Ultrasonic guided wave damage identification method based on twin data driving

Abstract

The invention discloses an ultrasonic guided wave damage identification method based on twin data driving, and the method comprises the steps: collecting an aluminum sheath damage detection twin model ultrasonic guided wave signal as a source domain data set 1 in a laboratory, extracting and constructing a source domain data set 2 through an ultrasonic guided wave signal mode, and constructing a target domain data set through a real detection signal; preprocessing a source domain data set and a target domain data set by adopting continuous wavelet transform according to the established data set; common feature extraction of the source domain data set and the target domain data set is realized through a ResNet50 network; a specific domain feature extractor is constructed, and two strategies of intra-domain alignment and inter-domain alignment are adopted to align data in different distributions so as to obtain domain invariant features and eliminate offset; and finally, designing a specific domain classifier to obtain category separable features, eliminating differences among different classifiers, and performing verification in damage detection of the cable sheath. According to the method, the interpretability, universality and generalization ability of the damage detection model are enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of ultrasonic guided wave nondestructive testing, and specifically relates to an ultrasonic guided wave damage identification method based on twin data driving. Background Art

[0002] With the development of ultrasonic guided wave nondestructive testing technology and AI technology, researchers have begun to use deep learning to explore the mapping relationship between guided wave signals and structural states. Deep learning technology can adaptively extract guided wave features, greatly reducing the need for prior knowledge and detection time. However, there are the following main problems in the application of deep learning technology to structural damage detection: (1) It is difficult to build a complete data set; (2) The use scenarios of structural damage identification methods are limited; (3) The physical interpretability of data-driven structural damage identification models is poor. Twin data-driven is a method that uses real-time data in the real world to create a virtual twin model to monitor, optimize and predict physical systems. It uses digital technology to build a virtual model to represent the actual behavior of physical entities. It uses a combination of technical means such as fusion analysis of diversified data, iterative optimization of the decision process, and interactive feedback between virtual and real to achieve deep fusion, symbiosis and all-round analysis between the real physical system and the virtual "twin" model. The twin data-driven method has higher accuracy, stronger generalization ability and better interpretability, and has greater research potential.

[0003] The existing cable aluminum sheath corrosion damage imaging method and device based on twin network and ultrasonic guided wave (CN115128166B) uses ultrasonic guided wave detection equipment to collect ultrasonic guided wave signals; standardizes the training signal set and the test signal set to obtain a standard training set and a standard test set; uses a deep convolutional neural network to reduce the dimension and reconstruct the standard training set to simulate the cable aluminum sheath corrosion damage signal set; constructs a twin network, performs deep learning and updates the twin network parameters to obtain a trained twin network; inputs the standard test set into the trained twin network to obtain the cable aluminum sheath corrosion damage index, and uses a probabilistic imaging method to present the cable aluminum sheath corrosion damage state. Ultrasonic guided waves are used to detect the cable aluminum sheath, and the twin network is trained, optimized and knowledge transferred by simulating corrosion damage signals to achieve corrosion damage imaging of the cable aluminum sheath. It uses the cable aluminum sheath corrosion damage data as the test set, and the cable aluminum sheath under normal conditions as the training set. The convolutional neural network is used to learn and construct a twin network to effectively identify and image the cable sheath corrosion damage. However, the method requires that the damage data is real sampling data, and the damage and normal data are collected on the same workpiece. The method is modeled by the original signal, which is difficult to use for structural health monitoring of workpieces in real situations, that is, it is impossible to perform structural health monitoring on different workpieces of the same material under unknown damage conditions. This method uses twin data to drive network training and adopts transfer learning methods to achieve ultrasonic guided wave damage identification. Summary of the invention

[0004] The main purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art and to provide an ultrasonic guided wave damage identification method based on twin data drive. The present invention combines the ultrasonic guided wave signal of the aluminum sheath damage detection twin model to construct a source domain data set 1, the ultrasonic guided wave signal modal extraction signal to construct a source domain data set 2, and the real detection signal to construct a target domain data set, and inputs its time-frequency characteristics into the twin data driven detection model, thereby realizing the identification of cable sheath damage using ultrasonic guided waves.

