, and positioning defect regions by using factor matrix gradient amplitudes and dynamic thresholds. According to the method, through multi-modal data fusion, dynamic graph regularization constraint and mixed tensor decomposition technologies, the detection sensitivity and the positioning precision of the small defects on the surface of the aircraft are improved, and meanwhile, the physical interpretability of the characteristics and the robustness of the algorithm to complex working conditions are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of non-destructive testing technology, and in particular to an aircraft defect recognition method and system based on tensor decomposition and attention mechanism. Background Art

[0002] Early and accurate detection of aircraft surface defects 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 unable to meet the high-sensitivity detection requirements of small surface damage and multi-physical field coupling effects. Especially in dynamic service environments, single sensor data is easily affected by noise and cannot fully characterize the multi-dimensional characteristics of defects (such as deformation, acoustic response, and strain distribution). There is an urgent need for an intelligent detection method that can unify and integrate multi-modal data and take into account physical laws and mathematical constraints.

[0003] Existing technologies for aircraft defect detection primarily utilize multimodal data segmentation processing strategies. For example, optical images and acoustic emission signals are aligned through time synchronization mechanisms, and features from each modality are extracted using independent classifiers. At the data fusion level, heterogeneous features are integrated using cascaded concatenation or weighted averaging. For strain field analysis, low-dimensional features are extracted based on traditional tensor decomposition methods, combined with threshold segmentation algorithms to locate defect areas. Additionally, some approaches attempt to incorporate graph neural networks to model sensor node associations or extract multi-resolution features through fixed-scale convolution.

[0004] However, there are still some deficiencies in the existing technology. When fusing multimodal data, the existing technology lacks a unified high-dimensional tensor modeling framework, which makes it difficult to effectively coordinate cross-modal spatiotemporal correlation features. In addition, the traditional tensor decomposition method relies too much on the mathematical low-rank assumption and ignores physical laws such as material mechanics continuity, 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 local abnormal dynamic changes under complex working conditions, and artificial threshold setting limits the adaptability to different defect types and noise environments. The present invention constructs a fourth-order spatiotemporal-modal tensor to uniformly represent multi-source data, designs a hybrid constraint tensor decomposition model to inject physical laws and dynamic graph correlation constraints, and combines a multi-scale attention mechanism with adaptive threshold positioning to effectively overcome the above-mentioned defects and achieve high-precision and robust aircraft defect detection. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides an aircraft defect identification method and system based on tensor decomposition and attention mechanism, which solves the problems in the existing technology of insufficient sensitivity in detecting tiny defects on the aircraft surface, difficulty in modeling the cross-temporal and spatial correlation of multimodal data, and high false detection rate and poor robustness caused by the separation of physical laws and mathematical constraints.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: an aircraft defect recognition method based on tensor decomposition and attention mechanism, comprising the following steps:

[0007] Obtain multimodal data of cracks, fissures, bulges, and dents on aircraft surfaces;

[0008] Constructing multimodal data into a fourth-order tensor and constructing 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 cross-modal feature differences;

[0009] Performing 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 a core tensor and a factor matrix;

[0010] Extracting multi-scale features from the core tensor, and performing weight assignment on the multi-scale features in combination with topological persistent homology analysis to obtain weighted fused multi-scale features;

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

[0012] The present invention provides an aircraft defect recognition method and system based on tensor decomposition and attention mechanism.

[0013] It has the following beneficial effects:

[0014] 1. The present invention constructs a fourth-order tensor structure to uniformly integrate multimodal data such as optical images, acoustic emission signals, and strain field distribution, effectively overcoming the limitations of single sensor data in traditional methods. At the same time, it uses dynamic graph regularization constraints and mixed constraint tensor decomposition technology to achieve collaborative optimization and noise suppression of cross-modal features, thereby improving the sensitivity and accuracy of detection of tiny defects (such as cracks and corrosion).

