Structure cross-domain damage identification method based on domain adaptation model

Abstract

The invention relates to the technical field of structural health monitoring, in particular to a structural cross-domain damage identification method based on a domain adaptation model, which comprises a signal acquisition stage, a source domain data set construction stage and a cross-domain damage identification model training stage, and the trained cross-domain damage identification model is used for damage detection of a target domain. Through a domain adaptive regression model, a guided wave sensing technology and transfer learning are innovatively combined, the problem of cross-domain feature offset is effectively solved, the data distribution difference between a source domain and a target domain is reduced, good generalization performance is still achieved in the face of cross-domain damage data, it is guaranteed that high precision is kept in damage identification of different monitoring areas, and the method is suitable for popularization and application. And the limitation of a traditional deep learning model in cross-scene application is broken through.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of structural health monitoring, and particularly relates to a structural cross-domain damage identification method based on a domain adaptation model. BACKGROUND

[0002] In recent years, structural health monitoring plays an increasingly important role in civil engineering, aerospace, mechanical manufacturing and other fields. Its core goal is to identify potential damage or degradation by monitoring the dynamic response of the structure in real time or periodically, so as to ensure the safety and durability of the structure. However, in actual engineering, structures often face complex and changing environmental and load conditions, resulting in significant cross-domain characteristics of monitoring data, that is, there is a distribution difference between training data (source domain) and actual monitoring data (target domain).

[0003] This difference may be due to factors such as environmental temperature changes, material aging, sensor layout differences or working condition changes, which greatly reduces the performance of traditional single-domain-based damage identification methods in actual applications. Existing structural health monitoring methods based on machine learning (such as support vector machines, neural networks, etc.) have achieved certain results in structural damage identification, but most of them assume that training data and test data are independent and identically distributed, which still cannot adapt to cross-domain scenarios.

[0004] It should be noted that the information disclosed in this BACKGROUND section is only intended to increase the understanding of the overall background of the present application, and should not be considered as acknowledging or implying in any form that this information constitutes prior art known to those of ordinary skill in the art. SUMMARY

[0005] To solve the technical problem that the traditional damage identification method is difficult to adapt to the cross-domain scenario, the present application provides a structural cross-domain damage identification method based on a domain adaptation model, comprising the following steps: Signal acquisition: planning a source domain on the surface of the measured structure, and the source domain at least includes a first region and a second region, for the first region and the second region, respectively collecting two types of guided wave signals: one is the first guided wave signal under the condition that the structure of the region is intact, and the other is the second guided wave signal under the condition that the structure of the region is damaged; Source domain dataset construction: preprocessing the first guided wave signal and the second guided wave signal of the first region to extract the scattering signal of the first region; preprocessing the first guided wave signal and the second guided wave signal of the second region to extract the scattering signal of the second region; labeling each scattering signal according to different damage positions and regions to form a complete and standardized source domain damage dataset; Cross-domain damage identification model training: import the source domain damage data set into the cross-domain damage identification model for training, the cross-domain damage identification model includes a feature extraction module and a domain adaptive regression module, the feature extraction module extracts features from the source domain damage data set to obtain a source domain damage feature vector; input the source domain damage feature vector into the domain adaptive regression module to align the damage feature vector distribution between different regions and optimize the cross-region adaptation capability; Target domain damage identification: collect the first guided wave signal and the second guided wave signal of the target domain, preprocess and extract the scattering signals from the two types of signals, and construct a target domain damage data set; import the target domain damage data set into the trained cross-domain damage identification model, and output the damage location coordinates of the target domain from the model.

[0006] Further, the extraction of the scattering signal includes the following steps: Remove noise interference in the first guided wave signal and the second guided wave signal in the region by digital filtering to obtain a filtered first guided wave signal and a filtered second guided wave signal; Normalize the filtered first guided wave signal and the filtered second guided wave signal to map the signal amplitude to the [-1, 1] interval, and obtain the processed first guided wave signal and the processed second guided wave signal; Take the processed first guided wave signal as a baseline and combine the processed second guided wave signal for signal difference processing to extract the scattering signal containing damage information.

[0007] Further, the damage information at least includes amplitude change information, phase shift information and time of flight information.

