Composite structure damage morphology monitoring method and system based on deep learning

Abstract

The application discloses a composite structure damage form monitoring method and system based on deep learning, comprising the following steps: collecting multi-source monitoring data of the composite structure under the loaded state, and performing pretreatment; adopting a delay coordinate embedding method to reconstruct a damage evolution trajectory in a high-dimensional phase space, and performing dimension reduction to generate a low-dimensional trajectory of chaotic dynamics; extracting a singular attractor feature to generate a singular attractor feature set; performing sequence modeling through an improved Linformer damage identification network to generate a prediction vector; training the improved Linformer damage identification network based on the prediction vector, and introducing a nonlinear dynamics constraint to generate a damage identification network with the nonlinear dynamics constraint; and performing damage form classification and damage evolution prediction. The application combines dynamics and deep learning, realizes composite structure damage monitoring, and has the advantages of high accuracy, strong stability and reliable early warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of monitoring and intelligent diagnosis, and particularly relates to a composite structure damage form monitoring method and system based on deep learning. BACKGROUND

[0002] The existing composite structure damage monitoring methods mostly rely on single sensing methods such as acoustic emission, ultrasonic wave, vibration signal or image detection, and the damage state is determined through feature extraction and traditional pattern recognition methods. Such methods are often difficult to capture nonlinear dynamic characteristics in the early damage evolution stage, and the potential chaotic characteristics and spatiotemporal coupling relationship in complex time series are not fully utilized, resulting in limited accuracy and stability of damage identification. In addition, some methods attempt to introduce deep learning models for automatic feature extraction, but mostly stay at the time series modeling level, ignoring the frequency spectrum characteristics and dynamic manifold constraints in the damage evolution process of the composite structure, and it is difficult to effectively predict the damage evolution trend.

[0003] Based on the above deficiencies of the prior art, it is difficult to simultaneously realize the fusion modeling of multi-source data, the reconstruction of chaotic dynamic low-dimensional trajectory and the effective utilization of strange attractor features, and it is also impossible to introduce nonlinear dynamic constraints in the deep learning framework to improve the physical consistency of the model. Therefore, the existing methods have obvious defects in the accuracy, stability and early warning ability of damage form classification and damage evolution prediction, and a new monitoring method and system combining deep learning and nonlinear dynamics theory is urgently needed to solve these problems. SUMMARY

[0004] One object of the present application is to provide a composite structure damage form monitoring method based on deep learning. The present application combines dynamics and deep learning to realize composite structure damage monitoring, and has the advantages of high accuracy, strong stability and reliable early warning.

[0005] According to the composite structure damage form monitoring method based on deep learning, the method comprises the following steps:

[0006] Collecting multi-source monitoring data of the composite structure under the loaded state, performing noise filtering, normalization, time alignment and space registration processing, and generating a standardized monitoring data sequence;

[0007] Based on the standardized monitoring data sequence, the delay coordinate embedding method is used to reconstruct the damage evolution trajectory in the high-dimensional phase space, and the damage evolution time series manifold is obtained. The damage evolution time series manifold is dimensionally reduced to generate a chaotic dynamic low-dimensional trajectory;

[0008] Extracting strange attractor features in the chaotic dynamic low-dimensional trajectory to generate a strange attractor feature set;

[0009] The improved Linformer damage identification network is inputted with the singular attractor feature set, sequence modeling is performed, damage morphology feature representation is generated, and a prediction vector is generated through a full connection network;

[0010] The improved Linformer damage identification network is trained based on the prediction vector, and a nonlinear dynamics constraint is introduced as a regularization condition embedded into a joint loss function to generate a damage identification network with nonlinear dynamics constraint;

[0011] The damage identification network with nonlinear dynamics constraint is used for damage morphology classification and damage evolution prediction to generate damage categories and damage evolution trends.

[0012] Optionally, the generation of the standardized monitoring data sequence specifically includes:

[0013] Under the load state of the composite structure, acoustic emission signals, ultrasonic detection signals, vibration response signals and image data are synchronously collected to obtain multi-source monitoring data;

[0014] Noise filtering processing is performed on the acoustic emission signals, ultrasonic detection signals and vibration response signals to remove environmental interference components to obtain a noise filtered signal set;

[0015] Noise filtering processing is performed on the image data to reduce background interference and imaging artifacts in the image acquisition process to obtain noise filtered image data;

[0016] Amplitude normalization processing is performed on the noise filtered signal set to obtain a normalized signal set;

[0017] Pixel gray scale normalization processing is performed on the noise filtered image data to map the pixel gray scale range to a unified interval to obtain normalized image data;

[0018] Time alignment processing is performed on the normalized signal set to keep it consistent in time index;

[0019] Spatial registration processing is performed on the normalized image data, a reference image frame is selected, and spatial alignment of all image frames is completed through pixel-level position adjustment;

[0020] At each time index, the amplitude component of the acoustic emission signal, the amplitude component of the ultrasonic detection signal, the amplitude component of the vibration response signal and the pixel gray scale component of the image data are combined into a monitoring vector, all monitoring vectors are arranged in time sequence to generate a standardized monitoring data sequence.

[0021] Optionally, the generation of the chaotic dynamics low-dimensional trajectory specifically includes:

[0022] The delay coordinate embedding method is used to perform delay coordinate embedding processing on the standardized monitoring data sequence, embedding dimension and time delay parameters are set, the monitoring vector at each time point is combined with the monitoring vectors at several delayed time points in turn to form a high-dimensional trajectory point set containing multiple time sequence components;

[0023] The high-dimensional trajectory point set is arranged in time sequence to construct a high-dimensional phase space reconstruction matrix;

[0024] In the high-dimensional phase space reconstruction matrix, an adjacency relationship is established according to the similarity between the high-dimensional trajectory points, and adjacent trajectory points are connected to form a damage evolution time sequence manifold;

[0025] Dimensionality reduction processing is performed on the damage evolution time sequence manifold, and a nonlinear mapping method is used to map the high-dimensional trajectory points to a low-dimensional space to obtain a low-dimensional trajectory point set;

[0026] The low-dimensional trajectory point set is arranged in time index order to generate a low-dimensional trajectory of chaotic dynamics.

