Adaptive double-entropy-multi-scale Transform unsupervised cross-domain composite material damage prediction method

Abstract

The invention provides a self-adaptive double-entropy-multi-scale Transform unsupervised cross-domain composite material damage prediction method, which comprises the following steps: preprocessing an acoustic emission signal and extracting a double-entropy feature matrix based on the acoustic emission signal acquired by a composite material test piece under a tensile-tensile fatigue load; constructing a data set based on the double-entropy feature matrix and training a damage evolution model; and inputting the double-entropy feature matrix corresponding to an acoustic emission signal into the trained damage evolution model based on the acoustic emission signal acquired by the composite material test piece under the shear fatigue load, so as to realize real-time prediction of the shear fatigue damage. According to the method, the cross-load high-reliability prediction is realized, the feature dimension redundancy is improved, the physical interpretability is improved, the supervision marking cost is reduced, and a brand new solution is provided for engineering application.

Description

Technical Field

[0001] The present invention belongs to the field of computer material science and technology, and relates to acoustic emission damage prediction of composite materials, and specifically to an adaptive dual-entropy-multi-scale Transformer unsupervised cross-domain composite material damage prediction method. Background Art

[0002] Composite materials have been widely used in aerospace, wind power generation, marine engineering, and other fields due to their high specific strength, high specific stiffness, and excellent corrosion resistance. However, composite materials have complex internal laminate structures, diverse stress coupling patterns, and typical damage modes (such as matrix cracks, fiber fractures, delamination, and interfacial debonding) often exhibit multi-scale, progressive evolution characteristics, posing significant challenges to the health monitoring and life prediction of in-service structures.

[0003] Acoustic emission (AE) technology can capture high-frequency elastic wave signals at the moment of microcrack initiation or propagation in materials, and is considered one of the only nondestructive testing methods that can reflect the origin and evolution of damage in real time. Traditional AE evaluation methods rely primarily on empirical thresholds or statistical parameters (count rate, ring count, amplitude distribution, etc.) to determine the extent of damage. However, these methods have limited ability to distinguish multi-source and multi-mode signals within composite materials and are highly sensitive to environmental noise and changes in stress patterns, making them difficult to accurately predict in complex service scenarios.

[0004] In recent years, data-driven machine learning models (such as support vector machines, convolutional neural networks, and long short-term memory networks) have been introduced into the field of acoustic emission signal processing. By learning high-dimensional features of acoustic emission waveforms or time-frequency spectrum characteristics, automatic identification of damage states under a single loading condition has been achieved. However, a large number of comparative studies have found that the existing technologies have the following main drawbacks:

[0005] First, they rely on a single load condition, making it difficult to achieve highly reliable predictions across loads. Existing models are typically trained on single stress pattern datasets, such as "tension-tension" or "compression-compression." The learned feature distributions are tightly coupled to the corresponding loading condition. When the stress state changes (such as switching to shear or multiaxial loading), model performance degrades dramatically, lacking the cross-load robustness required for engineering applications. Second, they suffer from feature dimensionality redundancy and insufficient physical interpretability. Existing models typically directly use high-dimensional raw waveforms or dense time-frequency spectra as model inputs. This not only leads to a large number of network parameters and high training overhead, but also lacks an effective characterization of damage physics mechanisms, limiting interpretability and industrial applicability. Third, supervised labeling is costly. Complete ground-truth damage labels are difficult to obtain at engineering sites, making models that rely on fully supervised training difficult to deploy and scale under real-world complex stress conditions. In summary, the difficulty in achieving highly reliable predictions across loads, the lack of feature dimensionality redundancy and physical interpretability, and the high cost of supervised labeling limit engineering applications. Summary of the Invention

[0006] In view of the defects and shortcomings of the existing technology, the purpose of the present invention is to provide an adaptive dual-entropy-multi-scale Transformer unsupervised cross-domain composite material damage prediction method to solve the technical problems of the existing neural network model in the field of acoustic emission signal processing, which is difficult to achieve high-reliability prediction across loads, lack of feature dimension redundancy and physical interpretability, and high cost of supervised annotation.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] An adaptive dual-entropy-multiscale Transformer unsupervised cross-domain composite material damage prediction method, the method comprising:

[0009] Based on the acoustic emission signals collected from composite test pieces under tension-tension fatigue load, the acoustic emission signals are preprocessed and a double entropy feature matrix is ​​extracted. Based on the double entropy feature matrix, a data set is constructed and a damage evolution model is trained.

[0010] The damage evolution model includes a multi-scale network and a Transformer framework. The multi-scale network is used to enhance and fuse the input dual entropy feature matrix to obtain an enhanced feature matrix. The Transformer framework includes an encoder and a decoder, which are used to compress and reconstruct the enhanced feature matrix. After obtaining the output matrix, linear projection is performed to finally obtain the damage index prediction value.