[0015] 2. The present invention introduces a joint optimization model of low-rank constraints, dynamic graph Laplace constraints and physical field continuity constraints to force the retention of the intrinsic spatiotemporal correlation and physical rationality of the data during the mathematical decomposition process, avoiding the 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. The present invention dynamically generates a graph structure adjacency matrix based on multimodal feature differences, adaptively captures abnormal correlation patterns in local areas (such as sudden changes in acoustic emission energy and mismatches in strain distribution), and combines graph Laplace regularization constraints to effectively solve the defect that traditional static graph models cannot adapt to changes in complex working conditions, thereby improving robustness to non-uniform background interference.

[0017] 4. The present 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 uses the gated attention mechanism to adaptively allocate weights, thereby achieving the complementary advantages of global distribution patterns and local detail features, and avoiding the problem of missed detection or false detection of small defects at a single scale.

[0018] 5. The present invention dynamically generates thresholds based on the calculation of spatial factor matrix gradient amplitude and local statistical characteristics, and combines morphological post-processing to eliminate noise interference. It does not rely on predefined thresholds or manual experience and can adapt to the defect positioning needs under different materials, lighting and noise environments, thereby improving the engineering applicability and generalization ability of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention;

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

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the specification of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0022] Please see the attached Figure 1 The embodiment of the present invention provides an aircraft defect recognition method and system based on tensor decomposition and attention mechanism, comprising the following steps:

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

[0024] Acquiring multimodal data on cracks, fissures, bulges, and dents on aircraft surfaces is accomplished through the following methods:

[0025] An industrial-grade area scan camera is used to capture optical images of the aircraft surface. The camera features a built-in synchronization trigger module to eliminate ambient light interference and ensure time synchronization. Optical images are captured across the entire area to be inspected, capturing surface texture and topographical features from multiple angles. The light source preferably utilizes a multi-band combination of wavelengths, spanning the visible to near-infrared spectrum, to enhance the differences in reflective properties of different material surfaces. The optical image generates an RGB three-channel pixel matrix at each time point, which is aligned with the remaining modal data using a timestamp.

[0026] Acoustic emission signals of aircraft structures are collected through a multi-channel acoustic emission sensor array, and sensors are placed in key load-bearing components and joints. The acoustic emission sensor has a wide-band response characteristic and can 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 through a synchronous trigger module to ensure the spatiotemporal consistency of multimodal data. After preprocessing, the acoustic emission signal is extracted to extract the time domain energy characteristics and frequency domain resonance peak parameters. The calculation formula is:

[0027]

[0028] Among them, STFT is short-time Fourier transform, which is used to convert the time domain signal into the time-frequency domain representation; s(t,h,w) is the time domain waveform of the acoustic emission signal at time t and spatial position (h,w); f min ,f max The lower and upper limits of the effective frequency range of the acoustic emission signal are determined according to the frequency response characteristics of the sensor; f is the frequency index variable, traversing [f min ,f max ] discrete frequency points within the interval.

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

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

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

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

[0033]

[0034] Where t is the time index variable, which indicates the time point number of the synchronous acquisition; h is the spatial height index variable, which indicates the row number in the two-dimensional grid space coordinate system; w is the spatial width index variable, which indicates the column number in the two-dimensional grid space coordinate system; m is the modal index variable; R(h,w), G(h,w), and B(h,w) are the red, green, and blue channel pixel intensities of the optical image at position (h,w).

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

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

[0037] Tensor representation of strain field: Mapping strain components into physical 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 to avoid feature fusion errors caused by timing deviations.

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

[0040] The multimodal data obtained in step S1 is mapped to a fourth-order tensor structure, which includes a time dimension, a space dimension, a modal dimension, and a physical quantity dimension. The time dimension represents the synchronous acquisition time series of the multimodal data, and each time point corresponds to a complete multimodal data acquisition cycle. The spatial dimension represents the physical position of the aircraft surface through a two-dimensional grid coordinate system. The grid resolution is determined by the distribution density of the digital image correlation technology (DIC) target pattern to ensure that each grid cell corresponds to a unique coordinate of the actual structure surface. The modal dimension is divided into three categories: optical image, acoustic emission signal, and strain field distribution, which carry data of different physical properties respectively. The physical quantity dimension is used to store quantitative 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 component of the strain tensor.

