An ultrasonic guided wave damage identification method for multi-structure data feature migration

Abstract

The application discloses a kind of multi-structure data feature migration ultrasonic guided wave damage identification methods, including collecting single structure ultrasonic guided wave scattering signal as source domain dataset, collect multi-structure ultrasonic guided wave scattering signal as target domain dataset;Using convolutional neural network obtains the characteristic information of structure damage;Adopt maximum mean difference method to construct the quantitative evaluation mechanism of the difference of different structure guided wave signals;Feature matching loss and source domain training label loss are used to optimize deep network model;For source domain and target domain signal, construct multi-layer scattering feature mapping relationship;Realize scattering feature adaptation in multiple hidden layers;Finally, combined with multi-sensor damage index, the effectiveness of damage feature is verified using structure image and damage positioning.The application uses a variety of structure real detection data modeling to realize the transfer detection and evaluation of different structure damage conditions, and enhances the universality and generalization ability of ultrasonic guided wave big data detection model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of ultrasonic guided wave structure health monitoring, and particularly relates to an ultrasonic guided wave damage identification method based on multi-structure data feature transfer. BACKGROUND

[0002] A multi-structure material rapid detection system is one of the engineering goals in the field of ultrasonic guided wave structure health monitoring. Multi-structure material damage identification has the following main problems: (1) For traditional mechanism analysis methods, the dispersion characteristics, wave speed and propagation characteristics of different materials are different, and the monitoring data collected in existing materials are not suitable for multi-material damage analysis, and the model universality and generalization are not good. (2) The large data training set of deep network modeling is difficult to meet the requirements of multi-structure materials. The damage modes of these new materials have certain correlation, but are not completely consistent, which makes it difficult for the same monitoring model to adapt to the data of different monitoring structures. (3) The change of material properties directly affects the ultrasonic propagation speed and attenuation characteristics, and then affects the amplitude and phase of the detection signal, making it difficult to distinguish damage characteristics and environmental characteristics. Therefore, a damage feature transfer strategy is established, and damage data of different structures are used in actual detection, which is a new idea to solve this problem. Through this way, a feature-based transfer learning method is used to reduce the feature gap between the source domain and the target domain. Combined with a multi-sensor damage index, a structure image and damage positioning are obtained.

[0003] A CFRP material structure damage area positioning method based on transfer learning (CN118380079A) uses composite plate simulation data as a source domain, composite plate actual data as a target domain, and uses a transfer learning conditional operator to fine-tune a deep learning model, thereby successfully realizing damage positioning of the target domain composite plate; but this method requires that the target domain data and the actual measured data are completely consistent, so this method cannot be used for structure health monitoring on other panels of the same type, that is, it cannot realize transfer adaptation without the participation of target domain data in training; and this method models through original signals, and the present method models through scattered signals, and extracts inherent scattering features of signals for damage identification. SUMMARY

[0004] The main purpose of the present application is to overcome the shortcomings and deficiencies of the prior art, and to provide an ultrasonic guided wave damage identification method based on multi-structure data feature transfer. The present application collects ultrasonic guided wave scattered signals of a single structure as a source domain data set, collects ultrasonic guided wave scattered signals of multiple structures as a target domain data set, inputs the hidden layer multi-structure information into a detection model, and thereby realizes monitoring of multi-structure cross-workpiece damage using ultrasonic guided waves.

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

[0006] The ultrasonic guided wave damage identification method based on multi-structure data feature transfer comprises the following steps:

[0007] S1, collect single structure ultrasonic guided wave scattering signals as source domain data sets, and collect multi-structure ultrasonic guided wave scattering signals as target domain data sets;

[0008] S2, use a convolutional neural network to obtain scattering feature information of structural damage;

[0009] S3, adopt a maximum mean difference method to construct a quantitative evaluation mechanism of differences in guided wave scattering signals of different structures; feature matching loss and label loss of source domain training are used to optimize the convolutional neural network; a multi-layer scattering feature mapping relationship is constructed for source domain and target domain signals; and scattering feature adaptation in multiple hidden layers is realized;

[0010] S4, combine a multi-sensor damage index, and use structural images and damage positioning to verify the effectiveness of damage features.

[0011] Further, in step S1, the source domain data set is wherein are the s-th sample and the corresponding damage condition in the source domain, s∈n s , n s is the number of labeled samples; the target domain data set is wherein n t is the number of unlabeled samples, is the n t th target sample; a band-pass filter is used to denoise the signal, and the airborne signal in the monitoring signal is removed to adapt to the influence of different position sensors in the array sensor; the i-th element X i in each guided wave sample X f is normalized by a normalization factor N

[0012]

[0013] Further, in step S1, the source domain data is single structure ultrasonic guided wave scattering signals, and the target domain data is multi-structure ultrasonic guided wave scattering signals; the single structure uses isotropic material structural parts, and a single sensor experiment is used to construct a structural health monitoring source domain data set; the multi-structure uses anisotropic material or composite material structural parts, and a single sensor experiment is used to construct a structural health monitoring target domain data set.