[0008] Further, the feature extraction module is composed of at least 4 cascaded convolution-pooling blocks, each block including a one-dimensional convolution layer and a maximum pooling layer, wherein the convolution kernel size is 1xK (K≥2) and the ReLU activation function is used, and the pooling kernel size is ≥2x2; Then the feature extraction module extracts features from the source domain damage data set, which includes the following steps: Local feature extraction: use multiple one-dimensional convolution kernels to scan and linearly transform the scattering signals in the source domain damage data set to generate a convolution feature map; apply the ReLU activation function to the convolution feature map to obtain a primary feature map; Feature fusion: stack a one-dimensional convolution layer on the basis of the primary feature map, use a deep convolution kernel to perform deep convolution processing on the primary feature map, and perform pooling operation through a maximum pooling layer and apply a ReLU activation function to output an intermediate feature map; Global feature encoding: establish global dependency relationships according to the intermediate feature map, learn to integrate all intermediate feature maps into a global representation strongly related to the physical nature of the damage, and output a global feature vector; Feature space mapping: the global feature vector is flattened and input into a fully connected layer, which learns to linearly combine all elements of the global feature to output a source domain damage feature vector.

[0009] Further, the source domain damage feature vector is input into the domain adaptive regression module to align the distribution of damage feature vectors between different regions, including the following steps: Extracting a batch feature matrix of the first region and a batch feature matrix of the second region from the source domain damage feature vector, respectively, and performing singular value decomposition to obtain an orthogonal basis matrix of the first region and an orthogonal basis matrix of the second region , and through the orthogonal basis mismatch penalty constraint the order correspondence between the orthogonal basis matrices of different domains; Based on the orthogonal basis matrix of the first region and the orthogonal basis matrix of the second region , calculate the principal angle of the first region subspace and the principal angle of the second region subspace ; According to the principal angle of the first region subspace and the principal angle of the second region subspace , construct the of the first region subspace and the of the second region subspace, respectively; Finally, by minimizing the and between the first region and the second region, a deep feature representation with high transferability is learned, expressed as

[0010] wherein, denotes the feature extractor, denotes the regressor, is the supervised loss of the source domain, is the loss, is the regularization loss, , is a hyperparameter for balancing the losses.

[0011] Further, the orthogonal basis matrix of the first region and the orthogonal basis matrix of the second region are expressed as:

[0012] wherein, and These are the feature matrices for the first and second regions, respectively. and These are the singular values ​​of the first and second regions, respectively. and These are the singular vector matrices for the first and second regions, respectively.

[0013] Furthermore, orthogonal basis mismatch penalty Expressed as:

[0014] in, This is a mismatch penalty term for orthogonal bases. , These are the basis vector weight matrices extracted from the subspaces of the first and second regions, respectively.

[0015] Furthermore, the principal angle of the subspace of the first region The principal angle of the subspace of the second region They are expressed as follows:

[0016]

[0017] in, and These are the orthonormal bases in the first and second region subspaces of dimension b, respectively.

[0018] Furthermore, different subspaces Expressed as:

[0019] in, This represents the relative subspace distance between the first and second subspace regions. The first subspace and the second subspace are the first subspace. From a personal perspective.

[0020] Furthermore, the supervised loss of the source domain is expressed as:

[0021] but loss Expressed as:

[0022] loss Expressed as:

[0023] wherein, represents sampling a batch of b sample instances from the source domain, is a sample input signal, is a sample corresponding label, is a batch distribution of source domain samples, is a batch distribution of target domain samples, is a square loss function, is a mathematical expectation operator, is a principal angle matrix between the first regional subspace and the second regional subspace.

[0024] Based on the above, the structural cross-domain damage identification method based on the domain adaptation model provided by the application, compared with the prior art, effectively solves the cross-domain feature deviation problem by the domain adaptation regression model, innovatively combines guided wave sensing technology and transfer learning, reduces the data distribution difference between the source domain and the target domain, still has good generalization performance when facing cross-regional damage data, guarantees to maintain high precision in damage identification in different monitoring regions, and breaks through the limitation of traditional deep learning models in cross-scene application. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings described in the following description are some embodiments of the application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor. In the following description, the positional relationship described in the drawings is based on the direction of the components shown in the drawings as the reference.