[0041] The filling rule of the fourth-order tensor is: for each time point t, spatial position (h, w) and mode m, fill the corresponding physical quantity vector into the tensor slice

[0042] A dynamic graph structure is constructed based on the multimodal features of the 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 association relationships. A node is defined as the multimodal feature vector of each spatial location at time t, and the calculation formula is:

[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 = 1.2.3 corresponds to the optical image, acoustic emission signal, and strain field distribution, respectively; : represents the full slice of the physical quantity dimension.

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

[0046] The weight of the edge is dynamically calculated based on the cross-modal feature differences between nodes, and the adjacency matrix elements The calculation formula is:

[0047]

[0048] in, is the edge weight between node i and node j at time t; and Represent the feature vectors of nodes i and j respectively; Represents the mean difference in neighborhood features between nodes i and j.

[0049] Generate graph Laplacian matrix L according to dynamic graph structure G , used as a regularization constraint for subsequent tensor decomposition. The calculation formula of the graph Laplacian matrix is:

[0050] L G =DA t ;

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

[0052] Among them, the structured expression 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 standardized input for subsequent tensor decomposition;

[0053] Adaptive edge weights in dynamic graphs: Dynamically adjust node connection strength based on cross-modal feature differences, effectively capturing abnormal correlation patterns in local areas;

[0054] Generation of graph Laplacian matrices: Converting dynamic graph structures into mathematical constraints ensures the coordinated optimization of the tensor decomposition process and the intrinsic correlation between the data, thereby improving the physical consistency of the decomposition results.

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

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

[0057] TuckEr decomposition model construction:

[0058] The fourth-order tensor constructed in step S2 Decomposition into core tensors and the factor matrix And meet: Among them, R T , R H , R W , R M is the decomposition rank parameter; the factor matrix U (1) 、U (2) 、U (3) They respectively characterize the time 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 jointly imposed. The objective function is defined as:

[0060]

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

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

[0063] Core tensor updates:

[0064]

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

[0066] Factor matrix update:

[0067]

[0068] Similarly update U (2) 、U (3) , where T (1) is a tensor The modulo-1 expansion matrix of G (1) Core tensor The matrix after expansion modulo -1; represents the Kronecker product;

[0069] Dynamic graph constraint synchronization update: according to the current factor matrix U (3) Adjust the dynamic graph Laplacian matrix L G , and inject it into the regularization term of the next iteration to achieve the coordinated optimization of the decomposition process and the dynamic graph structure.

[0070] The physical significance of the low-rank constraint is that aircraft defects usually appear as local anomalies, and the low-rank property of the core tensor can effectively separate the background normal mode from the defect features.

[0071] The role of dynamic graph regularization: constraining the modal interaction factor U through the dynamic graph Laplace matrix (3) , enhancing the interpretability of cross-modal correlations, such as the collaborative characterization of acoustic emission energy mutations and strain anomalies;

[0072] Introduction of physical field continuity constraints: The gradient constraint of the spatial factor matrix ensures that the decomposed strain distribution conforms to the material continuum assumption, avoiding physical discontinuities caused by mathematical decomposition;

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

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

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

[0076] The core tensor obtained from the decomposition of step S3 Multi-scale features are extracted from the core tensor, including coarse-grained, medium-grained and fine-grained features. Coarse-grained features are generated by the maximum pooling operation, and each modal slice of the core tensor is sliced Perform maximum pooling with a pooling kernel size of 2×2 and a step size of 2, and output the feature map The medium-grained features are generated by the average pooling operation, which performs average pooling on the core tensor with a pooling kernel size of 3×3 and a step size of 1, and outputs the feature map Fine-grained features directly retain the original core tensor As feature map F3. The multi-scale pooling operation captures the local details and global distribution patterns of defects through features of different granularities.