[0027] Optionally, the generation of the strange attractor feature set specifically includes:

[0028] The low-dimensional trajectory point set corresponding to the low-dimensional trajectory of chaotic dynamics is obtained, geometric distribution processing is performed on the low-dimensional trajectory point set to generate geometric morphological features, and the geometric morphological features include the average Euclidean distance between trajectory points, trajectory curvature distribution and trajectory envelope area;

[0029] The low-dimensional trajectory point set is subjected to adjacency relationship processing, a topological structure is established, and topological features are generated, and the topological features include node degree distribution, number of connected components and local clustering coefficient;

[0030] The low-dimensional trajectory point set is subjected to time index processing, a state transition matrix is constructed, and feature decomposition is performed to extract stability features, and the stability features include the maximum Lyapunov exponent, the spectral radius of the state transition matrix and the convergence rate of the principal eigenvector;

[0031] The geometric morphological features, topological features and stability features are subjected to normalization processing, and the normalized features are combined to generate a strange attractor feature set.

[0032] Optionally, the generation of the prediction vector specifically includes:

[0033] The strange attractor feature set is input into an improved Linformer damage identification network, and the improved Linformer damage identification network includes an input feature construction layer, a frequency domain transformation layer, a low-rank projection layer, a manifold coupling attention layer and a multi-layer stacked modeling layer;

[0034] In the input feature construction layer, the singular attractor feature set is arranged in time index order into an input feature matrix;

[0035] The input feature matrix is subjected to Fourier transform in the spectrum domain transformation layer to obtain a spectrum domain feature matrix;

[0036] The spectrum domain feature matrix is taken as the input of the low-rank projection layer to perform low-rank projection processing to obtain a query matrix, a key matrix and a value matrix, wherein the low-rank projection processing refers to multiplying the input of the low-rank projection layer with a corresponding low-rank projection weight matrix;

[0037] In the manifold coupled attention layer, a manifold coupled attention mechanism is used to calculate similarity weights based on the query matrix and the key matrix, and the similarity weights are weighted and fused with an adjacency relationship matrix constructed by the low-dimensional trajectory point set to generate a manifold coupled attention weight matrix after normalization processing:

[0038] ;

[0039] wherein, the manifold coupled attention weight matrix is represented by M, the normalization is represented by N, the key matrix is represented by K, the query matrix is represented by Q, the transposition operation is represented by T, the scaling factor is represented by S, the hyperparameter for adjusting the influence degree of the manifold adjacency information and the similarity weights is represented by a, the adjacency relationship matrix is represented by A;

[0040] The manifold coupled attention weight matrix is multiplied with the value matrix to obtain an attention weighted representation matrix, and the attention weighted representation matrix is subjected to residual connection and layer normalization with the input of the low-rank projection layer to generate a first layer feature representation;

[0041] In the multi-layer stacked modeling layer, the first layer feature representation is taken as the input of the low-rank projection layer to recalculate the query matrix, the key matrix and the value matrix, and the manifold coupled attention mechanism is repeated to obtain a multi-layer feature representation sequence, and the final layer feature representation is taken as the damage mode feature representation and subjected to full connection network processing to generate a prediction vector.

[0042] Optionally, the generation of the nonlinear dynamic constraint damage identification network specifically includes:

[0043] The improved Linformer damage identification network is trained;

[0044] A nonlinear dynamic equation is established as a reference evolution relationship;

[0045] A prediction error metric is constructed using the mean square error between the prediction vector and the true label vector;

[0046] The prediction vector is substituted into the nonlinear dynamics equation, and a dynamics consistency constraint metric is constructed;

[0047] The prediction error metric and the dynamics consistency constraint metric are combined by weighting, and a joint loss function is established;

[0048] Based on the joint loss function, the improved Linformer damage identification network is updated using the gradient descent method, and the network weight matrix and bias parameter are iteratively optimized until the joint loss function converges or the training round reaches the preset upper limit, generating a nonlinear dynamics constraint damage identification network.

[0049] Optionally, the generation of the damage category and the damage evolution trend specifically includes:

[0050] The nonlinear dynamics constraint damage identification network is called, and a prediction vector is obtained based on the real-time collected multi-source monitoring data, the prediction vector including a damage category probability value and a damage evolution intensity value;

[0051] The category corresponding to the maximum probability of the damage category probability value is taken as the damage category;

[0052] The damage evolution trend is composed of the damage evolution intensity value.

[0053] According to the composite structure damage morphology monitoring system based on deep learning, the composite structure damage morphology monitoring system based on deep learning comprises:

[0054] The acquisition and preprocessing module is used for acquiring multi-source monitoring data of the composite structure under a loaded state and generating a standardized monitoring data sequence;

[0055] The time sequence trajectory reconstruction module is used for reconstructing a damage evolution time sequence manifold based on the standardized monitoring data sequence and generating a chaotic dynamics low-dimensional trajectory;

[0056] The feature extraction module is used for extracting a singular attractor feature set from the chaotic dynamics low-dimensional trajectory;

[0057] The improved Linformer damage identification network is used for receiving the singular attractor feature set and generating a prediction vector;

[0058] The nonlinear dynamics constraint training module is used for establishing a nonlinear dynamics constraint and training the improved Linformer damage identification network in combination with a joint loss function, and generating a nonlinear dynamics constraint damage identification network;

[0059] The prediction output module is used for calling the nonlinear dynamics constraint damage identification network to generate a damage category and a damage evolution trend.

[0060] The beneficial effects of this invention are:

[0061] This invention introduces a technical approach combining chaotic dynamics and deep learning. By synchronously acquiring and standardizing multi-source monitoring data, it unifies acoustic emission signals, ultrasonic signals, vibration response signals, and image data into a time-series standardized monitoring data sequence. This overcomes the shortcomings of single sensing methods in terms of information content and robustness, and achieves comprehensive capture of multi-dimensional and multi-modal features of the damage process of composite structures, providing high-quality input for subsequent dynamic trajectory reconstruction and feature extraction.

[0062] Secondly, based on standardized monitoring data, this invention reconstructs high-dimensional phase space trajectories using a time-delayed coordinate embedding method and generates low-dimensional chaotic dynamic trajectories through dimensional reduction. This effectively reveals the nonlinear evolution law of composite structures during loading. This process can not only reconstruct damage evolution trajectories with chaotic characteristics, but also extract a set of singular attractor features at the geometric, topological, and stability levels. This allows the monitoring results to reflect the complex spatiotemporal behavior characteristics of the damage. Compared with traditional methods that rely solely on time-domain or frequency-domain analysis, this invention enhances the ability to reveal the essential characteristics of the damage.

[0063] Furthermore, this invention uses an improved Linformer damage recognition network to perform sequential modeling of the singular attractor feature set. It introduces a spectral domain transformation and a dynamic manifold coupled attention mechanism, which overcomes the limitation of existing deep learning methods that only directly model in the temporal space. The spectral domain transformation reveals potential periodic perturbations and energy distributions, while the dynamic manifold coupled attention mechanism embeds the adjacency constraint of low-dimensional trajectories in the calculation of attention weights. This allows the damage morphology feature representation to take into account both spectral space and dynamic manifold features. Through this improvement, the model can identify weak nonlinear damage signs in the early stages, improving the sensitivity and stability of the monitoring results.