[0011] Based on the acoustic emission signals collected from composite test pieces under shear fatigue load, the double entropy feature matrix corresponding to the acoustic emission signals is input into the trained damage evolution model to achieve real-time prediction of shear fatigue damage.

[0012] The present invention also includes the following technical features:

[0013] Specifically, the process of training the damage evolution model includes: using the acoustic emission signals collected under tensile-tensile fatigue load as the source domain data and preprocessing them; based on the preprocessed data, an adaptive dual-entropy analysis method is used to construct the dual-entropy feature matrix of the source domain; based on the multi-scale network-Transformer framework and an unsupervised cross-domain method, the model is trained, with the dual-entropy feature matrix of the source domain as input and the output being the tensile fatigue life percentage and damage degree; and based on the output, a damage evolution model is constructed.

[0014] Specifically, an adaptive dual-entropy analysis method is adopted to construct a dual-entropy feature matrix of the source domain, including: reconstructing the phase space of the preprocessed acoustic emission data, embedding dimensions and time lags, and calculating and obtaining the source domain permutation entropy; estimating the power spectral density of the preprocessed acoustic emission data, dividing it into multiple frequency bands, and calculating and obtaining the source domain spectral entropy; combining the permutation entropy and the spectral entropy to form a source domain dual-channel entropy vector, dynamically weighting it using the channel attention mechanism, and obtaining the source domain weighted weight vector; obtaining the source domain weighted entropy feature based on the source domain weighted weight vector and the source domain dual-channel entropy vector, and stacking the source domain weighted entropy features in chronological order to obtain the source domain dual-entropy feature matrix.

[0015] Specifically, the process of realizing real-time prediction of shear fatigue damage includes: using the acoustic emission signals collected under shear fatigue load as the target domain data and preprocessing them; based on the preprocessed data, an adaptive dual entropy analysis method is used to construct the dual entropy characteristic matrix of the target domain; prediction is performed through the damage evolution model, with the dual entropy characteristic matrix of the target domain as input, and the output is the degree of shear fatigue damage.

[0016] Specifically, an adaptive dual-entropy analysis method is adopted to construct a dual-entropy feature matrix of the target domain, including: reconstructing the phase space of the preprocessed acoustic emission data, embedding dimensions and time lags, and calculating and obtaining the permutation entropy of the target domain; estimating the power spectral density of the preprocessed acoustic emission data, dividing it into multiple frequency bands, and calculating and obtaining the spectral entropy of the target domain; combining the permutation entropy and the spectral entropy to form a target domain dual-channel entropy vector, dynamically weighting it using the channel attention mechanism, and obtaining the weighted weight vector of the target domain; obtaining the weighted entropy feature of the target domain based on the weighted weight vector of the target domain and the dual-channel entropy vector of the target domain, and stacking the weighted entropy features of the target domain in chronological order to obtain the dual-entropy feature matrix of the target domain.

[0017] Specifically, model training based on the multi-scale network-Transformer framework and the use of unsupervised cross-domain methods includes multi-scale network feature enhancement, Transformer long-range time series modeling and model training, and unsupervised cross-domain optimization of the model.

[0018] Specifically, the multi-scale network feature enhancement includes: linearly mapping the input dual entropy feature matrix to a high-dimensional space to obtain a dimensionally upgraded feature matrix; constructing an adjacency relationship matrix between features based on the dimensionally upgraded feature matrix; using the adjacency relationship matrix between features to perform residual diffusion propagation to obtain propagation features of each order; and fusing the propagation features of each order through a convolution operation to obtain an enhanced feature matrix.

[0019] Specifically, the Transformer long-range temporal series modeling and model training include: position encoding the enhanced feature matrix, and adding the position encoding to the enhanced feature matrix to obtain temporal perception features; embedding the temporal perception features into the layer as the input features of the Transformer encoder, and the input features are processed by the multi-head attention mechanism in the encoder, and then fused with the input features. The fused results are normalized at the layer and further transformed with nonlinear features by the feedforward neural network; the transformed results are fused with the input of the feedforward neural network, and then normalized at the layer again to output a temporal enhancement feature sequence; the temporal enhancement feature sequence output by the encoder is input into the decoder, and the decoder has the same structure as the encoder; the decoder outputs the reconstructed image feature vector, and finally generates the damage index prediction value at each moment through linear projection and nonlinear function; the damage index prediction value generated at each moment forms a real-time damage index prediction sequence.