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

[0078] Based on the strain field distribution data collected in step S1 Calculate its topological persistence homology to generate a significance index map. Persistent homology extracts topological features (such as connected branches and ring holes) in the strain field by constructing the Vietoris-Rips complex and calculates the persistence of each feature (i.e., its survival time). The calculation formula is:

[0079]

[0080] in, Centered at (h,w); c i is the topological feature falling into the neighborhood; persistence(c i )=death i -birth i Represents the topological feature c i Persistence, birth i and death i are the thresholds at which the feature appears and disappears during the filtering process, respectively.

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

[0082]

[0083] Among them, F k is the k-th scale feature map, Represents a channel splicing operation.

[0084] The attention weight is normalized and limited to the interval [0, 1] to represent the contribution of each scale feature to defect detection.

[0085]

[0086] Among them, α k is the attention weight of the k-th scale; F k is the k-th scale feature map; It is the fused multi-scale feature, which is used to input the subsequent classifier.

[0087] Among them, the multi-scale pooling design: maximum pooling (coarse granularity) suppresses noise and retains significant defect features; average pooling (medium granularity) captures local average responses, balancing details and smoothness; the original scale (fine granularity) retains high-resolution spatial information to ensure that tiny defects are not missed.

[0088] Topologically persistent homology analysis: Identifying material discontinuities through topological features of the strain field (e.g., holes, connected branches), consistent with the physical fracture mechanism; the saliency map P quantifies 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); the spliced ​​saliency map P introduces physical prior knowledge to enhance the model's sensitivity to topological anomalies.

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

[0091] S5. Identify the defect type based on the weighted fusion of multi-scale features and locate the defect position 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 Input the fully connected neural network for classification, where d = R1×R2×3R3. The fully connected network contains a hidden layer and a Softmax output layer. The calculation process is:

[0094] y = Softmax(W2·ReLU(W1F));

[0095] in, is a learnable weight matrix; ReLU is the rectified linear unit activation function; Softmax is a normalized exponential function that outputs category probability.

[0096] Based on the spatial factor matrix decomposed in step S3 Calculate its spatial gradient amplitude to locate the defect area. The gradient amplitude is calculated in the horizontal direction by the finite difference method. and vertical direction Partial derivatives of , and synthesize the gradient magnitude map:

[0097]

[0098] in, The horizontal and vertical directions are calculated by finite difference method respectively.

[0099] Preferably, the dynamic threshold τ(h,w) of the defect area is adaptively determined by local statistical characteristics, and the calculation formula is:

[0100] τ(h,w)=μ P +β·σ P ;

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

[0102] The binary mask D(h,w)∈{0,1} of the defect area is generated based on the comparison result of the gradient amplitude and the dynamic threshold. The calculation formula is:

[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) are the mean and standard deviation of the significance index map P, respectively; β is the adjustment coefficient; D(h, w) indicates that there is a defect at position (h, w).

[0105] Pixels with a median value of 1 in the mask are marked as potential defect areas, and morphological closing operations (such as dilation-erosion) are used to remove isolated noise points and connect broken areas.

[0106] Among them, the fully connected network design: the hidden layer dimension h is tuned through the validation set to avoid overfitting or underfitting; the ReLU activation function introduces sparsity to enhance the nonlinear expression ability of the model; the Softmax output layer normalizes the classification results into probabilities to support multi-category decision-making.

[0107] Gradient amplitude positioning mechanism: spatial factor matrix U (2) It contains the material deformation distribution pattern, and its gradient mutation reflects the discontinuity of the physical field caused by the defect; the finite difference method approximates the spatial derivative, which is consistent with the discrete data characteristics in engineering practice.