[0005] The present invention is achieved by at least one of the following technical solutions.

[0006] A twin data-driven ultrasonic guided wave damage identification method comprises the following steps:

[0007] S1. Collect the ultrasonic guided wave signals of the damage detection twin model as the source domain dataset 1, construct the source domain dataset 2 by extracting the ultrasonic guided wave signal modality, and construct the target domain dataset by the real detection signal;

[0008] S2, using continuous wavelet transform to preprocess the source domain data set and the target domain data set;

[0009] S3, the source domain dataset and the target domain dataset realize common feature extraction through deep learning network;

[0010] S4. Build a domain-specific feature extractor and use intra-domain alignment and inter-domain alignment strategies to align data with different distributions to obtain domain-invariant features and eliminate offsets.

[0011] S5. Design domain-specific classifiers to obtain category-separable features and eliminate the differences between different classifiers to achieve twin data-driven ultrasonic guided wave damage detection.

[0012] Furthermore, in step S1, the source domain dataset is is a twin model ultrasonic guided wave signal dataset, the source domain dataset is Construct a data set for ultrasonic guided wave mode extraction, where x i S,r and i S,r (p,q) are the i-th sample and the corresponding damage condition in the r-th source domain, r∈(1,2), p is the damage distance, q is the damage degree, n s is the number of labeled samples, the real detection signal constructs the target domain dataset where n t is the number of unlabeled samples, x j T is the jth target sample.

[0013] Furthermore, in step S1, the twin model is fine-tuned and calibrated in combination with the actual signal to achieve consistency between the twin model signal and the actual ultrasonic guided wave signal. The source domain data set S1 is constructed from the twin model signal group.

[0014] Furthermore, in step S1, the ultrasonic guided wave signal modal extraction is constructed, and a source domain data set S2 is created by establishing a peak-frequency modulation dictionary using a matching pursuit algorithm. The dictionary atoms are composed of sinusoidal modulation functions as basis functions, and their final shape is adjusted by three main parameters: frequency, number of peaks, and attenuation coefficient.

[0015] Furthermore, in step S2, the localization characteristics C(a, b) of the waveguide signal in the data set in the time-frequency domain are obtained by a two-dimensional representation obtained by continuous wavelet transform:

[0016]

[0017] Among them, a is the scaling parameter, b is the translation parameter, ψ(t) is the wavelet mother function, f(t) is the input signal, t is the time, and R is the local time domain.

[0018] Furthermore, in step S3, the source domain dataset and the target domain dataset are subjected to common feature extraction through a ResNet50 network, and the feature extraction process is as follows:

[0019] g S,r =G(x S,r ),g T =G(x T )

[0020] Among them, G(·) represents the ResNet50 public feature extraction network, g S,r is the feature learned from the rth source domain, g T is the learned target feature, x S,r is the rth source domain sample, x T is the target domain sample.

[0021] Furthermore, in step S4, the domain-specific feature extractor uses two strategies, intra-domain alignment and inter-domain alignment, to align data with different distributions to obtain domain-invariant features and eliminate offsets. The intra-domain alignment strategy uses the multi-core maximum mean difference as the distance metric between the source domain and the target domain to evaluate the distribution offset. Given the rth source feature and target features Then the square value of the maximum mean difference of multiple cores in the domain is for:

[0022]

[0023] Among them, n sis the number of labeled samples, n t is the number of unlabeled samples, h(·) is the feature kernel, and Denote the features learned by the i-th and j-th pairs of source and target domains respectively. The inter-domain alignment strategy is used to eliminate the distribution offset of all source and target domain pairs. and Represent the features learned by the i-th and j-th pairs of the source and target domains respectively, then the square value of the maximum mean difference of the multi-cores between the domains is for:

[0024]

[0025] in, and n fj are the number of features learned in the i-th and j-th pairs of source and target domains, respectively. The subscripts m and n are the accumulation coefficients.

[0026] Furthermore, in step S5, the domain-specific classifier is composed of R predictors Each predictor C k is a Softmax classifier that accepts the learned feature g from the rth source domain S,r , and output the corresponding classification label. The classification loss with R source domains is recorded as:

[0027]

[0028] in, is the cross entropy classification loss of the r-th source domain.