[0020] Specifically, the unsupervised cross-domain optimization of the model includes: constructing a total loss function based on the damage prediction degree error, reconstruction loss and discriminator loss, and weighting the reconstruction loss and discriminator loss; taking the normalized cumulative acoustic emission energy value as the true value and calculating the damage prediction degree error based on the real-time damage index prediction sequence; calculating the reconstruction loss based on the eigenvector corresponding to each moment in the enhanced feature matrix and the reconstructed image eigenvector.

[0021] Specifically, the amplitude, peak frequency, rise time, duration, and ring count of the acoustic emission signal collected under tension-tension fatigue loading were selected as the source domain acoustic emission data. The amplitude, peak frequency, rise time, duration, and ring count of the acoustic emission signal collected under shear fatigue loading were selected as the target domain emission data.

[0022] Compared with the prior art, the present invention has the following beneficial technical effects:

[0023] This invention adopts the permutation entropy-spectral entropy lightweight attention fusion for the first time, constructs a new dual-entropy adaptive feature representation, simplifies the feature dimension redundancy and improves physical interpretability, thereby significantly weakening the sensitivity of acoustic emission signals to load patterns; the multi-scale network used in modeling captures the nonlinear coupling between entropy features, and uses the Transformer framework to accurately characterize long-range dependencies. The combination of the two realizes multi-scale graph-time series collaborative modeling and improves prediction accuracy; in addition, the invention adopts an unsupervised cross-domain migration mechanism, that is, no target domain damage label is required, and high-reliability cross-load prediction can be completed only through self-supervised reconstruction and domain adversarial alignment, providing a new solution for significantly reducing labeling costs on engineering sites. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A flow chart of an unsupervised cross-domain composite material damage prediction method based on acoustic emission adaptive dual entropy-multiscale Transformer is shown.

[0025] Figure 2 A flowchart of the adaptive dual-entropy dimensionless feature in step 2 of the embodiment is shown.

[0026] Figure 3 The flowchart of the multi-scale graph network feature enhancement process in step three of an embodiment of the present invention is shown.

[0027] Figure 4 The flowchart of the Transformer timing prediction process in step 4 of the embodiment of the present invention is shown.

[0028] Figure 5 The test site of the embodiment of the present invention and the size details of the composite material test piece are shown.

[0029] Figure 6 The results of real-time prediction of shear fatigue damage based on the trained damage evolution model are presented.

[0030] The technical solution of the present invention is further described below in conjunction with embodiments. DETAILED DESCRIPTION

[0031] In this paper, "adaptive dual entropy" refers to a signal analysis method that combines permutation entropy and spectral entropy. "Multi-scale" refers to the enhancement and fusion of feature matrices through a multi-scale network architecture. Transformer unsupervised cross-domain learning refers to the use of the Transformer model to achieve migration and model adaptation between source and target domains within an unsupervised learning framework.

[0032] It should be noted that, unless otherwise specified, the components and raw materials used in the present invention are conventional components and raw materials known in the prior art.

[0033] In accordance with the above technical solution, specific embodiments of the present invention are given below. It should be noted that the present invention is not limited to the following specific embodiments, and all equivalent changes made on the basis of the technical solution of this application fall within the protection scope of the present invention.

[0034] Example:

[0035] This embodiment provides an adaptive dual entropy-multiscale Transformer unsupervised cross-domain composite material damage prediction method, such as Figure 1 As shown, the method includes: preparing the composite material into a test piece for a tensile-tensile fatigue test, collecting and obtaining acoustic emission signals; constructing and training a damage evolution model based on the acoustic emission signals collected under tensile-tensile fatigue load; preparing the composite material into a test piece for a shear fatigue test, deploying the damage evolution model in a real-time shear fatigue test environment, collecting and obtaining acoustic emission signals; and realizing real-time prediction of shear fatigue damage through the damage evolution model based on the acoustic emission signals collected under shear fatigue load.

[0036] As a specific solution of this embodiment, Figure 5 As shown in the figure, the DS5 series full-information AE signal analyzer of Soft Island Technology is used to collect AE signals. The RS-54A sensor head of the DS5 AE signal analyzer is attached to the composite test piece with vaseline as a coupling agent. The AE acquisition equipment is adjusted and the AE threshold value is adjusted from low to high until there is no voltage fluctuation to avoid recording the ambient noise of the laboratory.

[0037] As a specific embodiment of this embodiment, the test piece was made of a two-dimensional plain-weave glass fiber-reinforced composite laminate. It consisted of 20 single-layer composite layers, each 0.2 mm thick, for a total composite laminate thickness of approximately 4 mm. The cross-section of the fiber bundles in the single-layer composite material was approximately elliptical, with a major axis of approximately 700 μm and a minor axis of approximately 90 μm.