[0014] Further, in step S2, the convolutional neural network is used to obtain feature information of structural damage, and a multi-layer convolutional neural network with N L convolutional layers is constructed to obtain hidden mapping features of the original signal X i in different domains; for the convolutional network, the k-th element h i,k of the convolution result obtained by the i-th convolution kernel of the first layer is:

[0015]

[0016] wherein m1 is the size of the convolution kernel, l is the number of convolution layers, is the zero padding vector of the input sample, K i,j is the jth element of the ith convolution kernel, n1 is the number of convolution kernels; the convolution result is activated using a Relu function.

[0017] Further, the step S3 adopts the maximum mean difference method to construct the quantitative evaluation mechanism of the difference of the different structure guided wave signals, and the specific process is: the maximum mean difference is used to calculate the scattering feature distance between the source domain X S and the target domain X T .

[0018]

[0019] wherein f(.) represents a feature mapping function, each sample is mapped to a Hilbert space through the function, n s is the number of labeled samples, n t is the number of unlabeled samples; X i , X j represents the ith, jth element in each guided wave sample X.

[0020] Further, the feature matching loss and the label loss of the source domain training are used to optimize the deep network model, and the hidden feature distance between the source domain feature h S and the target domain feature h T . is calculated in the following manner:

[0021]

[0022] The overall loss function of the adaptive deep transfer learning method of the scatter feature is calculated in the following manner:

[0023]

[0024] wherein L total is the regression loss, is the weight coefficient of the hidden feature distance loss in the convolution layer N l , c is the amplification factor of the regression loss, N S is the source domain, N T is the target domain, y pre is the network output, y s is the known label.

[0025] Further, in step S3, a multi-layer feature mapping relationship is constructed for the source domain and target domain signals to realize feature adaptation in multiple hidden layers.

[0026] Further, in step S4, the multi-sensor damage index is combined to obtain a structural image and damage positioning, and the specific steps are as follows: a ring array anisotropic probability imaging method with excitation end correction is improved through discrete point-by-point calculation, and the pixel value of the discrete point depends on the damage index and the transient imaging weight P(x, y):

[0027]

[0028] Wherein, the damage probability P r in the discrete coordinate point (x, y) is estimated through superposition calculation of all transmission paths N P , DI r is the damage index obtained by the convolutional neural network, W D (x, y) is the excitation end imaging correction coefficient, W r (x, y) is the linear attenuation imaging weight. The difference between the damaged signal and the normal signal is used to obtain the time of flight, and then the imaging path where the damage is located is determined.

[0029] Further, in step S4, the multi-sensor damage index is combined to obtain a structural image and damage positioning, and the specific steps are as follows: a ring array anisotropic probability imaging method with excitation end correction is improved through discrete point-by-point calculation, and the pixel value of the discrete point depends on the damage index and the transient imaging weight P(x, y):

[0030] Further, in step S4, the multi-sensor damage index is combined to obtain a structural image and damage positioning, and the specific steps are as follows: a ring array anisotropic probability imaging method with excitation end correction is improved through discrete point-by-point calculation, and the pixel value of the discrete point depends on the damage index and the transient imaging weight P(x, y):

[0031] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0032] The present application adopts feature migration technology to realize ultrasonic guided wave structural health monitoring, extracts multiple structural features in hidden layers through convolutional neural network and maximum average difference algorithm, uses feature optimization algorithm to reduce the feature gap between the source domain and the target domain, optimizes the model, and finally combines the multi-sensor damage index to obtain a structural image and damage positioning, realizing ultrasonic guided wave structural health monitoring of multiple structural data. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0034] Figure 1 The flow chart of the ultrasonic guided wave damage identification method of the multi-structure data feature migration of the embodiment of the present application;

[0035] Figure 2 The flow chart of the feature information of the structure damage obtained by the convolutional neural network of the embodiment of the present application;

[0036] Figure 3 The flow chart of the multi-hidden layer feature adaptation of the embodiment of the present application. DETAILED DESCRIPTION

[0037] In order to make the person skilled in the art better understand the present application, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0038] In the present application, the phrase "embodiment" means that the specific features, structures or characteristics described in combination with the embodiment can be contained in at least one embodiment of the present application. The appearance of this phrase in the specification does not necessarily mean the same embodiment, nor is it an independent or alternative embodiment to other embodiments. Those skilled in the art explicitly and implicitly understand that the embodiments described in the present application can be combined with other embodiments.