[0029] Furthermore, in step S5, the domain-specific feature extractor has an intra-domain offset alignment loss of:

[0030]

[0031] The inter-domain offset alignment loss is:

[0032]

[0033] The total loss of the deep adaptive representation module is:

[0034] L MMD =Loss+Loss'.

[0035] Furthermore, in step S5, the classifier difference metric of the specific domain classifier is composed of the absolute value difference of all classifier probability outputs of the target domain data, and the classifier difference metric loss L dl :

[0036]

[0037] Among them, E is the mathematical expectation, R is the number of classifiers, and g T is the learned target feature, C i (g T ) and C j (g T ) performs Softmax classification on the i-th and j-th pairs of learned features in the source and target domains, respectively.

[0038] Furthermore, in step S5, the loss of the domain-specific classifier network consists of three parts: classification loss, MK-MMD loss, and metric loss. By minimizing the classification loss, the network can accurately classify the source domain data; by minimizing the MK-MMD loss, the domain-invariant representation is learned; by minimizing the metric loss, all classifier differences can be removed and the same target samples can get the same prediction. The total loss formula is:

[0039] L=L cl +γ(L MMD +L dl )

[0040] Where γ is the loss weight.

[0041] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0042] The present invention combines twin models, modal extraction and real detection data to realize multi-domain migration identification of different damage conditions, and enhances the interpretability, versatility and generalization ability of the damage detection model.

[0043] The present invention adopts twin data-driven ultrasonic guided wave detection technology for detection, extracts the time-frequency features in the guided wave signal through continuous wavelet transform, and then uses the deep learning network as a common feature extractor. It also constructs a specific domain feature extractor and a specific domain classifier to realize the identification of the damage location and degree. It has been verified on the aluminum sheath of the cable, realizing the accurate detection and timely maintenance of the cable sheath. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0045] Figure 1 This is a flow chart of an ultrasonic guided wave damage identification method driven by twin data according to an embodiment of the present invention;

[0046] Figure 2This is a flow chart of continuous wavelet transform according to an embodiment of the present invention;

[0047] Figure 3 This is a flow chart of the ResNet50 common feature extractor according to an embodiment of the present invention;

[0048] Figure 4 This is a flow chart of a domain-specific feature extractor according to an embodiment of the present invention;

[0049] Figure 5 This is a flow chart of a specific domain classifier according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application.

[0051] Reference to "embodiments" in this application means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments.

[0052] See also Figure 1 to Figure 5 In this embodiment, a twin data-driven ultrasonic guided wave damage identification method includes the following steps:

[0053] S1. The laboratory collects ultrasonic guided wave signals of the damage detection twin model as source domain dataset 1, constructs source domain dataset 2 through ultrasonic guided wave signal modal extraction, and constructs target domain dataset through real detection signals;

[0054] S2, using continuous wavelet transform to preprocess the source domain data set and the target domain data set;

[0055] S3, source domain dataset and target domain dataset realize common feature extraction through ResNet50 network;

[0056] S4. Build a domain-specific feature extractor and use intra-domain alignment and inter-domain alignment strategies to align data with different distributions to obtain domain-invariant features and eliminate offsets.

[0057] S5. Design a domain-specific classifier to obtain category-separable features and eliminate the differences between different classifiers to achieve twin data-driven ultrasonic guided wave damage detection.

[0058] As an embodiment, in this embodiment, aluminum sheath damage detection is taken as an example. In step S1

[0059] The twin model is fine-tuned and calibrated in combination with the actual signal to achieve consistency between the twin model signal and the actual ultrasonic guided wave signal. The source domain data set S1 is constructed from the twin model signal group. The ultrasonic guided wave signal modal extraction is constructed by using the matching pursuit algorithm to establish a peak-frequency modulation dictionary to create a source domain data set S2. The dictionary atom is a sinusoidal modulation function as the basis function, and its final shape is adjusted by three main parameters: frequency, number of peaks, and attenuation coefficient; the target domain data set constructed by the real detection signal is composed of signal samples collected by the ultrasonic guided wave detector.