[0026] Figure 1 A flowchart of a structural cross-domain damage identification method based on a domain adaptation model provided by an embodiment of the application; Figure 2 A damage layout diagram of a first region and a second region provided by an embodiment of the application; Figure 3 An architecture diagram of a feature extraction module provided by an embodiment of the application; Figure 4 An architecture diagram of a domain self-adaptive regression module provided by an embodiment of the application. DETAILED DESCRIPTION

[0027] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0028] In addition, the terms "first", "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, unless otherwise specified, the meaning of "multiple" is two or more. In addition, the term "comprising" and any variations thereof mean "at least including".

[0029] Please refer to Figure 1 , Figure 1 A flowchart of a structural cross-domain damage identification method based on a domain adaptation model is provided for an embodiment of the present application.

[0030] Embodiment one To solve the technical problem that the traditional damage identification method is difficult to adapt to the cross-domain scene, or to achieve at least one of the advantages or other advantages, an embodiment of the present application provides a structural cross-domain damage identification method based on a domain adaptation model, as shown in the figure. The structural cross-domain damage identification method based on the domain adaptation model comprises the following steps: Signal acquisition: planning a source domain on the surface of the measured structure, and the source domain at least includes a first region and a second region. For the first region and the second region, two types of guided wave signals are collected respectively: one is the first guided wave signal under the condition that the structure of the region is intact, and the other is the second guided wave signal under the condition that the structure of the region is damaged.

[0031] Source domain data set construction: pre-processing the first guided wave signal and the second guided wave signal of the first region to extract the scattering signal of the first region; pre-processing the first guided wave signal and the second guided wave signal of the second region to extract the scattering signal of the second region; and labeling each scattering signal according to different damage positions and regions to form a complete and standardized source domain damage data set.

[0032] Cross-domain damage identification model training: the source domain damage data set is imported into the cross-domain damage identification model for training, the cross-domain damage identification model includes a feature extraction module and a domain self-adaptive regression module, the feature extraction module extracts features from the source domain damage data set to obtain a source domain damage feature vector; the source domain damage feature vector is input into the domain self-adaptive regression module to align the damage feature vector distribution between different regions and optimize the cross-region adaptation capability.

[0033] Target domain damage identification: collect the first guided wave signal and the second guided wave signal of the target domain, and sequentially preprocess and extract the scattered signals of the two types of signals to construct a target domain damage data set; the target domain damage data set is imported into the trained cross-domain damage identification model, and the damage position coordinates of the target domain are output by the model.

[0034] In this embodiment, Lamb wave (Lamb wave) is preferably used as the ultrasonic guided wave signal. Lamb wave contains rich information during propagation and has no significant dissipation, so it can carry damage information and propagate a long distance, which is convenient for large-area specimen detection. Of course, SH wave, surface wave, axisymmetric guided wave propagating in tubular structure, etc. can also be used, which are all within the protection scope of the present application.

[0035] In the signal collection stage, first, a boundary is planned on the surface of the measured structure to divide the first region and the second region clearly as the source domain. As shown in Figure 2 different sizes of damage are designed in the first region and the second region, and the damage positions in the first region are different from those in the second region. The “☆” in the figure represents different damage positions, and the “○” is a PZT sensor.

[0036] The PZT sensors are uniformly arranged at the four corners of the first region and the second region. For the first region and the second region, two types of guided wave signals are collected: one is the first guided wave signal of the intact structure in the region, and the other is the second guided wave signal of the damaged structure in the region.