[0038] As a specific solution of this embodiment, the tensile-tensile fatigue test uses a composite material open-hole laminate as a test piece and is performed on an MTS810±100kN electro-hydraulic servo material testing machine, referring to the ASTM D5766 standard published by the American Society for Testing and Materials; the shear fatigue test uses a V-notch specimen as a test piece and is performed on an Instron 8872±25kN electro-hydraulic servo material testing machine, referring to the ASTM D5379 standard published by the American Society for Testing and Materials.

[0039] As a specific solution of this embodiment, the source data tensile fatigue test was conducted at four stress levels, each level including two parallel test specimens; the stress levels were set to the range of 50% to 80% of the tensile strength, covering typical low, medium and high stress level conditions to ensure coverage of the entire life cycle of fatigue damage; acoustic emission monitoring was used to track the fatigue damage response at each stress level. The cumulative acoustic emission energy is the sum of the energy of all acoustic emission signals over a period of time. It can more sensitively reflect the entire process of fatigue damage in composite laminates from initiation to expansion to failure, and is an important characterization parameter for characterizing the fatigue damage evolution stage and damage severity; the cumulative acoustic emission energy is normalized to a dimensionless damage variable (DI) ranging from 0 to 1, and the number of loading cycles is normalized to a percentage of fatigue life, that is, a fatigue life percentage-damage degree diagram is constructed.

[0040] like Figures 2 to 4 As shown, the method specifically includes the following steps:

[0041] Step 1: Use the acoustic emission signal collected under tensile-tensile fatigue load as the source domain data and perform preprocessing:

[0042] The amplitude, peak frequency, rise time, duration and ring count are selected from the acoustic emission signal collected under tensile-tensile fatigue load, and the original acoustic emission waveform sequence s(t) collected is cut into a sliding window with a window length of L to form s k ∈R L ; Each window sequence is normalized to obtain a normalized sequence, as shown in the following formula 1;

[0043]

[0044] In formula 1:

[0045] Represents the normalized signal value, which represents the normalized result of the i-th event in the k-th window; that is, the value after removing the mean and scaling, which is used to enhance the robustness of the model;

[0046] s k (i) represents the i-th acoustic emission event in the k-th sliding window, i.e., the local number (from 1 to L);

[0047] μ k represents the mean of the kth sliding window;

[0048] σ k represents the standard deviation of the k-th sliding window;

[0049] The above-mentioned source domain acoustic emission data are organized into five-dimensional original feature vectors in chronological order, as shown in Formula 2:

[0050] x(t)=[A(t),F p (t),RT(t),DT(t),RC(t)] Formula 2;

[0051] In formula 2:

[0052] x(t) represents the five-dimensional original feature vector composed of the source domain acoustic emission data in time sequence;

[0053] A represents the amplitude in dB; A(t) represents the original eigenvector of the amplitude at time t;

[0054] F p Indicates the peak frequency in kHz; F p (t) represents the original eigenvector of the peak frequency at time t;

[0055] RT represents the rise time in μs; RT(t) represents the original eigenvector of the rise time at time t;

[0056] DT represents the duration in μs; DT(t) represents the original feature vector of the duration at time t;

[0057] RC is the ring count, in times; RC(t) represents the original eigenvector of the ring count at time t;

[0058] Then, the window length L = 256 events and the step size Δ = 64 events are used to slide the window, and the sequence block X is obtained. k ∈R L×5 ;

[0059] The z-score normalization is performed on the components of the five-dimensional original feature vector within each window, as shown in Equation 3 below:

[0060]

[0061] In formula 3:

[0062] represents the normalized eigenvalue of the i-th acoustic emission event in the j-th feature dimension within the k-th sliding window;

[0063] X k (i, j) represents the original eigenvalue of the i-th acoustic emission event in the j-th feature dimension within the k-th sliding window;

[0064] μ k,j Represents the mean value of the j-th dimension feature in the k-th sliding window;

[0065] σ k,j Represents the standard deviation of the j-th dimension feature in the k-th sliding window.

[0066] Step 2: Based on the preprocessed data, the adaptive dual entropy analysis method is used to construct the dual entropy feature matrix of the source domain:

[0067] Step 2.1, calculate the permutation entropy: reconstruct the normalized eigenvalue sequence obtained in step 1 into phase space, select embedding dimension d = 4 and time lag τ = 1, and obtain the window permutation entropy as shown in Equation 4 below:

[0068]

[0069] In formula 4:

[0070] PE k Represents the normalized permutation entropy, which is used to quantify the time-domain complexity of the sequence within the window, and its value range is 0 to 1.