[0060] like Figure 2 As shown, in this embodiment, the wavelet transform in step S2 pre-processes the source domain data set and the target domain data set, specifically:

[0061] The two-dimensional representation obtained by continuous wavelet transform captures the localization characteristics of the waveguide signal in the data set in the time and frequency domain.

[0062]

[0063] Among them, a is the scaling parameter, b is the translation parameter, ψ(t) is the wavelet mother function, f(t) is the input signal, and t is time.

[0064] like Figure 3 As shown, in this embodiment, the source domain dataset and the target domain dataset described in step S3 realize common feature extraction through a deep learning network, specifically:

[0065] As an embodiment, the deep learning network uses the ResNet50 network. The public feature extractor learns discriminative features from multiple source data and target data. ResNet50 has five stages from input to output, and has a total of 50 layers: 2 layers (7×7) in Stage0, 9 layers (3×3) in Stage1, 12 layers (3×4) in Stage2, 18 layers (3×6) in Stage3, and 9 layers (3×3) in Stage4. Let G(·) represent the ResNet50 public feature extraction network, and the feature extraction process can be expressed as:

[0066] g S,r =G(x S,r ),g T =G(x T )

[0067] Among them, G(·) represents the ResNet50 public feature extraction network, g S,r is the feature learned from the rth source domain, g T is the learned target feature.

[0068] like Figure 4 As shown, in this embodiment, a specific domain feature extractor is constructed in step S4, and two strategies, intra-domain alignment and inter-domain alignment, are used to align data with different distributions to obtain domain-invariant features and eliminate offsets, specifically:

[0069] The domain-specific feature extractor is divided into two parts: intra-domain alignment and inter-domain alignment. The multi-kernel maximum mean difference (MK-MMD) in the intra-domain alignment strategy is used as the distance metric between the source domain and the target domain to evaluate the distribution shift. The smaller the MK-MMD value, the closer the two distributions are. Given the rth source feature and target features Then the square value of MK-MMD is:

[0070]

[0071] Where h(·) is the feature kernel. The intra-domain alignment strategy can significantly reduce the intra-domain offset between each pair of source and target domains, but ignores the inter-domain offset between all pairs of source and target domains. Therefore, we further combine the inter-domain alignment strategy to eliminate the distribution offset of all pairs of source and target domains. Let and denote the features learned by the i-th and j-th pairs of the source and target domains, respectively. Then the square value of the inter-domain MK-MMD is:

[0072]

[0073] in, and n fj are the number of features learned for the i-th and j-th pairs of source and target domains, respectively.

[0074] like Figure 5 As shown, in this embodiment, step S5 designs a specific domain classifier to obtain class-separable features and eliminates the differences between different classifiers to achieve cable sheath damage detection, specifically:

[0075] Design a classifier difference metric loss L consisting of the absolute value difference of the classifier probability output of all pairs of target domain data dl :

[0076]

[0077] Among them, E is the mathematical expectation, R is the number of classifiers, C is the Softmax classifier, and i and j are the features learned from the i-th and j-th pairs of source and target domains, respectively.

[0078] It should be noted that, for the sake of convenience, the aforementioned method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present invention is not limited to the described order of actions, because according to the present invention, certain steps can be performed in other orders or simultaneously.

[0079] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0080] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well.

Claims

1. A twin data-driven ultrasonic guided wave damage identification method, characterized in that: The steps include: S1. Collect the ultrasonic guided wave signals of the damage detection twin model as the source domain dataset 1, construct the source domain dataset 2 by extracting the ultrasonic guided wave signal modality, and construct the target domain dataset by the real detection signal; S2, using continuous wavelet transform to preprocess the source domain data set and the target domain data set; S3, the source domain dataset and the target domain dataset realize common feature extraction through deep learning network; S4. Build a domain-specific feature extractor and use intra-domain alignment and inter-domain alignment strategies to align data with different distributions to obtain domain-invariant features and eliminate offsets. S5. Design domain-specific classifiers to obtain category-separable features and eliminate the differences between different classifiers to achieve twin data-driven ultrasonic guided wave damage detection.