[0064] Finally, this invention introduces nonlinear dynamic constraints into model training, establishes a joint loss function, and weights and fuses the prediction error with the dynamic consistency metric to ensure that the network output not only approximates the true label numerically, but also maintains consistency with the nonlinear dynamic characteristics in terms of evolution law. This physical consistency constraint avoids the overfitting and physical distortion problems that may occur in pure data-driven models, making the generated damage recognition network more reliable and practical in damage morphology classification, evolution trend prediction, and disaster risk early warning. Attached Figure Description

[0065] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0066] Fig. 1 This is a flowchart of a deep learning-based method for monitoring the damage morphology of composite structures proposed in this invention.

[0067] Fig. 2 This is a schematic diagram of the improved Linformer damage recognition network generating prediction vectors for a deep learning-based composite structure damage morphology monitoring method proposed in this invention. Detailed Implementation

[0068] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0069] refer to Figs. 1-2 A deep learning-based method for monitoring the damage morphology of composite structures includes the following steps:

[0070] Collect multi-source monitoring data of the composite structure under load, perform noise filtering, normalization, time alignment and spatial registration processing to generate a standardized monitoring data sequence;

[0071] Based on standardized monitoring data sequences, the damage evolution trajectory is reconstructed in high-dimensional phase space using a delayed coordinate embedding method to obtain the damage evolution time-series manifold. Dimension reduction is then performed on the damage evolution time-series manifold to generate a low-dimensional trajectory of chaotic dynamics.

[0072] Extracting singular attractor features from low-dimensional trajectories in chaotic dynamics to generate a set of singular attractor features;

[0073] The strange attractor feature set is input into the improved Linformer damage recognition network for sequence modeling, generating damage morphology feature representation, and then generating prediction vectors through a fully connected network.

[0074] The improved Linformer damage recognition network is trained based on the prediction vector, and nonlinear dynamic constraints are introduced as regularization conditions to embed the joint loss function, thereby generating a damage recognition network with nonlinear dynamic constraints.

[0075] A damage identification network with nonlinear dynamic constraints is used to classify damage morphology and predict damage evolution, generating damage categories and damage evolution trends.

[0076] In this embodiment, the generation of the standardized monitoring data sequence specifically includes:

[0077] Under the condition of composite structure under load, acoustic emission signals, ultrasonic detection signals, vibration response signals and image data are collected simultaneously to obtain multi-source monitoring data;

[0078] Noise filtering is performed on acoustic emission signals, ultrasonic detection signals, and vibration response signals to remove environmental interference components and obtain a set of noise-filtered signals.

[0079] Noise filtering is performed on the image data to reduce background interference and imaging artifacts during image acquisition, resulting in noise-filtered image data.

[0080] The amplitude normalization process is performed on the noise-filtered signal set to obtain a normalized signal set;

[0081] Pixel grayscale normalization is performed on the noise-filtered image data to map the pixel grayscale range to a uniform interval, resulting in normalized image data.

[0082] Perform time alignment on the normalized signal set to keep it consistent in time index;

[0083] Spatial registration processing is performed on the normalized image data. A reference image frame is selected, and spatial alignment of all image frames is completed through pixel-level position adjustment.

[0084] Under each time index, the amplitude components of the acoustic emission signal, the ultrasonic detection signal, the vibration response signal, and the pixel grayscale components of the image data are combined into a monitoring vector. All monitoring vectors are arranged in chronological order to generate a standardized monitoring data sequence.

[0085] In this embodiment, the generation of the low-dimensional trajectory of chaotic dynamics specifically includes:

[0086] The standardized monitoring data sequence is processed using a delayed coordinate embedding method. The embedding dimension and time delay parameter are set, and the monitoring vector at each time point is sequentially combined with the monitoring vectors at several delayed time points to form a high-dimensional trajectory point set containing multiple time-series components. Specifically, the embedding dimension and time delay parameter are determined, where the embedding dimension defines the number of time-series components contained in each high-dimensional trajectory point, and the time delay parameter determines the interval between adjacent time-series components. Under each time index, the monitoring vector at that time point is extracted and sequentially combined with the monitoring vectors of several preceding delayed time points to obtain a high-dimensional trajectory point containing that time point and its delay history. This process is repeated for all time indices of the standardized monitoring data sequence to ultimately form a high-dimensional trajectory point set. This high-dimensional trajectory point set is a collection of all high-dimensional state vectors obtained from the standardized monitoring data sequence using the delayed coordinate embedding method, used to characterize the temporal evolution of the composite structure in a high-dimensional phase space.

[0087] The high-dimensional trajectory point set is arranged sequentially in time order to construct a high-dimensional phase space reconstruction matrix, so that each row of the high-dimensional phase space reconstruction matrix corresponds to a high-dimensional trajectory point and each column corresponds to a certain monitoring component or its delayed state, thereby completely representing the temporal state of the composite structure damage evolution in the matrix.

[0088] In the high-dimensional phase space reconstruction matrix, adjacency relationships are established based on the similarity metric between high-dimensional trajectory points. Adjacent trajectory points are connected to form a damage evolution time-series manifold. Specifically, each row in the matrix is ​​considered a high-dimensional trajectory point, representing the temporal state of the composite structure at a certain time index. The similarity metric between any two trajectory points is calculated, represented by the Euclidean distance between them; a smaller value indicates closer proximity. For each trajectory point, several nearest trajectory points are selected as its neighbors, constructing an adjacency set. Based on this, all trajectory points and their adjacency sets are connected in temporal order to form a topological structure composed of points and edges. This topological structure serves as the damage evolution time-series manifold of the composite structure, characterizing the continuous evolution of standardized monitoring data sequences in the high-dimensional phase space.

[0089] A dimensionality reduction process is performed on the temporal manifold of damage evolution. A nonlinear mapping method is used to map high-dimensional trajectory points to a low-dimensional space, resulting in a low-dimensional trajectory point set. Specifically, the dimension of the low-dimensional target space is selected, and the dimension of the low-dimensional target space is less than the embedding dimension. Based on the high-dimensional trajectory point set, a similarity matrix is ​​constructed between all trajectory points. The similarity matrix is ​​measured by the Euclidean distance between trajectory points and is used to characterize the adjacency relationship of the high-dimensional trajectory point set. A nonlinear mapping method is used to reduce the dimensionality of the similarity matrix by preserving the local neighborhood structure of the trajectory points, mapping each high-dimensional trajectory point to the low-dimensional target space. After mapping, all mapped points form a low-dimensional trajectory point set. This low-dimensional trajectory point set is the set of all low-dimensional state vectors formed after mapping the high-dimensional trajectory point set to the low-dimensional target space using the nonlinear dimensionality reduction method. It represents the temporal trajectory distribution of the composite structure damage evolution process in the low-dimensional target space, and each low-dimensional trajectory point corresponds to the damage state of the composite structure within a specific time window.