[0071] d! represents the total number of permutation patterns; r

[0072] r represents the index of the permutation pattern, indicating the rth pattern (of the same species);

[0073] p r Indicates the probability of occurrence of the rth arrangement pattern in the current window;

[0074] Step 2.2, calculate the spectral entropy: Use the Welch method to estimate the power spectral density of the normalized sequence obtained in step 1, divide it into N frequency bands, and calculate the spectral entropy using the following formula 5:

[0075]

[0076] In formula 5:

[0077] SpE k represents the spectral entropy of the kth window;

[0078] N represents the total number of frequency bands; in this embodiment, N=64;

[0079] q i Indicates the relative power ratio of the i-th frequency band. This parameter is used to characterize the frequency-domain energy dispersion within the window and its value range is 0 to 1.

[0080] In this embodiment, q i The following formula 6 is used to calculate and obtain:

[0081] q b =P k (f b ) / ∑ j P k (f b ) Equation 6;

[0082] In formula 6:

[0083] b represents the frequency band number index, indicating the bth frequency band (64 in total);

[0084] f b Indicates the center frequency corresponding to the bth frequency band;

[0085] P k (f b ) represents the frequency f b The power spectral density value of the frequency band;

[0086] Step 2.3, channel attention weighting: construct a dual-channel entropy vector e k =[PE k ,SpE k ], the weights are obtained through lightweight full connection-Softmax, and the weighted weight vector is expressed as follows:

[0087] w k =Softmax(ReLU(e k ×W1)×W2) Formula 7;

[0088] In formula 7:

[0089] w k Represents the weighted weight vector, satisfying that the sum of the two elements is 1;

[0090] e k Represents the dual-channel entropy vector, e k =[PE k ,SpE k ];

[0091] Softmax() represents the normalized exponential function;

[0092] ReLU stands for ReLU (Rectified Linear Unit) function;

[0093] W1∈R 2×h , W2∈R h×2 is a trainable parameter of the model, where h is the dimension of the attention hidden layer, which is usually a small value, for example, 4;

[0094] Step 2.4, obtain weighted entropy features: obtain weighted entropy features based on the weighted weight vector and the dual-channel entropy vector, as shown in the following formula 8:

[0095] z k =w k ⊙e k ∈R 2 Formula 8;

[0096] In formula 8:

[0097] zk represents the weighted entropy feature;

[0098] ⊙ represents element-wise multiplication;

[0099] The weighted entropy features are stacked in time order to obtain the double entropy feature matrix Z∈R T×2 , T represents the sequence length; subsequently, the model is trained based on the multi-scale network-Transformer framework and an unsupervised cross-domain method, taking the double entropy feature matrix Z of the source domain as input and the output as the tensile fatigue life percentage and damage degree; a damage evolution model is constructed based on the output.

[0100] Step 3: Multi-scale network feature enhancement:

[0101] Step 3.1, feature dimension upgrade: linearly map the input dual entropy feature matrix to the high-dimensional space to obtain the dual entropy feature matrix after dimension upgrade, as shown in the following formula 8:

[0102]

[0103] In formula 8:

[0104] H (0) Represents the double entropy feature matrix after dimensionality increase;

[0105] Z represents the double entropy feature matrix;

[0106] W0 represents the weight matrix, d0 represents the feature dimension, and d0=32 is often used to improve the feature expression capability;

[0107] b0 represents the bias vector.

[0108] Step 3.2, Adjacency Matrix Adaptive Learning: Construct the adjacency relationship matrix between features based on the dimension-enhanced feature matrix, as shown in Equation 8 below:

[0109] A=Softmax(H (0) ×U1×(H (0) ×U2) T ) Equation 8;

[0110] In formula 8:

[0111] U1 and U2 represent low-rank decomposition matrices, r represents the decomposition rank; U1 and U2 are two independent trainable projection matrices that transform the node feature H (0) Projected to a low-rank subspace; U1 is used to generate a 'query vector', which represents the direction that the current node actively pays attention to in the graph structure; U2 is used to generate a 'key vector', which represents the characteristic attributes of other nodes that may be paid attention to.

[0112] A represents the adjacency matrix between features, reflecting the dynamic connection relationship between features.

[0113] Step 3.3, second-order diffusion propagation: Use the adjacency relationship matrix between features to perform residual diffusion propagation to obtain the propagation features of each order, as shown in the following formula 9:

[0114] D (1) =α×A×H (0) +(1-α)×H (0) , D (2) =α×A×D (1) +(1-α)×D (1) Formula 9;

[0115] In formula 9:

[0116] D (1) Represents the diffusion characteristics of the first-order residual graph;

[0117] D (2) Represents the diffusion characteristics of the second-order residual graph;

[0118] α represents the control feature propagation strength, α∈[0,1], and the diffusion order is fixed to 2 to balance the computational complexity and feature capture capability.