2. The twin data-driven ultrasonic guided wave damage identification method according to claim 1 is characterized in that: In step S1, the source domain dataset is is a twin model ultrasonic guided wave signal dataset, the source domain dataset is Construct a data set for ultrasonic guided wave mode extraction, where x i S,r and i S,r (p,q) are the i-th sample and the corresponding damage condition in the r-th source domain, r∈(1,2), p is the damage distance, q is the damage degree, n s is the number of labeled samples, the real detection signal constructs the target domain dataset Where n t is the number of unlabeled samples, x j T is the jth target sample.

3. The twin data-driven ultrasonic guided wave damage identification method according to claim 1 is characterized in that: In step S1, the twin model is fine-tuned and calibrated with actual signals to achieve consistency between the twin model signal and the actual ultrasonic guided wave signal. The source domain data set S1 is constructed from the twin model signal group.

4. The twin data-driven ultrasonic guided wave damage identification method according to claim 1 is characterized in that: In step S1, the ultrasonic guided wave signal modal extraction is constructed, and a source domain data set S2 is created by establishing a peak-frequency modulation dictionary using a matching pursuit algorithm. The dictionary atoms are based on sinusoidal modulation functions, and their final shape is adjusted by three main parameters: frequency, number of peaks, and attenuation coefficient.

5. The twin data-driven ultrasonic guided wave damage identification method according to claim 1, characterized in that: In step S2, the localization characteristics C(a, b) of the waveguide signal in the data set in the time-frequency domain are obtained by a two-dimensional representation obtained by continuous wavelet transform: Among them, a is the scaling parameter, b is the translation parameter, ψ(t) is the wavelet mother function, f(t) is the input signal, t is the time, and R is the local time domain.

6. The twin data-driven ultrasonic guided wave damage identification method according to claim 1 is characterized in that: Step S3, the source domain dataset and the target domain dataset are subjected to common feature extraction through the ResNet50 network, and the feature extraction process is as follows: g S,r =G(x S,r ),g T =G(x T ) Among them, G(·) represents the ResNet50 public feature extraction network, g S,r is the feature learned from the rth source domain, g T is the learned target feature, x S,r is the rth source domain sample, x T is the target domain sample.

7. The twin data-driven ultrasonic guided wave damage identification method according to claim 1 is characterized in that: In step S4, the domain-specific feature extractor uses two strategies, intra-domain alignment and inter-domain alignment, to align data with different distributions to obtain domain-invariant features and eliminate offsets. The intra-domain alignment strategy uses the multi-core maximum mean difference as the distance metric between the source domain and the target domain to evaluate the distribution offset. and target features Then the square value of the maximum mean difference of multiple cores in the domain is for: Among them, n s is the number of labeled samples, n t is the number of unlabeled samples, h(·) is the feature kernel, and Denote the features learned by the i-th and j-th pairs of source and target domains respectively. The inter-domain alignment strategy is used to eliminate the distribution offset of all source and target domain pairs. and Represent the features learned by the i-th and j-th pairs of the source and target domains respectively, then the square value of the maximum mean difference of the multi-cores between the domains is for: Among them, n fi and n fj are the number of features learned in the i-th and j-th pairs of source and target domains, respectively. The subscripts m and n are the accumulation coefficients.

8. The twin data-driven ultrasonic guided wave damage identification method according to claim 1, characterized in that: In step S5, the domain-specific classifier consists of R predictors Each predictor C k is a Softmax classifier that accepts the learned feature g from the rth source domain S,r , and output the corresponding classification label, the classification loss with R source domains is recorded as: in, is the cross entropy classification loss of the r-th source domain.

9. The twin data-driven ultrasonic guided wave damage identification method according to claim 1, characterized in that: In step S5, the domain-specific feature extractor has an intra-domain offset alignment loss of: The inter-domain offset alignment loss is: The total loss of the deep adaptive representation module is: L MMD =Loss+Loss'。 10. The twin data-driven ultrasonic guided wave damage identification method according to claim 1, characterized in that: In step S5, the classifier difference metric of the specific domain classifier is composed of the absolute value difference of the classifier probability output of all pairs of target domain data, and the classifier difference metric loss L dl : Among them, E is the mathematical expectation, R is the number of classifiers, and g T is the learned target feature, C i (g T ) and C j (g T ) performs Softmax classification on the i-th and j-th pairs of learned features in the source and target domains, respectively.

Images