[0108] Dynamic threshold adaptation: Local statistical characteristics (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 positioning collaboration: The classifier identifies defect types based on fused features, while the positioning module relies on the physical interpretability of the spatial factor matrix. The two complement each other to improve detection reliability. The binary mask and classification probability are jointly output to support visualization and subsequent maintenance decisions.

[0110] The aircraft defect recognition system based on tensor decomposition and attention mechanism described below and the aircraft defect recognition method based on tensor decomposition and attention mechanism described above can be referenced to each other.

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

[0112] Multimodal data acquisition module, used to obtain optical images, acoustic emission signals and strain field distribution;

[0113] Tensor dynamic graph building module, used to model multimodal data as high-dimensional tensors and dynamic graph structures;

[0114] Hybrid Constrained Decomposition module, which performs tensor decomposition optimization with low rank, dynamic graph regularization, and physics continuity;

[0115] Multi-scale decoupling module, combined with topological persistent homology analysis to decouple defect characteristics;

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

[0117] The system of this embodiment can be used to execute the above method embodiments, and its principles and technical effects are similar, so they will not be repeated here.

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

Claims

1. Aircraft defect recognition method based on tensor decomposition and attention mechanism, characterized by: The following steps are involved: Obtain multimodal data of cracks, fissures, bulges, and dents on aircraft surfaces; Constructing multimodal data into a fourth-order tensor and constructing 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 cross-modal feature differences; Performing 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 a core tensor and a factor matrix; Extracting multi-scale features from the core tensor, and performing weight assignment on the multi-scale features in combination with topological persistent homology analysis to obtain weighted fused multi-scale features; The defect type is identified according to the weighted fused multi-scale features, and the defect position is located based on the gradient change of the factor matrix.

2. The aircraft defect recognition method based on tensor decomposition and attention mechanism according to claim 1, characterized in that: The step of obtaining multimodal data of cracks, fissures, bulges and depressions on the aircraft surface includes: Collect optical images through industrial cameras; Acoustic emission signals are collected through a multi-channel acoustic emission sensor array equipped with a synchronous trigger module; The DIC system is combined with a high frame rate camera for area array measurement, and the strain field distribution is collected in conjunction with the target pattern attached to the surface of the structure.

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

4. The aircraft defect recognition method based on tensor decomposition and attention mechanism according to claim 3, characterized in that: The constructing of a dynamic graph structure based on the modal features of the tensor includes: Node definition: Multimodal feature vector of each spatial position (h, w) at time t As a dynamic graph node, 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 = 1.2.3 corresponds to the optical image, acoustic emission signal, and strain field distribution, respectively; : represents the full slice of the physical quantity dimension; Edge weight calculation: Dynamically update the adjacency matrix based on cross-modal feature differences The edge weight calculation formula is: in, is the edge weight between node i and node j at time t; and Represent the feature vectors of nodes i and j respectively; Represents the mean difference in neighborhood features between nodes i and j.

5. The aircraft defect recognition method based on tensor decomposition and attention mechanism according to claim 1, characterized in that: The step of performing tensor decomposition on the fourth-order tensor comprises: Decomposition model construction: Tucker decomposition model is used to decompose the fourth-order tensor Decomposition into core tensors and the factor matrix And meet: Among them, R T , R H , R W , R M is the decomposition rank parameter; Hybrid constraint introduction: The following constraints are imposed during the decomposition process: Low rank constraint: minimize the core tensor The nuclear norm of To suppress noise interference, dynamic graph regularization constraints: through the dynamic graph Laplace matrix L G Constraint factor matrix U (3) ,Right now Force the factor matrix to be consistent with the dynamic graph structure and the physical field continuity constraint: minimize the spatial factor matrix U (2) The gradient norm of Ensure that the decomposition results are material mechanics continuity; The following variables are updated alternately using the improved alternating direction multiplier method: Core Tensor Updates: in, is a tensor; k is the number of ADMM iterations; λ1 is the regularization coefficient of the low-rank constraint; U (1) , U (2) , U (3) is the factor matrix of the current iteration; Factor matrix update: Similarly update U (2) 、U (3) , where T (1) is the modulo-1 expansion matrix of the tensor T; G (1) Core tensor The matrix after expansion modulo -1; represents the Kronecker product; Dynamic graph constraint synchronization update: according to the current factor matrix U (3) Adjust the dynamic graph Laplacian matrix L G .