[0039] As shown in the figure, the multi-structure data feature migration ultrasonic guided wave damage identification method of the embodiment includes the following steps: Figure 1

[0040] S1, collect single-structure ultrasonic guided wave scattering signals as source domain data set, and collect multi-structure ultrasonic guided wave scattering signals as target domain data set;

[0041] S2, use a convolutional neural network to obtain scattering feature information of structure damage.

[0042] ​S3, a maximum mean difference method is used to construct a quantitative evaluation mechanism for the differences between different structural guided wave signals; feature matching loss and source domain training label loss are used to optimize the deep network model; a multi-layer scattering feature mapping relationship is constructed for source domain and target domain signals; scattering feature adaptation in multiple hidden layers is achieved;

[0043] S4, combined with the multi-sensor damage index, the effectiveness of the damage feature is verified by using the structural image and damage positioning;

[0044] In waveguide structures made of different materials, the differences in ultrasonic guided wave mode response are reflected in the number, type, amplitude and position of the ultrasonic guided wave mode. The proposed feature transfer method uses the regenerated Hilbert space to obtain the feature distance and feature distribution distance of different materials, and constructs a similarity measure of multi-layer mapping features to improve the consistency of feature representation. This method can be divided into four stages: signal preprocessing, convolution feature extraction, scattering feature adaptation and multi-sensor probability imaging.

[0045] In this embodiment, the signals collected from aluminum plates are taken as the source domain, and the signals collected from multi-sensor composite plates are taken as the target domain. In step S1, the single structure belongs to the aluminum plate, and artificial crack damage is simulated. The thickness of the aluminum plate sample is 7mm, and the damage size on the back is 5mm long and 1mm wide. Different damages are set by gradually increasing the damage depth, and the crack damage depth is increased from 3mm to 6mm every 1mm. The PZT exciter and receiver are installed on the damaged side. A total of 150 samples are collected, of which 30 data samples are in normal and 4 damaged states respectively. The source domain dataset where x S i and y S i are the i-th sample in the source domain and the corresponding damage, respectively, and n s is the number of labeled samples.

[0046] As an embodiment, the multi-structure is a multi-material composite plate, and 150 samples of T300 composite plate structures in normal and 4 damage states are collected in this embodiment. The sample size is 300*300mm, and the thickness is 3mm. A total of 15 layers of layered carbon fiber layers are used to construct the sample, and the thickness of each layer of carbon fiber is 0.2mm. Due to the layering of this composite structure, the dispersion characteristics of ultrasonic guided waves and the speed of each direction are different. The dispersion characteristics and anisotropy of the sample are further analyzed. At a frequency of 200 kHz, the S0 mode propagating along the 0° direction has a speed of 7289m / s, while the S0 mode propagating along the 90° direction has a speed of 6917m / s. The same artificial crack damage is set, and the PZT is installed. Different damage levels with crack damage depths of 1.5mm to 3mm are set by increasing the crack damage depth (0.5mm each time). The target domain dataset where nt is the number of unlabeled samples, x Tj is the jth target sample. Normalization of the signal is required to balance the signals from multiple sensors. A band-pass filter of 20-800 kHz is used to denoise the signal. The on-board signal in the monitoring signal is removed to adapt to the influence of sensors at different positions in the array. The ith element in each guided wave sample X i will be normalized by a normalization factor N f is normalized, as an example, N f is set to 10.

[0047]

[0048] A target domain detection experiment was conducted in the same type of composite panel. The sample is a plate structure with a side length of 600 mm and a thickness of 3 mm. Sixteen PZT sensors are installed on the composite panel to form a ring array, with PZT 1 as the driving PZT. Artificial crack damage (four levels, including 5 mm, 10 mm, 15 mm and 20 mm) is set on the sample. In each condition, each of the 15 receiving PZT sensors collects 30 samples. A total of 2250 guided wave signals are collected, including 450 normal samples and 1800 damage samples under 4 damage conditions.

[0049] As Figure 2 shown, in this embodiment, the feature information of the structure damage is obtained using the convolutional neural network in step S2, and a multi-layer feature mapping relationship is constructed for the source domain and the target domain signal. A multi-layer convolutional neural network with N L convolutional layers is constructed to obtain the hidden mapping features of the original signal X i in different domains.