[0090] The set of low-dimensional trajectory points is arranged in time index order to generate a low-dimensional trajectory of chaotic dynamics, which reflects the nonlinear dynamic law of damage evolution of composite structure with a simplified continuous expression.

[0091] In this embodiment, the generation of the strange attractor feature set specifically includes:

[0092] Obtain the set of low-dimensional trajectory points corresponding to the low-dimensional trajectory of chaotic dynamics, perform geometric distribution processing on the set of low-dimensional trajectory points, and generate geometric morphological features, including the average Euclidean distance between trajectory points, trajectory curvature distribution, and trajectory envelope area.

[0093] The generation of the geometric morphological features specifically includes: in the low-dimensional trajectory point set, selecting any two trajectory points in sequence, extracting the corresponding low-dimensional state vectors, calculating the sum of the squares of the differences between the two vectors and taking the square root to obtain the Euclidean distance, summing the Euclidean distances of all trajectory point pairs and dividing by the total number of trajectory point pairs to obtain the average Euclidean distance between trajectory points; selecting three adjacent trajectory points in time index order, calculating the trajectory curvature value based on the vector angle relationship, and sequentially counting all curvature values ​​on the entire low-dimensional trajectory point set to form a trajectory curvature distribution; performing convex hull operation on the low-dimensional trajectory point set to obtain the smallest convex polygon containing all trajectory points, and calculating the area of ​​the convex polygon as the area of ​​the trajectory envelope region to generate the geometric morphological features;

[0094] Adjacency processing is performed on the low-dimensional trajectory point set to establish a topological structure and generate topological features, including node degree distribution, number of connected components, and local clustering coefficients.

[0095] The generation of the topological features specifically includes: establishing adjacency relationships based on similarity metrics between trajectory points, constructing a topological structure from all trajectory points and their adjacency relationships; then performing node degree calculation on the topological structure, counting the number of adjacent trajectory points for each trajectory point to form a node degree distribution; performing connectivity analysis on the topological structure to determine whether all trajectory points can be connected into a single subgraph through adjacency relationships, counting the number of independent connected subgraphs to obtain the number of connected components; finally performing local clustering analysis on the topological structure, calculating the ratio of the actual number of edges between each trajectory point and the maximum number of edges that can be formed, and taking the average value of all trajectory points as the local clustering coefficient. The node degree distribution refers to the statistical distribution formed by the number of adjacent trajectory points of all trajectory points in the topological structure constructed from the low-dimensional trajectory point set, used to characterize the overall connectivity characteristics of the trajectory points.

[0096] A time indexing process is performed on the low-dimensional trajectory point set to construct a state transition matrix and perform eigenvalue decomposition to extract stability features, including the maximum Lyapunov exponent, the spectral radius of the state transition matrix, and the convergence rate of the principal eigenvector.

[0097] The generation of the stability features specifically includes: constructing a state transition matrix according to the time index order, representing the state transition relationship between adjacent trajectory points in matrix form; based on the state transition matrix, calculating the maximum Lyapunov exponent through trajectory point perturbation evolution analysis, specifically by applying a small perturbation to the low-dimensional state vector at the initial time, tracking the growth of the perturbation vector under iterative state transitions, calculating the logarithmic rate of change of the perturbation vector length with the number of iterations, and taking its limit as the maximum Lyapunov exponent, which is used to characterize the sensitivity of the system to the initial conditions; performing eigenvalue decomposition on the state transition matrix, extracting the magnitude of all eigenvalues, and taking the maximum value as the spectral radius, which is used to characterize the stability boundary of the system during the iteration process; iteratively calculating the principal eigenvector of the state transition matrix, statistically analyzing the decay rate of the difference between the principal eigenvector and the stable direction with the number of iterations during the iteration process, and taking the decay rate as the convergence rate of the principal eigenvector, which is used to characterize the speed at which the system tends to a stable state;

[0098] Geometric features, topological features, and stability features are normalized, and the normalized features are combined to generate a set of singular attractor features.

[0099] In this embodiment, the generation of the prediction vector specifically includes:

[0100] The set of singular attractor features is input into an improved Linformer damage recognition network, which includes an input feature construction layer, a spectral domain transformation layer, a low-rank projection layer, a manifold coupled attention layer, and a multi-layer stacked modeling layer.

[0101] The improvement of the Linformer damage identification network lies in that it no longer limits itself to directly performing sequence modeling on the singular attractor feature set in the time domain. Instead, it performs Fourier transform on the singular attractor feature set to construct a spectral domain feature matrix, enabling the input features to reveal the potential periodic perturbations and energy distribution characteristics in the evolution of composite structure damage in the frequency dimension. On this basis, the network's low-rank self-attention mechanism further introduces an adjacency relation matrix based on a dynamic manifold constructed from a set of low-dimensional trajectory points. This integrates the physical topological constraints between trajectory points with the attention weight calculation process, forming a dynamic manifold coupled attention mechanism. This ensures that the generated damage morphology feature vector satisfies the dual constraints of spectral space features and dynamic manifold features when characterizing composite structure damage. Through this improvement, the network can capture the weak nonlinear evolution characteristics of composite structure damage in the early stage and improve the stability and predictability of damage monitoring results.

[0102] In the input feature construction layer, the set of singular attractor features is arranged into an input feature matrix in time index order;

[0103] The input feature matrix is ​​subjected to a Fourier transform through a spectral domain transform layer to obtain a spectral domain feature matrix, which is used to characterize the energy distribution of the singular attractor feature set in the spectral space.

[0104] The spectral domain feature matrix is ​​used as the input to the low-rank projection layer. Low-rank projection processing is performed to obtain the query matrix, key matrix, and value matrix. The low-rank projection processing refers to multiplying the input of the low-rank projection layer with the corresponding low-rank projection weight matrix.

[0105] In the manifold coupled attention layer, a manifold coupled attention mechanism is used. Similarity weights are calculated based on the query matrix and key matrix, and then weighted and fused with an adjacency relation matrix constructed from a low-dimensional trajectory point set. After normalization, a manifold coupled attention weight matrix is ​​generated. The adjacency relation matrix represents the topological connectivity between low-dimensional trajectory points, and fusion hyperparameters are used to adjust the influence between adjacency relations and similarity weights.