6. The method for aircraft defect recognition based on tensor decomposition and attention mechanism according to claim 1, characterized in that: The step of extracting multi-scale features from the core tensor comprises: Multi-scale pooling operation: Scale 1: Core Tensor Each spatial slice of Perform maximum pooling, where k is the modal dimension index of the core tensor, the pooling kernel is 2×2, the stride is 2, and the output feature map Scale 2: Core Tensor Perform average pooling with a pooling kernel of 3×3 and a step size of 1 to output the feature map Scale 3: directly retain the core tensor As the original scale feature F3; Feature pyramid construction: Multi-scale feature maps F1, F2, and F3 are spliced ​​along the channel dimension to form a multi-scale feature pyramid Among them, R′1 and R′2 are the spatial dimensions after pooling.

7. The aircraft defect recognition method based on tensor decomposition and attention mechanism according to claim 6, characterized in that: The step of performing weight distribution on the multi-scale features in combination with topological persistent homology analysis includes: Topologically persistent coherent computation: Input: strain field distribution data Obtained through the collected multimodal data; Significant persistence strip generation: Calculate the persistence homology of the strain field E and extract the persistence index (c i ) and generate a significance index map It is defined as: in, Centered at (h, w); c i is the topological feature that falls into the neighborhood; Gated Attention Mechanism: Input: multi-scale feature pyramid Fpyrami and saliency index map P; Weight distribution: through learnable parameter matrix Calculate the attention weight α k : Among them, F k is the k-th scale feature map, Represents a channel splicing operation; Feature fusion: weighted multi-scale features Calculated using the following formula: Among them, α k is the attention weight of the k-th scale; F k is the k-th scale feature map; It is the fused multi-scale feature, which is used to input the subsequent classifier.

8. The method for aircraft defect recognition based on tensor decomposition and attention mechanism according to claim 1, characterized in that: The step of identifying the defect type according to the weighted fused multi-scale features includes: Feature vectorization: multi-scale fusion features Flattened to a one-dimensional vector, where d = R1 × R2 × 3R3; Classifier design: Output defect type probability distribution through a fully connected network Where C is the number of defect categories, and the calculation formula is: y = Softmax(W2·ReLU(W1F)); in, is a learnable weight matrix; ReLU is the rectified linear unit activation function; Softmax is a normalized exponential function that outputs category probability.

9. The aircraft defect recognition method based on tensor decomposition and attention mechanism according to claim 8, characterized in that: The step of locating the defect position based on the gradient change of the factor matrix includes: Gradient calculation: for spatial factor matrix Calculating spatial gradients It is defined as: in, Calculated by finite difference method in horizontal and vertical directions respectively; Adaptive threshold positioning: based on gradient amplitude And the topological significance index P(h,w) generates a binary defect mask D(h,w)∈{0,1}: 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) are the mean and standard deviation of the significance index map P, respectively; β is the adjustment coefficient; D(h, w) indicates that there is a defect at position (h, w).

10. Aircraft defect recognition system based on tensor decomposition and attention mechanism, characterized by: The aircraft defect recognition method based on tensor decomposition and attention mechanism according to any one of claims 1 to 9 comprises: Multimodal data acquisition module, used to obtain optical images, acoustic emission signals and strain field distribution; Tensor dynamic graph building module, used to model multimodal data as high-dimensional tensors and dynamic graph structures; Hybrid Constrained Decomposition module, which performs tensor decomposition optimization with low rank, dynamic graph regularization, and physics continuity; Multi-scale decoupling module, combined with topological persistent homology analysis to decouple defect characteristics; The defect recognition module outputs the defect type and location.

Images