[0050] As an example, the convolutional neural network model uses 5 convolutional layers and 3 fully connected layers, in which feature distance matching is performed in the fully connected layer. The data sample is intercepted, and the signal length is set to 1024. The element distance loss weight of each layer is set to 0.01. In terms of optimizer selection, the Adam optimizer with a learning rate of 0.001 is used. Finally, the total loss of backpropagation is obtained by the label distance and the feature distance. During the training process, the single PZT monitoring data of the aluminum plate and the composite plate are respectively taken as the source domain and the target domain for network training. The PCA algorithm is used to extract the main components of the hidden features, and the feature dimension is reduced to realize visualization. For the convolutional network, the kth element h i,k of the convolution result obtained by the ith convolution kernel of the first layer is:

[0051]

[0052] where m1 is the size of the convolution kernel, l is the number of convolution layers, is the zero padding vector of the input sample, K i,j is the jth element of the ith convolution kernel, and n1 is the number of convolution kernels. As an embodiment, the convolution result is activated using the Relu function. In the damage assessment, the model adopts a regression algorithm instead of a classification algorithm, so slight fluctuations can be seen under different damage degrees.

[0053] In this embodiment, feature adaptation is achieved in the plurality of hidden layers described in step S3. The specific process is as follows: first, the source domain and the target domain data are respectively input into the convolutional neural network, and then the maximum mean difference loss is calculated for the data of each layer of the two domains. The maximum mean difference loss is added to the model training loss, and finally the model is optimized to realize transfer learning. The maximum mean difference loss is calculated between the source domain X S and the target domain X T :

[0054]

[0055] where f(.) represents a feature mapping function, and each sample is mapped to a Hilbert space through the function. In the reproducing kernel Hilbert space, MMD can be realized by bringing the source domain data full matrix [x i S x i* S ] and the target domain data full matrix [x j T x j* T ] into the Gaussian kernel k(x i ,x j ):

[0056]

[0057] The hidden feature distance between the source domain feature h S and the target domain feature h T is calculated as follows:

[0058]

[0059] In order to enhance the similarity of the features learned from different fields and ensure the consistency of the model recognition effect, the overall loss function of the scatter feature adaptive deep transfer learning method can be calculated as follows:

[0060]

[0061] where L totalis the regression loss, is convolutional layer N l the weight coefficient of the hidden feature distance loss. c is the magnification factor of the regression loss, N S is the source domain, y pre is the network output, y s is the known label.

[0062] Therefore, the similarity of the features learned from different domains is enhanced, and the consistency of the model recognition effect is ensured.

[0063] As Figure 3 shown in the embodiment, the combination of the multi-sensor damage index in the S4 step obtains the structural image and the damage positioning. In this verification, another similar composite plate with multiple sensors is used as a new monitoring sample. After the model training on the source domain samples of the aluminum plate and the target domain sample, the damage index is obtained. In the pre-training process, the aluminum plate and the composite plate are used as the source domain and the target domain, and the deep learning network for guided wave feature extraction is calibrated to obtain an automatic damage recognition model. The specific steps are as follows: through discrete point-by-point calculation, the anisotropic probability imaging method with excitation end correction of the ring array is improved. The pixel value of the discrete point depends on the damage index and the transient imaging weight:

[0064]

[0065] Among them, the damage probability P r (x,y) in the discrete coordinate point (x,y) can be estimated by superposition calculation of all transmission paths N P , W D (x,y) is the excitation end imaging correction coefficient, W r (x,y) is the linear attenuation imaging weight. DI r is the damage index obtained by the model. The higher the DI value, the higher the probability of structural damage positioning in the corresponding imaging path. For each receiving PZT sensor, the damage index can accurately reflect four different damage degrees. The difference between the damaged signal and the normal signal is used to obtain the time of flight, and then the imaging path where the damage is located is determined. Through the discretization of the monitored structure, the conversion time from the discrete coordinate point to the driving PZT and the receiving PZT can be obtained. The imaging time coefficient C T is defined as:

[0066] C T =(t s '-t s ) / t0=[(t1'+t2')-(t1+t2)] / t0

[0067] Where t0 is the time of flight of the direct signal. t1 and t2 are the time of flight of the wave signal from the transmitter to the damage location and the time of flight of the wave signal from the damage location to the receiving PZT, respectively. Through CT The value and the scaling factor calculate the linear attenuation imaging weight γ. When the discrete coordinate point is far away from the actual damage point, the damage probability decreases, C T The value becomes larger:

[0068]

[0069] The amplitude threshold is used to obtain the time of flight (ToF), and the damage path is determined by the ToF. The damage index is superimposed on the damage path to form the structural image of the composite panel. In traditional probability imaging, the imaging path is an ellipse because the sound speed is the same in all directions. The composite panel used in this work is 0 / 90 layered, and the speed difference in each direction is large, so the imaging path forms a rectangular path instead of an elliptical path. After setting a 96% threshold to filter the image, the real damage location can be accurately located in the structural image. After calculating the damage probability value of each discrete point through each monitoring path, the structural image of the monitored structure is finally obtained. The position with the maximum pixel value in the image should be the damage location.