[0037] After the signal collection is completed, the source domain data set construction stage is entered. The first guided wave signal and the second guided wave signal of the first region are preprocessed to extract the scattered signal of the first region; the first guided wave signal and the second guided wave signal of the second region are preprocessed to extract the scattered signal of the second region; Specifically, the extraction of the scattered signal includes the following steps: The noise interference in the first guided wave signal and the second guided wave signal of the region is removed by digital filtering to obtain the filtered first guided wave signal and the filtered second guided wave signal; The normalized first guided wave signal and the normalized second guided wave signal are obtained by normalizing the filtered first guided wave signal and the filtered second guided wave signal, so as to eliminate the dimensional difference between the data characteristics, and the processed first guided wave signal and the processed second guided wave signal are obtained, which lays a foundation for subsequent analysis and model training. The processed first guided wave signal is taken as a baseline, and the processed second guided wave signal is combined to perform signal difference processing, and a scattering signal containing damage information is extracted. Preferably, the damage information at least includes amplitude change information, phase offset information and time of flight information, which can effectively represent the damage state of the measured structure.

[0038] According to different damage positions and regions, each scattering signal is artificially labeled to form a complete and standardized source domain damage data set, which provides reliable data support for subsequent structure health monitoring model training and verification based on domain adaptive regression.

[0039] After obtaining the source domain damage data set, the application proposes a cross-domain damage identification model based on domain adaptive regression, which adopts a double-module cascade architecture and includes a feature extraction module and a domain adaptive regression module.

[0040] The feature extraction module is used to extract a damage feature vector, and is composed of at least 4 cascaded convolution-pooling blocks, each block including a one-dimensional convolution layer and a maximum pooling layer, wherein the convolution kernel size is 1xK (K>=2) and the ReLU activation function is adopted, and the pooling kernel size is >=2x2, so as to realize deep feature learning from guided wave signals to high-level abstract features.

[0041] The domain adaptive regression module is used to realize cross-domain feature distribution alignment, and the core idea is to realize transfer learning between different domains through feature space distribution alignment. Therefore, a domain adaptive regression module based on representation subspace distance (RSD) is constructed.

[0042] From the geometric perspective, the module aligns the feature subspaces of different domains based on the Grassmann manifold theory, avoiding the feature norm disturbance problem caused by directly using instance representation for feature alignment in traditional methods, so as to improve the transfer ability and cross-domain regression performance of the model while maintaining the stability of the feature structure. This collaborative architecture design of "feature extraction-domain adaptive regression" breaks through the limitations of traditional deep learning models in cross-scene applications, and significantly improves the accuracy and generalization ability of structure damage identification.

[0043] In specific implementation, as Figure 3As shown, the source domain damage data set is imported into the cross-domain damage identification model, and the source domain damage data set is subjected to feature extraction by the feature extraction module. The feature extraction module includes the following steps: Local feature extraction: the scattering signals in the source domain damage data set are scanned and linearly transformed using multiple one-dimensional convolution kernels to generate a convolution feature map; a ReLU activation function is applied to the convolution feature map to obtain a primary feature map; Feature fusion: stacking a one-dimensional convolution layer on the basis of the primary feature map, performing deep convolution processing on the primary feature map using a deep convolution kernel, and performing a pooling operation through a maximum pooling layer and applying a ReLU activation function to output an intermediate feature map; Global feature encoding: establishing a global dependency relationship according to the intermediate feature map, learning to integrate all intermediate feature maps into a global representation strongly related to the physical nature of the damage, and outputting a global feature vector; Feature space mapping: flattening the global feature vector and inputting it into a fully connected layer to learn to linearly combine all elements of the global feature, and outputting a source domain damage feature vector.

[0044] In the local feature extraction stage, each convolution kernel acts as a digital filter to capture micro fluctuations and transient features in the scattering signal. Meanwhile, the essence of convolution operation is linear transformation, and the generated feature map is a similarity map of the input data and the convolution kernel pattern. The area with high numerical value represents a high degree of match with the convolution kernel pattern.

[0045] The ReLU activation function is applied to the convolution feature map to output a one-dimensional feature sequence, and the primary feature map has a shape of [batch size, number of convolution kernels, length after convolution]. Each feature sequence represents the output of the scattering signal after a specific filter, which can be visualized as an intensity distribution map of a certain local fluctuation pattern (such as a specific shape of wave peak, trough, or oscillation) in the scattering signal.