[0106] ;

[0107] in, This represents the manifold-coupled attention weight matrix. Indicates normalization, Represents the key matrix, Represents the query matrix. This indicates the transpose operation. Indicates the scaling factor. The hyperparameter represents the degree to which manifold adjacency information and similarity weights influence each other. Represents the adjacency matrix;

[0108] The manifold coupled attention mechanism introduces the manifold structure information contained in the low-dimensional trajectory of chaotic dynamics into the attention weight generation process during the low-rank self-attention calculation process, thereby realizing the constrained modeling of damage evolution characteristics.

[0109] The generation of the adjacency relation matrix specifically includes: obtaining a low-dimensional set of trajectory points, sequentially calculating the Euclidean distance between any two trajectory points, determining trajectory point pairs with an Euclidean distance less than a preset similarity threshold as adjacent trajectory point pairs, and assigning a value of one to the corresponding position in the adjacency relation matrix; determining trajectory point pairs with a distance greater than or equal to the preset similarity threshold as non-adjacent trajectory point pairs, and assigning a value of zero to the corresponding position in the adjacency relation matrix; and establishing adjacency relations between all trajectory points through this process to obtain the adjacency relation matrix.

[0110] Multiplying the manifold-coupled attention weight matrix by the value matrix yields the attention weighted representation matrix. This attention weighted representation matrix is ​​then coupled with the input of the low-rank projection layer via a residual connection and layer normalization to generate the first layer feature representation.

[0111] ;

[0112] in, This represents the first layer of feature representation. Representation layer normalization, This represents the attention-weighted representation matrix. Represents the characteristic matrix in the spectral domain. Represents the input feature matrix;

[0113] In the multi-layer stacked modeling layer, the first layer feature representation is used as the input of the low-rank projection layer. The query matrix, key matrix, and value matrix are recalculated, and the manifold coupling attention mechanism is repeated to obtain a multi-layer feature representation sequence. The final layer feature representation is used as the damage morphology feature representation. After processing by a fully connected network, a prediction vector is generated.

[0114] In this embodiment, the generation of the damage recognition network based on nonlinear dynamic constraints specifically includes:

[0115] Train the improved Linformer damage recognition network;

[0116] Establish a nonlinear dynamic equation as a reference evolution relationship:

[0117] ;

[0118] in, Indicates time index The reference vector below, Indicates time index The reference vector below, Represents the linear evolution coefficients. Represents the nonlinear quadratic coupling coefficient;

[0119] The generation of the linear evolution coefficient and the nonlinear quadratic coupling coefficient specifically includes: taking the low-dimensional trajectory point set generated by delay coordinate embedding and dimensional reduction of the standardized monitoring data sequence as input data, selecting trajectory point pairs with adjacent time indices as training samples, constructing a least squares optimization problem to fit the nonlinear dynamic equation, in which the linear evolution coefficient and the nonlinear quadratic coupling coefficient are set as parameters to be determined, and obtaining the optimal parameter estimates by iteratively minimizing the square error between the predicted value and the actual value, and finally forming the linear evolution coefficient and the nonlinear quadratic coupling coefficient that accurately describe the nonlinear evolution relationship of the low-dimensional trajectory point set;

[0120] The prediction error metric is constructed using the mean squared error between the predicted vector and the true label vector.

[0121] ;

[0122] in, This represents a measure of prediction error. Represents the total number of samples. Indicates time index The actual label vector below, Indicates time index The predicted vector below;

[0123] Substitute the predicted vector into the nonlinear dynamic equations and construct a dynamic consistency constraint metric:

[0124] ;

[0125] in, This represents a measure of dynamic consistency constraints. Indicates time index The predicted vector below;

[0126] A joint loss function is established by weighting and combining the prediction error metric and the dynamic consistency constraint metric:

[0127] ;

[0128] in, Denotes the joint loss function. The weighting coefficients represent the measurement of prediction error. The weighting coefficients representing the dynamic consistency constraint measure;

[0129] The weight coefficients are determined through cross-validation, specifically including: first, dividing the singular attractor feature set into a training set and a validation set; then, setting different candidate weight coefficient values ​​on the training set, and weighting the prediction error metric and the dynamic consistency constraint metric to obtain multiple candidate joint loss functions; iteratively optimizing the network parameters using the training set, and calculating the value of each candidate joint loss function on the validation set to obtain candidate joint loss function values ​​under different weight coefficient combinations; selecting the weight coefficient corresponding to the smallest candidate joint loss function value as the final weight coefficient, and fixing it for the weighted combination of the joint loss function, thereby ensuring that prediction accuracy and dynamic constraint consistency are considered simultaneously during the training process;

[0130] Based on the joint loss function, gradient descent is used to update the parameters of the improved Linformer damage recognition network, iteratively optimizing the network weight matrix and bias parameters until the joint loss function converges or the training epochs reach a preset upper limit, thus generating a nonlinear dynamically constrained damage recognition network.

[0131] In this embodiment, the generation of the damage category and damage evolution trend specifically includes:

[0132] A damage identification network constrained by nonlinear dynamics is invoked to obtain a prediction vector based on real-time multi-source monitoring data. The prediction vector includes a damage category probability value and a damage evolution intensity value.

[0133] The specific categories of damage include:

[0134] The undamaged state indicates that no significant damage was detected in the composite structure, and it is in normal working condition.

[0135] Minor damage: This indicates that the composite structure has early microcracks, slight fiber breakage, or slight delamination of the matrix, but has not yet affected the overall load-bearing capacity.

[0136] Moderate damage: This indicates that obvious crack propagation, fiber breakage accumulation, or local delamination have formed inside the composite structure, and the structural stiffness has decreased.

[0137] Severe damage: This indicates that the composite structure has large-scale cracks, obvious delamination or fiber bundle damage, the structural performance is greatly reduced, and there is a risk of failure.

[0138] Failure state: This indicates that the composite structure cannot maintain its load-bearing capacity under the design conditions, and is experiencing catastrophic damage or is about to fail as a whole.

[0139] The category corresponding to the highest probability of the damage category probability value is taken as the damage category, and the damage category probability value is used to characterize the probability of the occurrence of the composite structure under different damage categories.

[0140] The damage evolution trend is constructed from the damage evolution intensity values. Specifically, the damage evolution intensity values ​​are arranged into a continuous numerical sequence according to the time index, which serves as the basis for the damage evolution trend. Time series analysis is performed on the continuous numerical sequence to calculate the rate of change and cumulative change amplitude between adjacent time points. When the rate of change shows a continuous increase, it is determined that the damage is in an accelerated development stage. When the rate of change tends to level off, it is determined that the damage evolution has entered a stable stage. When the rate of change decreases, it is determined that the damage evolution has slowed down. The damage evolution intensity value change sequence over time is used as the damage evolution trend to characterize the evolution process of the composite structure from early minor damage to mid-term expansion and even potential catastrophic events in the later stage, thereby realizing dynamic monitoring of the direction and speed of damage development in the composite structure.