[0119] Step 3.4, feature fusion output: The convolution operation is used to fuse the propagation features of each order (i.e., the first-order residual map diffusion feature and the second-order residual map diffusion feature) to obtain the enhanced feature matrix, as shown in the following formula 10:

[0120] H G =GELU([H (0) ||D (1) ||D (2) ]W f ) Equation 10;

[0121] In formula 10:

[0122] H G represents the enhanced feature matrix;

[0123] GELU() represents the GELU (Gaussian Error Linear Unit) activation function;

[0124] W f represents the mapping weight matrix, Where d represents the output dimension after GELU mapping, that is, the input dimension of Transformer. In this embodiment, d = 64;

[0125] || indicates feature splicing;

[0126] Step 4: Transformer long-term temporal modeling and model training:

[0127] Step 4.1, Position Encoding: Since the Transformer model lacks inherent perception of sequence order, it is necessary to first add position encoding to the enhanced feature matrix to clarify the position relationship of each feature vector in the sequence and provide basic information for the Transformer's long-term dependency modeling. In this embodiment, the enhanced feature matrix H output in step 3 is G ∈R T×d , define the position encoding matrix P∈R T×d , encoded using the sine-cosine function, as shown in Equation 11:

[0128]

[0129] In formula 11:

[0130] P t,2i Represents the position encoding matrix corresponding to the c-th dimension in the feature vector at the t-th position in the sequence;

[0131] t represents the tth position in the sequence, i is the cth dimension in the feature vector, and d is the feature dimension.

[0132] The position code is added to the enhanced feature matrix to obtain the temporal perception feature, as shown in Equation 12:

[0133] H T =H G +P formula 12;

[0134] In formula 12:

[0135] H T Represents temporal perception features;

[0136] Step 4.2, implementation of Transformer encoder: Transformer encoder is used to extract long-term dependencies and global temporal information in the sequence:

[0137] In this embodiment, the encoder-decoder adopts L e = 3-layer encoder and L d = 3-layer decoder. The temporal perception feature H T After the embedding layer, it is used as the input feature of the Transformer encoder. The input feature is processed using the multi-head attention mechanism in the encoder, as shown in the following formula 12:

[0138] MHA(Q,K,V)=Concat(head1,...,head m )×W O Formula 12;

[0139] In formula 12:

[0140] MHA(Q,K,V) represents the weighted output feature of the multi-head attention mechanism after calculating the input vector (Q,K,V), Q represents the input feature H T Query vector, K represents the input feature H T Key vector, V represents the input feature H T Value vector, H T Represents temporal perception features;

[0141] Concat() represents the concatenation operation, which concatenates the outputs of multiple attentions in the feature dimension;

[0142] The calculation method of each attention head is shown in the following formula 13:

[0143]

[0144] In formula 13:

[0145] head n represents the output matrix of the nth attention sub-head;

[0146] Represents the input feature H of the nth attention head T The projection matrix of the query vector, Represents the input feature H of the nth attention head T The projection matrix of the key vector, Represents the input feature H of the nth attention head T The projection matrix of the value vector,

[0147] d k represents the key vector dimension, d k =d / m, where m is the total number of attention heads. In this embodiment, m=8;

[0148] W O represents the output mapping matrix;

[0149] After the multi-head self-attention mechanism outputs features, it is fused with the input features. The fused results are normalized at the layer level and then further transformed into nonlinear features by the feedforward neural network. The nonlinear feature transformation is shown in Equation 14 below:

[0150] FFN(X)=max(0,XW1+b1)W2+b2 Equation 14;

[0151] In formula 14:

[0152] X represents the input feature matrix of the original FFN;

[0153] FFN(X) represents the output feature matrix, which is used for residual connection;

[0154] W1 represents the first layer weight, W2 represents the second layer weight, d f is the hidden layer dimension of FFN, usually set to 4d.

[0155] Each sublayer (MHA and FFN) in the encoder uses residual connection and layer normalization for stable training. After the result of feedforward neural network transformation is fused with the input of the feedforward neural network, it is normalized again and the output feature sequence H after time series enhancement is obtained. E , used as decoder input for the next step.

[0156] As a specific solution of this embodiment, the layer normalization process is shown in the following formula 15:

[0157]

[0158] In formula 15:

[0159] LayerNorm represents the normalized output matrix;

[0160] Sublayer(X) represents the sublayer output, that is, the submodule with X as input;

[0161] Step 4.3, Transformer decoder damage index prediction value:

[0162] The feature sequence H after the encoder output is enhanced E Input to the decoder, the decoder also uses a multi-head self-attention mechanism and a feedforward network (the same structure as the encoder), and the decoder outputs the reconstructed image feature vector H G The decoder uses a mask to ensure that the damage prediction at time t depends only on the current and previous information to avoid future data leakage; the reconstructed image feature vector H G Finally, the damage index prediction value at each moment is generated through linear projection and nonlinear function, as shown in the following formula 16:

[0163]

[0164] In formula 16:

[0165] represents the real-time damage prediction value at time t;

[0166] σ() represents the Sigmoid function;

[0167] Represents the reconstructed graph feature vector H G The feature vector output by the decoder at time t;

[0168] W d 、b d is the trainable parameter of the decoder, where W d Represents the regression weight matrix, W d ∈R d×1 ; b d Represents the bias term, which limits the damage index prediction value to the range of [0,1].