[0046] The introduction of the ReLU activation function can enhance the expression ability of non-linear features. The ReLU activation function is expressed as:

[0047] The ReLU activation function sets all negative values in the feature map to zero and retains positive values. It is equivalent to a "switch" that only activates features (positive values) that exceed a certain threshold and suppresses unimportant features (negative values). This allows the network to learn and express more complex non-linear feature relationships. Without the ReLU activation function, no matter how many layers of convolution are stacked, the entire network is ultimately equivalent to a linear transformation, with limited expression ability and unable to fit complex data distributions.

[0048] In the feature fusion stage, one-dimensional convolutional layers are continued to be stacked on the basis of the primary feature map. Multi-layer convolutional operations can gradually expand the receptive field, meaning that longer time segments in the original signal can be "seen". At the same time, deep convolutional kernels fuse multiple primary local features of the previous layer to form more structured intermediate time domain features. For example: the typical morphology of damage scattering wave packets.

[0049] The pooling operation can make the network insensitive to the slight shift of wave packets on the time axis. As long as the main energy (maximum value) of the wave packet is captured, the slight change in the exact arrival time will not affect feature extraction. This is crucial for overcoming timing errors in signal acquisition, while effectively reducing sequence length and computational load. Ultimately, a one-dimensional feature sequence with shorter length but more channels is output to form an intermediate feature map. These features represent wave packets or signal segments combined from basic waves, which are more discriminative.

[0050] In the global feature encoding stage, through the previous convolution and pooling, the receptive field of the deepest convolutional kernel can cover most or even the entire time length of the original input signal. By establishing global dependencies, deep neurons no longer focus on local details, but analyze the mutual relationship between different wave packets in the entire signal, i.e., long-range temporal dependencies. For example, the time difference (time of flight) between the incident wave S0 mode and the reflected wave S0 mode implies damage location information.

[0051] The network learns to integrate all the time domain features extracted in the previous stage into a global representation strongly related to the physical nature of the damage in this stage. It automatically learns advanced abstract features such as "damage index", "damage location encoding", "damage type encoding", etc. Finally, it outputs a high-level semantic feature vector to form a global feature vector. At this time, the global feature vector is still one-dimensional, but the information at each point is based on the analysis of the entire signal and is a highly abstract global feature representation.

[0052] Among them, establishing global dependencies means that deep neurons of the neural network can understand and integrate the interactions and internal relationships between different parts of the entire input signal, thereby forming a holistic and context-dependent understanding of the data. There are two main ways to establish global dependencies: (1) Physically: through the stacking of multi-layer network structures, the receptive field of deep neurons covers the entire input, providing global information. (2) Algorithmically: by learning the weight parameters of the convolutional kernel, the network learns to focus on which combinations of global information (such as time difference, energy ratio, phase difference) are crucial for the final task.

[0053] In the feature space mapping stage, first, the global feature vector (whose shape is [batch size, feature map length, feature channel number]) is flattened into a one-dimensional long vector. This one-dimensional long vector contains all high-level abstract features (such as various damage indices, mode energy ratios, etc.).

[0054] Next, the one-dimensional long vector is input into the fully connected layer. Each neuron of the fully connected layer is fully connected with each element of the input vector, and all features are combined through linear weighted summation (the calculation formula is: output = (w1*x1) + (w2*x2)+... + (wn*xn) + b). Among them, the weight (w) and the bias (b) are automatically learned and optimized through network training (such as back propagation algorithm). The essence is to learn how to assign different importance weights to different features—amplify the key features that are strongly related to the damage category, and suppress redundant or noisy features.

[0055] Finally, after one or more layers of linear combination and nonlinear activation (such as ReLU), the network projects the high-dimensional global feature vector to a new low-dimensional space (latent space), in which the distance between feature vectors of different damage categories is maximized, and feature vectors of the same damage are highly clustered. The final output is a low-dimensional, fixed-length source domain damage feature vector. The source domain damage feature vector is a highly condensed and abstracted representation of the scattering signal, directly representing the damage state of the scattering signal, and can be used for final classification or migration to the target domain.