[0141] A deep learning-based composite structure damage morphology monitoring system, comprising:

[0142] The acquisition and preprocessing module is used to acquire multi-source monitoring data of the composite structure under load and generate standardized monitoring data sequences.

[0143] The temporal trajectory reconstruction module is used to reconstruct the temporal manifold of damage evolution based on standardized monitoring data sequences and generate low-dimensional trajectories of chaotic dynamics.

[0144] The feature extraction module is used to extract a set of singular attractor features from low-dimensional trajectories in chaotic dynamics;

[0145] An improved Linformer damage recognition network is used to receive a set of singular attractor features and generate a prediction vector;

[0146] The nonlinear dynamics constraint training module is used to establish nonlinear dynamic constraints and train an improved Linformer damage recognition network in combination with a joint loss function, thereby generating a damage recognition network with nonlinear dynamic constraints.

[0147] The prediction output module is used to call the damage identification network with nonlinear dynamic constraints to generate damage categories and damage evolution trends.

[0148] Example 1:

[0149] To verify the feasibility of this invention in practice, it was applied to the health monitoring of composite material components in an aerospace manufacturing company. This company has long faced the problem of difficulty in timely detection of internal crack initiation in composite materials during the manufacturing and service stages of wing spars. Traditional ultrasonic non-destructive testing and manual inspection methods often rely on fixed-cycle testing, resulting in delayed detection results that are easily affected by noise environments and cannot accurately capture the evolution trend of internal damage in composite materials. This invention aims to solve the problem of existing detection technologies being insensitive to early damage in complex structures and having difficulty predicting evolution trends by constructing a deep learning-based composite structure damage morphology monitoring method.

[0150] In practical applications, researchers set up a controlled load environment in the test hall to continuously monitor multiple composite structural specimens under different loading conditions. By deploying acoustic emission sensors, ultrasonic probes, accelerometers, and high-resolution cameras, they collected multi-source monitoring data, including acoustic emission signals, ultrasonic detection signals, vibration response signals, and surface images. Subsequently, the data was sent to the acquisition and preprocessing module to perform noise filtering, normalization, time alignment, and spatial registration, generating a standardized monitoring data sequence. This ensured that all signals remained consistent in both time and space, eliminating the impact of equipment differences and environmental interference.

[0151] Building upon this foundation, the system reconstructs the high-dimensional phase space evolution trajectory of the composite structure using a delayed coordinate embedding method. It then generates a damage evolution temporal manifold by establishing adjacency relationships. To facilitate subsequent in-depth modeling, the system further employs a nonlinear dimensionality reduction method to map the high-dimensional trajectory to a low-dimensional space, obtaining a low-dimensional trajectory of chaotic dynamics. Through this low-dimensional trajectory, researchers can intuitively observe the dynamic changes of the composite material under different loading stages and extract a set of singular attractor features, including geometric morphological features, topological features, and stability features. These features effectively characterize the dynamic changes of the composite structure during the initiation, propagation, and stabilization stages of internal microcracks, providing reliable input for subsequent network modeling.

[0152] In the modeling phase, the singular attractor feature set is input into the improved Linformer damage recognition network. Unlike traditional methods, this invention introduces a spectral domain feature matrix coupled with a dynamic manifold attention mechanism, enabling the network to simultaneously consider topological constraints during damage evolution while capturing signal energy distribution features. Through residual connections and layer normalization design, the network ensures the stability of the training process and progressively extracts deeper damage morphology features in multi-layer stacked modeling. Finally, the system generates a prediction vector containing damage category probability, damage evolution intensity, and catastrophic risk probability score, providing researchers with a quantitative assessment of structural health status.

[0153] During training, the system further introduces nonlinear dynamic constraints and establishes a joint loss function, so that the predicted output not only conforms to the label distribution of historical monitoring data, but also maintains consistency with the actual evolution law of the composite structure in terms of dynamic consistency. Through iterative optimization on multiple samples, the network parameters converge continuously, forming a damage recognition network with nonlinear dynamic constraints. This improvement effectively overcomes the limitations of traditional deep learning models that rely solely on data fitting and lack physical constraints, enabling the prediction results to still have credibility and robustness in complex environments.

[0154] To verify the performance of the present invention, it was compared with the traditional method. The comparison results are shown in Table 1.

[0155] Table 1. Performance Comparison of the Invention and Traditional Methods

[0156]

[0157] As can be seen from Table 1, the method of the present invention has significant advantages over traditional methods in the identification and prediction of damage in composite structures. In terms of damage identification accuracy, the present invention achieves 91.8%, which is about 10.5 percentage points higher than the traditional ultrasonic detection method and 5.3 percentage points higher than the traditional deep learning method. This improvement comes from the deep fusion of multi-source monitoring data in the present invention, as well as the capture of small nonlinear damage features inside the composite structure through delayed coordinate embedding and low-dimensional trajectory construction of chaotic dynamics, thereby achieving higher accuracy identification.

[0158] In terms of the accuracy of damage evolution trend prediction, the method of this invention is higher than that of traditional methods, reaching 88.3%, while the traditional deep learning method is only 74.1%. The fundamental reason for this difference is that this invention not only relies on data-driven model training, but also introduces nonlinear dynamic constraints, embedding the laws of physical evolution into the network learning process, avoiding the problem of bias accumulation in long-term prediction of pure deep learning models. Therefore, it can describe the damage development trend more accurately.

[0159] Regarding early warning capabilities, this invention achieves an average early warning time of 4.7 hours, a significant improvement compared to the traditional method's less than 1.2 hours. This means that during actual structural operation, engineers have ample time to take maintenance measures, effectively reducing the risk of catastrophic damage. This improvement stems from the embedding mechanism of the catastrophic risk probability score in the prediction vector, enabling the system to capture signals of damage transitioning from a stable to an unstable state earlier.

[0160] This invention also demonstrates better control over the false alarm rate, at only 6.4%, a significant decrease compared to the 14.2% of traditional ultrasonic methods. This is mainly attributed to the multi-dimensional extraction of exotic attractor features, which effectively distinguishes real damage signals from noise or environmental disturbance signals, thereby reducing misjudgments.

[0161] In terms of computation time, this invention achieves 6.9 seconds per batch, which is slightly lower than traditional deep learning methods and significantly better than the 9.8 seconds per batch of traditional ultrasonic detection methods. This advantage is attributed to the low-rank projection and manifold coupled attention mechanism used in the improved Linformer network structure, which reduces computational complexity while ensuring modeling accuracy and improving the feasibility of real-time monitoring.