[0169] The damage index prediction value generated at each moment forms a real-time damage index prediction sequence Used for subsequent real-time damage tracking and life estimation.

[0170] Step 5: Unsupervised cross-domain optimization of the model:

[0171] The domain discriminator is used to align the feature distributions of the source and target domains and reduce the differences across load conditions. The total loss function consists of the damage prediction error, reconstruction loss, and discriminator loss, and the reconstruction loss and discriminator loss are weighted. The total loss function is shown in Equation 17:

[0172]

[0173] In formula 17:

[0174] Represents the total loss, that is, the multi-objective loss function;

[0175] λ and γ (Greek letter gamma) represent the domain adversarial loss weight coefficients, usually ranging from 0.1 to 1;

[0176] Indicates the damage prediction error; when calculating the damage prediction error based on the real-time damage index prediction sequence, the normalized cumulative acoustic emission energy value DI t is the true value, minimizing the error in damage prediction, as shown in Equation 18:

[0177]

[0178] represents the reconstruction loss, which measures the mean square error between the original features and the reconstructed features; as shown in Equation 19:

[0179]

[0180] In formula 19:

[0181] Denotes the enhanced feature matrix HG The feature vector of the t-th time point in ;

[0182] Represents the reconstructed graph feature vector H G The feature vector of the t-th time point in ;

[0183] Represents the discriminator loss, which is specifically characterized by domain adversarial loss.

[0184] In this example, step five leverages the Transformer's powerful ability to capture long-range dependencies, significantly improving damage prediction accuracy. An unsupervised cross-domain optimization strategy enables the model to effectively generalize across loads and operating conditions without requiring target domain damage labels. This method has broad applicability in industrial settings and can be directly deployed for real-time composite structural health monitoring.

[0185] Step 6: Deploy the trained damage evolution model in the real-time environment of shear fatigue test, and select the amplitude, peak frequency, rise time, duration and ring count of the acoustic emission signal collected under shear fatigue load as the target domain data and preprocess it. The preprocessing process is the same as step 1 above. Based on the preprocessed data, the adaptive dual entropy analysis method is used to construct the dual entropy feature matrix of the target domain. The process is the same as step 2 above. The damage evolution model is used to predict the double entropy characteristic matrix of the target domain. As input, the output is the degree of shear fatigue damage When the shear fatigue damage level of a certain output reaches the preset threshold of 0.9, a failure warning is triggered. The system also records the number of trigger cycles and estimates the remaining life span, providing a basis for structural health management decisions.

[0186] As a specific solution of this embodiment, in order to further improve the credibility of the cross-load damage prediction results, a Monte Carlo Dropout layer is embedded in the Transformer decoding end to quantitatively evaluate the uncertainty of the prediction output.

[0187] like Figure 6 As shown in the figure, the present invention can achieve high-precision real-time tracking of the damage evolution of the shear fatigue target domain when only the tensile fatigue source domain data is used for training; the predicted curve is highly consistent with the normalized damage index of the shear measured cumulative acoustic emission energy, verifying the excellent generalization performance of the proposed method under complex stress states.

Claims

1. An adaptive dual entropy-multiscale Transformer unsupervised cross-domain composite material damage prediction method, characterized by: The method includes: Based on the acoustic emission signals collected from composite test pieces under tension-tension fatigue load, the acoustic emission signals are preprocessed and a double entropy feature matrix is ​​extracted. Based on the double entropy feature matrix, a data set is constructed and a damage evolution model is trained. The damage evolution model includes a multi-scale network and a Transformer framework; wherein the multi-scale network is used to enhance and fuse the input dual entropy feature matrix to obtain an enhanced feature matrix; the Transformer framework includes an encoder and a decoder, which are used to compress and reconstruct the enhanced feature matrix, obtain an output matrix for linear projection, and ultimately obtain a damage index prediction value; Based on the acoustic emission signals collected from composite test pieces under shear fatigue load, the double entropy feature matrix corresponding to the acoustic emission signals is input into the trained damage evolution model to achieve real-time prediction of shear fatigue damage.