[0056] On the basis of the above, as shown in Figure 4 , the source domain damage feature vector is input into the domain adaptive regression module to align the distribution of damage feature vectors between different regions, including the following steps: The batch feature matrix of the first region and the batch feature matrix of the second region are extracted from the source domain damage feature vector, respectively, and singular value decomposition is performed to obtain the orthogonal basis matrix of the first region and the orthogonal basis matrix of the second region . The orthogonal basis matrix of the first region and the orthogonal basis matrix of the second region are expressed as:

[0057] wherein, and are the feature matrices of the first region and the second region, respectively, and and each of all column vectors respectively span the subspace of the first region and the subspace of the second region, and singular values of the first and second regions, respectively, and singular vector matrices of the first and second regions, respectively.

[0058] orthogonal basis matrices of the first region and orthogonal basis matrices of the second region principal angles of the first region subspace and principal angles of the second region subspace principal angles of the first region subspace principal angles of the second region subspace are respectively expressed as:

[0059]

[0060] where, and are the standard orthogonal bases in the b-dimensional first region subspace and the b-dimensional second region subspace, respectively.

[0061] Of course, according to the number of regions contained in the planned source domain, the number of subspaces corresponds to the number of regions, and different subspaces are expressed as

[0062] According to the principal angles of the first region subspace and the principal angles of the second region subspace , the of the first region subspace and the of the second region subspace are constructed, respectively, which are used to measure the degree of feature subspace offset between different domains. The principal angle is a classic subspace distance measurement defined on the Grassmann manifold. When all principal angles are zero, it means that the two subspaces are completely coincident. Then the of different subspaces is expressed as:

[0063] where, is the relative subspace distance between the first region subspace and the second region subspace, is the principal angle between the first region subspace and the second region subspace.

[0064] To further maintain the structural consistency of the subspaces and improve the performance of the domain adaptation regression, a bases mismatch penalization (BMP) is introduced. The regularizer avoids the mismatch between the important and unimportant bases during the minimization of the RSD by constraining the sequential correspondence between the orthogonal basis matrices of the first and second regions, thereby better maintaining the semantic consistency of the feature representation. The bases mismatch penalization is expressed as

[0065] wherein is the bases mismatch penalization term, , are the basis weight matrices extracted in the subspace of the first region and the subspace of the second region, respectively.

[0066] Finally, the deep feature representation with high transferability is learned by minimizing the and between the first region and the second region, and the end-to-end cross-domain damage regression modeling is achieved, which is expressed as

[0067] wherein denotes the feature extractor, denotes the regressor, is the supervised loss of the source domain, is the loss, is the regularization loss, , is the hyperparameter for balancing the losses.

[0068] ensures that the model learns good regression prediction performance on the source domain. The minimization of the geometric distance between the source domain and the target domain representation subspaces can avoid changing the scale of the features by aligning the directions of the subspaces rather than directly aligning the instance features. The auxiliary and enhanced alignment effect. In the SVD decomposition, the size of the singular value represents the "importance" or "energy" of the corresponding orthogonal basis. ensures that when matching the two subspaces, the orthogonal bases with similar importance rankings are aligned with each other, rather than the important bases being matched with the unimportant bases, thereby maintaining the geometric structure of the deep representation.

[0069] The supervised loss of the source domain is expressed as

[0070] loss , expressed as:

[0071] regularization loss expressed as:

[0072] wherein, denotes sampling a batch of b sample instances from the source domain, is a sample input signal, is a sample corresponding label, is a batch distribution of source domain samples, is a batch distribution of target domain samples, is a square loss function (L2 loss), is a mathematical expectation operator, is a principal angle matrix between the first regional subspace and the second regional subspace.

[0073] After training, the trained cross-domain damage identification model is used for structural damage identification. First, the first guided wave signal and the second guided wave signal of the target domain are collected, and the two types of signals are pretreated and scattered signal extraction in turn to construct a target domain damage data set. The target domain damage data set is imported into the trained cross-domain damage identification model, and finally the damage position coordinate label of the target domain is output, realizing high-precision cross-domain damage identification.

[0074] Embodiment two In some preferred embodiments, the present application also provides simulation verification for the above method. The test object is an aviation structure with a size of 600mmx600mmx2.0mm, and a tape is used to simulate the damage on the aviation structure. Four boundaries are planned on the surface of the aviation structure to clearly divide the first region, the second region, the third region and the fourth region, and the damage positions in the four regions are completely different. Among them, the first region and the second region are the source domain, and the third region and the fourth region are the target domain.