[0162] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring the damage morphology of composite structures based on deep learning, characterized in that, Includes the following steps: Collect multi-source monitoring data of the composite structure under load, perform noise filtering, normalization, time alignment and spatial registration processing to generate a standardized monitoring data sequence; Based on standardized monitoring data sequences, the damage evolution trajectory is reconstructed in high-dimensional phase space using a delayed coordinate embedding method to obtain the damage evolution time-series manifold. Dimension reduction is then performed on the damage evolution time-series manifold to generate a low-dimensional trajectory of chaotic dynamics. Extracting singular attractor features from low-dimensional trajectories in chaotic dynamics to generate a set of singular attractor features; The strange attractor feature set is input into the improved Linformer damage recognition network for sequence modeling, generating damage morphology feature representation, and then generating prediction vectors through a fully connected network. The improved Linformer damage recognition network is trained based on the prediction vector, and nonlinear dynamic constraints are introduced as regularization conditions to embed the joint loss function, thereby generating a damage recognition network with nonlinear dynamic constraints. A damage identification network with nonlinear dynamic constraints is used to classify damage morphology and predict damage evolution, generating damage categories and damage evolution trends. The generation of the standardized monitoring data sequence specifically includes: Under the condition of composite structure under load, acoustic emission signals, ultrasonic detection signals, vibration response signals and image data are collected simultaneously to obtain multi-source monitoring data; Noise filtering is performed on acoustic emission signals, ultrasonic detection signals, and vibration response signals to remove environmental interference components and obtain a set of noise-filtered signals. Noise filtering is performed on the image data to reduce background interference and imaging artifacts during image acquisition, resulting in noise-filtered image data. The amplitude normalization process is performed on the noise-filtered signal set to obtain a normalized signal set; Pixel grayscale normalization is performed on the noise-filtered image data to map the pixel grayscale range to a uniform interval, resulting in normalized image data. Perform time alignment on the normalized signal set to keep it consistent in time index; Spatial registration processing is performed on the normalized image data. A reference image frame is selected, and spatial alignment of all image frames is completed through pixel-level position adjustment. Under each time index, the amplitude components of the acoustic emission signal, the ultrasonic detection signal, the vibration response signal, and the pixel grayscale components of the image data are combined into a monitoring vector. All monitoring vectors are arranged in chronological order to generate a standardized monitoring data sequence. The generation of the low-dimensional trajectory in the chaotic dynamics specifically includes: The standardized monitoring data sequence is processed by delay coordinate embedding method. The embedding dimension and time delay parameters are set, and the monitoring vector of each time point is combined with the monitoring vectors of several delayed time points in sequence to form a high-dimensional trajectory point set containing multiple time-series components. Arrange the high-dimensional trajectory point set in chronological order to construct a high-dimensional phase space reconstruction matrix; In the high-dimensional phase space reconstruction matrix, adjacency relationships are established based on the similarity measure between high-dimensional trajectory points, and adjacent trajectory points are connected to form a damage evolution time-series manifold; The damage evolution time-series manifold is subjected to dimensionality reduction processing, and high-dimensional trajectory points are mapped to low-dimensional space using a nonlinear mapping method to obtain a set of low-dimensional trajectory points; Arrange the set of low-dimensional trajectory points in time index order to generate low-dimensional trajectories of chaotic dynamics; The generation of the prediction vector specifically includes: The set of singular attractor features is input into an improved Linformer damage recognition network, which includes an input feature construction layer, a spectral domain transformation layer, a low-rank projection layer, a manifold coupled attention layer, and a multi-layer stacked modeling layer. In the input feature construction layer, the set of singular attractor features is arranged into an input feature matrix in time index order; The input feature matrix is ​​subjected to a Fourier transform through a spectral domain transform layer to obtain the spectral domain feature matrix; The spectral domain feature matrix is ​​used as the input to the low-rank projection layer. Low-rank projection processing is performed to obtain the query matrix, key matrix, and value matrix. The low-rank projection processing refers to multiplying the input of the low-rank projection layer with the corresponding low-rank projection weight matrix. In the manifold coupled attention layer, a manifold coupled attention mechanism is used to calculate similarity weights based on the query matrix and the key matrix. These similarity weights are then weighted and fused with an adjacency relation matrix constructed from a low-dimensional set of trajectory points. After normalization, a manifold coupled attention weight matrix is ​​generated. A = SoftmaxQKTr + μM; Where A represents the manifold coupling attention weight matrix, Softmax represents normalization, K represents the key matrix, Q represents the query matrix, T represents the transpose operation, r represents the scaling factor, μ represents the hyperparameter that adjusts the influence of manifold adjacency information and similarity weights, and M represents the adjacency relation matrix. Multiply the manifold-coupled attention weight matrix with the value matrix to obtain the attention weighted representation matrix. Then, perform residual connection and layer normalization on the attention weighted representation matrix and the input of the low-rank projection layer to generate the first layer feature representation. In the multi-layer stacked modeling layer, the first layer feature representation is used as the input of the low-rank projection layer. The query matrix, key matrix and value matrix are recalculated, and the manifold coupling attention mechanism is repeated to obtain a multi-layer feature representation sequence. The final layer feature representation is used as the damage morphology feature representation. After processing by a fully connected network, a prediction vector is generated. The generation of the damage recognition network under nonlinear dynamic constraints specifically includes: Train the improved Linformer damage recognition network; Establish a nonlinear dynamic equation as a reference evolution relationship; A prediction error metric is constructed using the mean squared error between the predicted vector and the true label vector. Substitute the predicted vector into the nonlinear dynamic equation and construct a dynamic consistency constraint metric; A joint loss function is established by weighting and combining the prediction error metric and the dynamic consistency constraint metric. Based on the joint loss function, gradient descent is used to update the parameters of the improved Linformer damage recognition network, iteratively optimizing the network weight matrix and bias parameters until the joint loss function converges or the training epochs reach a preset upper limit, thus generating a nonlinear dynamically constrained damage recognition network.

2. The method for monitoring the damage morphology of composite structures based on deep learning according to claim 1, characterized in that, The generation of the strange attractor feature set specifically includes: Obtain the set of low-dimensional trajectory points corresponding to the low-dimensional trajectory of chaotic dynamics, perform geometric distribution processing on the set of low-dimensional trajectory points, and generate geometric morphological features, including the average Euclidean distance between trajectory points, trajectory curvature distribution, and trajectory envelope area. Adjacency processing is performed on the low-dimensional trajectory point set to establish a topological structure and generate topological features, including node degree distribution, number of connected components, and local clustering coefficients. A time indexing process is performed on the low-dimensional trajectory point set to construct a state transition matrix and perform eigenvalue decomposition to extract stability features, including the maximum Lyapunov exponent, the spectral radius of the state transition matrix, and the convergence rate of the principal eigenvector. Geometric features, topological features, and stability features are normalized, and the normalized features are combined to generate a set of singular attractor features.