2. The adaptive dual entropy-multiscale Transformer unsupervised cross-domain composite material damage prediction method according to claim 1 is characterized in that: The specific process of training the damage evolution model includes: The acoustic emission signals collected under tensile-tensile fatigue loading are used as source domain data and preprocessed; Based on the preprocessed data, the adaptive dual entropy analysis method is used to construct the dual entropy feature matrix of the source domain; Based on the multi-scale network-Transformer framework and an unsupervised cross-domain method, the model is trained. The double entropy feature matrix of the source domain is used as input, and the output is the tensile fatigue life percentage and damage degree. A damage evolution model is constructed based on the output.

3. The adaptive dual entropy-multiscale Transformer unsupervised cross-domain composite material damage prediction method according to claim 1 is characterized in that: The specific process of achieving real-time prediction of shear fatigue damage includes: The acoustic emission signals collected under shear fatigue load are used as target domain data and preprocessed; Based on the preprocessed data, the adaptive dual entropy analysis method is used to construct the dual entropy feature matrix of the target domain; The prediction is performed through the damage evolution model, which takes the double entropy characteristic matrix of the target domain as input and outputs the shear fatigue damage degree.

4. The adaptive dual entropy-multiscale Transformer unsupervised cross-domain composite material damage prediction method according to claim 2 or 3, characterized in that: The adaptive dual entropy analysis method includes: The pre-processed acoustic emission data are reconstructed in phase space, the dimensions and time delays are embedded, and the permutation entropy is calculated and obtained; The power spectral density of the pre-processed acoustic emission data is estimated, divided into multiple frequency bands, and the spectral entropy is calculated and obtained; The permutation entropy and spectral entropy are combined into a dual-channel entropy vector, which is dynamically weighted using the channel attention mechanism to obtain a weighted weight vector. The weighted entropy features are obtained based on the weighted weight vector and the dual-channel entropy vector, and the weighted entropy features are stacked in chronological order to obtain a dual-entropy feature matrix.

5. The adaptive dual entropy-multiscale Transformer unsupervised cross-domain composite material damage prediction method according to claim 2 is characterized in that: Model training based on the multi-scale network-Transformer framework and using unsupervised cross-domain methods includes multi-scale network feature enhancement, Transformer long-range temporal series modeling and model training, and unsupervised cross-domain optimization of the model; The multi-scale network feature enhancement specifically includes: linearly mapping the input dual entropy feature matrix to a high-dimensional space to obtain a dimensionally upgraded feature matrix; constructing an adjacency relationship matrix between features based on the dimensionally upgraded feature matrix; using the adjacency relationship matrix between features to perform residual diffusion propagation to obtain propagation features of each order; and fusing the propagation features of each order through a convolution operation to obtain an enhanced feature matrix.

6. The adaptive dual entropy-multiscale Transformer unsupervised cross-domain composite material damage prediction method according to claim 5, characterized in that: The Transformer long-range temporal series modeling and model training include: Perform position encoding on the enhanced feature matrix and add the position encoding to the enhanced feature matrix to obtain temporal perception features; The time-series-aware features are embedded in the layer and used as the input features of the Transformer encoder. The input features are processed using the multi-head attention mechanism within the encoder and fused with the input features. The fused results are normalized at the layer and then further transformed by the feedforward neural network for nonlinear features. The transformed results are fused with the input of the feedforward neural network and normalized again at the layer to output a time-enhanced feature sequence. The time-enhanced feature sequence output by the encoder is input into the decoder, which has the same structure as the encoder. The decoder outputs the reconstructed image feature vector, and finally generates the damage index prediction value at each moment through linear projection and nonlinear function. The generated damage index prediction value at each moment forms a real-time damage index prediction sequence.

7. The adaptive dual entropy-multiscale Transformer unsupervised cross-domain composite material damage prediction method according to claim 6, characterized in that: The unsupervised cross-domain optimization of the described model includes: A total loss function is constructed based on the damage prediction error, reconstruction loss, and discriminator loss, and weights are assigned to the reconstruction loss and discriminator loss. The normalized cumulative acoustic emission energy value is taken as the true value, and the damage prediction error is calculated based on the real-time damage index prediction sequence; The reconstruction loss is calculated based on the eigenvector corresponding to each moment in the enhanced feature matrix and the reconstructed image eigenvector.

8. The adaptive dual entropy-multiscale Transformer unsupervised cross-domain composite material damage prediction method according to claim 2, characterized in that: From the acoustic emission signals collected under tension-tension fatigue loading, the amplitude, peak frequency, rise time, duration, and ring count were selected as the source domain acoustic emission data.

9. The adaptive dual entropy-multiscale Transformer unsupervised cross-domain composite material damage prediction method according to claim 3, characterized in that: From the acoustic emission signals collected under shear fatigue loading, the amplitude, peak frequency, rise time, duration and ring count are selected as the target domain emission data.

Images