[0075] Convolutional neural network (CNN), Bayesian neural network (DD-BNN), modular artificial neural network (M-ANN:) and the structural cross-domain damage identification method based on domain adaptation model of the present application are used for damage identification and positioning respectively, and the test results are recorded as shown in Tables 1, 2 and 3: Table 1 Cross-domain damage identification and positioning results of each model in the first region the second region

[0076] Table 2 Positioning results of cross-domain damage identification of each model in the first area Positioning results of cross-domain damage identification of the second area

[0077] Table 3 Positioning results of cross-domain damage identification of each model in the first area Positioning results of cross-domain damage identification of the second area

[0078] The experimental results show that, compared with the existing damage identification model, the structural cross-domain damage identification method based on the domain adaptation model has the smallest positioning error (MAE, MAPE and LRE) of cross-domain damage.

[0079] In summary, the structural cross-domain damage identification method based on the domain adaptation model provided by the application has the following advantages: compared with the prior art, the guided wave sensing technology is innovatively combined with the transfer learning through the domain adaptation regression model, the cross-domain feature deviation problem is effectively solved, the data distribution difference between the source domain and the target domain is reduced, when facing cross-regional damage data, the method still has good generalization performance, ensures high accuracy in damage identification in different monitoring areas, and breaks through the limitations of traditional deep learning models in cross-scene applications.

[0080] Although the terms such as source domain, target domain and the like are used more in this paper, the possibility of using other terms is not excluded. The use of these terms is only to facilitate the description and explanation of the essence of the application; any additional limitation is contrary to the spirit of the application.

[0081] In addition, those skilled in the art should understand that although there are many problems in the prior art, each embodiment or technical solution of the application can only be improved in one or several aspects, and it is not necessary to solve all the technical problems listed in the prior art or background art at the same time. Those skilled in the art should understand that what is not mentioned in a claim should not be regarded as a limitation on the claim.

[0082] Finally, it should be pointed out that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solution deviate from the scope of the technical solutions of the embodiments of the application.

Claims

1. A method for identifying cross-domain structural damage based on a domain adaptation model, characterized in that: Includes the following steps: Signal acquisition: A source domain is planned on the surface of the structure under test, and the source domain includes at least a first region and a second region. For the first region and the second region, two types of guided wave signals are acquired respectively: one is the first guided wave signal when the structure of the region is intact, and the other is the second guided wave signal when the structure of the region is damaged. Source domain dataset construction: Preprocess the first and second guided wave signals of the first region to extract the scattered signal of the first region; The first and second guided wave signals of the second region are preprocessed to extract the scattered signals of the second region; each scattered signal is labeled according to different damage locations and regions to form a complete and standardized source domain damage dataset. Cross-domain damage recognition model training: The source domain damage dataset is imported into the cross-domain damage recognition model for training. The cross-domain damage recognition model includes a feature extraction module and a domain adaptive regression module. The feature extraction module extracts features from the source domain damage dataset to obtain source domain damage feature vectors. The source domain damage feature vectors are then input into the domain adaptive regression module to align the distribution of damage feature vectors between different regions and optimize cross-regional adaptation capabilities. Target domain damage identification: The first and second guided wave signals of the target domain are collected, and the two types of signals are preprocessed and the scattered signals are extracted in sequence to construct the target domain damage dataset; the target domain damage dataset is imported into the trained cross-domain damage identification model, and the model outputs the damage location coordinates of the target domain.

2. The structural cross-domain damage identification method based on the domain adaptation model according to claim 1, characterized in that: Extracting the scattering signal involves the following steps: Noise interference in the first and second guided wave signals in this region is removed by digital filtering to obtain the filtered first and second guided wave signals. The filtered first guided wave signal and the filtered second guided wave signal are normalized to map the signal amplitude to the [-1,1] interval, thus obtaining the processed first guided wave signal and the processed second guided wave signal. Using the processed first guided wave signal as a baseline, and combining it with the processed second guided wave signal, signal differential processing is performed to extract the scattered signal containing damage information.