3. The method for monitoring the damage morphology of composite structures based on deep learning according to claim 1, characterized in that, The generation of the damage category and damage evolution trend specifically includes: A damage identification network constrained by nonlinear dynamics is invoked to obtain a prediction vector based on real-time multi-source monitoring data. The prediction vector includes a damage category probability value and a damage evolution intensity value. The category corresponding to the highest probability value of the damage category is taken as the damage category. The damage evolution trend is determined by the damage evolution intensity value.

4. A deep learning-based composite structure damage morphology monitoring system, executing the deep learning-based composite structure damage morphology monitoring method according to any one of claims 1 to 3, characterized in that, include: The acquisition and preprocessing module is used to acquire multi-source monitoring data of the composite structure under load and generate standardized monitoring data sequences. The temporal trajectory reconstruction module is used to reconstruct the temporal manifold of damage evolution based on standardized monitoring data sequences and generate low-dimensional trajectories of chaotic dynamics. The feature extraction module is used to extract a set of singular attractor features from low-dimensional trajectories in chaotic dynamics; An improved Linformer damage recognition network is used to receive a set of singular attractor features and generate a prediction vector; The nonlinear dynamics constraint training module is used to establish nonlinear dynamic constraints and train an improved Linformer damage recognition network in combination with a joint loss function, thereby generating a damage recognition network with nonlinear dynamic constraints. The prediction output module is used to call the damage identification network with nonlinear dynamic constraints to generate damage categories and damage evolution trends; The generation of the standardized monitoring data sequence specifically includes: Under the condition of composite structure under load, acoustic emission signals, ultrasonic detection signals, vibration response signals and image data are collected simultaneously to obtain multi-source monitoring data; Noise filtering is performed on acoustic emission signals, ultrasonic detection signals, and vibration response signals to remove environmental interference components and obtain a set of noise-filtered signals. Noise filtering is performed on the image data to reduce background interference and imaging artifacts during image acquisition, resulting in noise-filtered image data. The amplitude normalization process is performed on the noise-filtered signal set to obtain a normalized signal set; Pixel grayscale normalization is performed on the noise-filtered image data to map the pixel grayscale range to a uniform interval, resulting in normalized image data. Perform time alignment on the normalized signal set to keep it consistent in time index; Spatial registration processing is performed on the normalized image data. A reference image frame is selected, and spatial alignment of all image frames is completed through pixel-level position adjustment. Under each time index, the amplitude components of the acoustic emission signal, the ultrasonic detection signal, the vibration response signal, and the pixel grayscale components of the image data are combined into a monitoring vector. All monitoring vectors are arranged in chronological order to generate a standardized monitoring data sequence. The generation of the low-dimensional trajectory in the chaotic dynamics specifically includes: The standardized monitoring data sequence is processed by delay coordinate embedding method. The embedding dimension and time delay parameters are set, and the monitoring vector of each time point is combined with the monitoring vectors of several delayed time points in sequence to form a high-dimensional trajectory point set containing multiple time-series components. Arrange the high-dimensional trajectory point set in chronological order to construct a high-dimensional phase space reconstruction matrix; In the high-dimensional phase space reconstruction matrix, adjacency relationships are established based on the similarity measure between high-dimensional trajectory points, and adjacent trajectory points are connected to form a damage evolution time-series manifold; The damage evolution time-series manifold is subjected to dimensionality reduction processing, and high-dimensional trajectory points are mapped to low-dimensional space using a nonlinear mapping method to obtain a set of low-dimensional trajectory points; Arrange the set of low-dimensional trajectory points in time index order to generate low-dimensional trajectories of chaotic dynamics; The generation of the prediction vector specifically includes: The set of singular attractor features is input into an improved Linformer damage recognition network, which includes an input feature construction layer, a spectral domain transformation layer, a low-rank projection layer, a manifold coupled attention layer, and a multi-layer stacked modeling layer. In the input feature construction layer, the set of singular attractor features is arranged into an input feature matrix in time index order; The input feature matrix is ​​subjected to a Fourier transform through a spectral domain transform layer to obtain the spectral domain feature matrix; The spectral domain feature matrix is ​​used as the input to the low-rank projection layer. Low-rank projection processing is performed to obtain the query matrix, key matrix, and value matrix. The low-rank projection processing refers to multiplying the input of the low-rank projection layer with the corresponding low-rank projection weight matrix. In the manifold coupled attention layer, a manifold coupled attention mechanism is used to calculate similarity weights based on the query matrix and the key matrix. These similarity weights are then weighted and fused with an adjacency relation matrix constructed from a low-dimensional set of trajectory points. After normalization, a manifold coupled attention weight matrix is ​​generated. A = SoftmaxQKTr + μM; Where A represents the manifold coupling attention weight matrix, Softmax represents normalization, K represents the key matrix, Q represents the query matrix, T represents the transpose operation, r represents the scaling factor, μ represents the hyperparameter that adjusts the influence of manifold adjacency information and similarity weights, and M represents the adjacency relation matrix. Multiply the manifold-coupled attention weight matrix with the value matrix to obtain the attention weighted representation matrix. Then, perform residual connection and layer normalization on the attention weighted representation matrix and the input of the low-rank projection layer to generate the first layer feature representation. In the multi-layer stacked modeling layer, the first layer feature representation is used as the input of the low-rank projection layer. The query matrix, key matrix and value matrix are recalculated, and the manifold coupling attention mechanism is repeated to obtain a multi-layer feature representation sequence. The final layer feature representation is used as the damage morphology feature representation. After processing by a fully connected network, a prediction vector is generated. The generation of the damage recognition network under nonlinear dynamic constraints specifically includes: Train the improved Linformer damage recognition network; Establish a nonlinear dynamic equation as a reference evolution relationship; A prediction error metric is constructed using the mean squared error between the predicted vector and the true label vector. Substitute the predicted vector into the nonlinear dynamic equation and construct a dynamic consistency constraint metric; A joint loss function is established by weighting and combining the prediction error metric and the dynamic consistency constraint metric. Based on the joint loss function, gradient descent is used to update the parameters of the improved Linformer damage recognition network, iteratively optimizing the network weight matrix and bias parameters until the joint loss function converges or the training epochs reach a preset upper limit, thus generating a nonlinear dynamically constrained damage recognition network.

Images