3. The structural cross-domain damage identification method based on the domain adaptation model according to claim 2, characterized in that: Damage information includes at least amplitude variation information, phase shift information, and time of flight information.

4. The structural cross-domain damage identification method based on the domain adaptation model according to claim 1, characterized in that: The feature extraction module consists of at least four cascaded convolution-pooling blocks, each containing a one-dimensional convolutional layer and a max pooling layer, wherein the convolutional kernel size is 1×K (K≥2) and uses the ReLU activation function, and the pooling kernel size is ≥2×2; The feature extraction module performs feature extraction on the source domain damage dataset, which includes the following steps: Local feature extraction: The scattering signals in the source domain damage dataset are scanned and linearly transformed using multiple one-dimensional convolution kernels to generate convolutional feature maps; the ReLU activation function is applied to the convolutional feature maps to obtain primary feature maps; Feature fusion: One-dimensional convolutional layers are stacked on the primary feature maps. Deep convolutional kernels are used to perform deep convolution processing on the primary feature maps. Pooling operations are performed through max pooling layers and ReLU activation function is applied to output intermediate feature maps. Global feature encoding: Establish global dependencies based on intermediate feature maps, learn to integrate all intermediate feature maps into a global representation that is strongly correlated with the physical nature of the damage, and output a global feature vector; Feature space mapping: The global feature vector is flattened and then input into the fully connected layer. The learning process linearly combines all elements of the global features to output the source domain damage feature vector.

5. The structural cross-domain damage identification method based on the domain adaptation model according to claim 4, characterized in that: The process of aligning the distribution of damage feature vectors between different regions by inputting the source domain damage feature vector into the domain adaptive regression module includes the following steps: Batch feature matrices of the first region are extracted from the source domain damage feature vector. Batch feature matrix of the second region Then, singular value decomposition is performed on each region to obtain the orthogonal basis matrix of the first region. Orthogonal basis matrix of the second region And through orthogonal basis mismatch penalty Constrain the order correspondence between orthogonal basis matrices of different domains; Orthogonal basis matrix based on the first region Orthogonal basis matrix of the second region Calculate the principal angle of the first subspace region. The main angle of the second region subspace ; Based on the principal angle of the first subspace The main angle of the second region subspace Construct the first region subspace respectively The second region subspace ; Ultimately, by minimizing the difference between the first and second regions... and To learn highly transferable deep feature representations, expressed as in, Indicates feature extractor, Indicates the regressor. For the source domain's supervision loss, for loss, for Regular loss, , Hyperparameters used to weigh various losses.

6. The structural cross-domain damage identification method based on the domain adaptation model according to claim 5, characterized in that: Orthogonal basis matrix of the first region Orthogonal basis matrix of the second region Expressed as: in, and These are the feature matrices for the first and second regions, respectively. and These are the singular values ​​of the first and second regions, respectively. and These are the singular vector matrices for the first and second regions, respectively.

7. The structural cross-domain damage identification method based on the domain adaptation model according to claim 5, characterized in that: Orthogonal basis mismatch penalty Expressed as: in, This is a mismatch penalty term for orthogonal bases. , These are the basis vector weight matrices extracted from the subspaces of the first and second regions, respectively.

8. The structural cross-domain damage identification method based on the domain adaptation model according to claim 5, characterized in that: The main angle of the subspace of the first region The principal angle of the subspace of the second region They are expressed as follows: in, and These are the orthonormal bases in the first and second region subspaces of dimension b, respectively.

9. The structural cross-domain damage identification method based on the domain adaptation model according to claim 8, characterized in that: Different subspaces Expressed as: in, This represents the relative subspace distance between the first and second subspace regions. The first subspace and the second subspace are the first subspace. From a personal perspective.

10. The structural cross-domain damage identification method based on the domain adaptation model according to claim 7, characterized in that: The supervised loss of the source domain is expressed as: but loss Expressed as: loss Expressed as: in, This indicates sampling a batch of b sample instances from the source domain. For the sample input signal, The labels corresponding to the samples. For the batch distribution of source domain samples, For the batch distribution of samples in the target domain, The squared loss function, For mathematical expectation operators, It is the principal angle matrix between the first and second subspace regions.

Images