[0070] It should be noted that, for the foregoing method embodiments, in order to facilitate description, they are all described as a series of action combinations, but those skilled in the art should know that the present application is not limited to the action sequence described, because according to the present application, certain steps can be performed in other sequences or simultaneously.

[0071] The technical features of the above embodiments can be combined in any way. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not contradict, they should be considered as within the scope of the present application.

[0072] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present application should be considered as equivalent replacement modes, and are all included in the protection scope of the present application.

Claims

1. A method for ultrasonic guided wave damage identification with multi-structure data feature migration, characterized in that, Comprise the following steps: S1, collect single structure ultrasonic guided wave scattering signal as source domain data set, collect multi-structure ultrasonic guided wave scattering signal as target domain data set; S2, using a convolutional neural network to obtain scattering characteristic information of structural damage, specifically including constructing a multi-layer convolutional neural network with N L Convolutional layers of the convolutional neural network to obtain the original signal X i Hidden mapping features in different domains; for Convolutional network, first layer i The convolution result obtained from the convolution kernel is the first... k element for: wherein is the convolution kernel size, l is the convolution layer number, is the zero padding vector of the input sample, K i,j is the element of the i th convolution kernel, j th element, is the number of convolution kernels; the convolution result is activated using the Relu function. S3, a quantitative evaluation mechanism of the difference of guided wave scattering signals of different structures is constructed by using the maximum mean difference method; the feature matching loss and the label loss of the source domain training are used to optimize the convolutional neural network; multi-layer scattering feature mapping relationship is constructed for source domain and target domain signals; Realize scattering feature adaptation in multiple hidden layers; The quantitative evaluation mechanism of differences of different structure guided wave signals is constructed by using the maximum mean difference method, and the specific process is as follows: the maximum mean difference is used to calculate the source domain X S The scattering feature distance between the target domain X T and the target domain wherein represents a feature mapping function by which each sample is mapped to a Hilbert space, n s is the number of labeled samples, n t is the number of unlabeled samples; , represents the i-th element in the j-th guided wave sample X , i , j . The feature matching loss and the label loss of the source domain training are used to optimize the deep network model, and the hidden feature distance between the source domain feature h S and the target domain feature h T is calculated by the following way: ​ The overall loss function of the scattering feature adaptive deep transfer learning method is calculated by the following method: wherein is a regression loss, is convolutional layers N l a weight coefficient of the hidden feature distance loss in the middle, c is a magnification factor of the regression loss, N S is a source domain, N T is a target domain, y pre is a network output, y s is a known label; S4, combined with multi-sensor damage index, the effectiveness of damage feature is verified by using structure image and damage positioning.

2. The method of claim 1, wherein, In step S1, the source domain dataset is ,in These are the first in the source domain s Each sample and its corresponding damage status, s∈n s , n s To label the number of samples; the target domain dataset is ,in n t This represents the number of unlabeled samples. For the first n t One target sample; The signal is denoised by a band-pass filter, and the airborne signal in the monitoring signal is removed to adapt to the influence of different position sensors in the array sensor, and the first element in each guided wave sample X i X i Normalization is performed by a normalization factor N f :​​ 。 3. The ultrasonic guided wave damage identification method of claim 1, wherein, In step S1, the source domain data is single structure ultrasonic guided wave scattering signal, and the target domain data is multi-structure ultrasonic guided wave scattering signal; single structure uses isotropic material structure, and single sensor experiment is used to construct structure health monitoring source domain data set; multi-structure uses anisotropic material or composite material structure, and single sensor experiment is used to construct structure health monitoring target domain data set.

4. The ultrasonic guided wave damage identification method of claim 1, wherein, In step S3, multi-layer feature mapping relationship is constructed for source domain and target domain signals, and feature adaptation in multiple hidden layers is realized; the parameter features of source domain and target domain are extracted by convolutional neural network, the model loss is optimized, the feature distance is reduced, and the scattering feature adaptation in hidden layer is realized.

5. The ultrasonic guided wave damage identification method of claim 1, wherein, In step S4, the combined multi-sensor damage index is used to obtain a structural image and damage localization. The steps are: the ring array anisotropic probabilistic imaging method with excitation end correction is improved by discrete point-by-point calculation, and the pixel value of the discrete point depends on the damage index and the transient imaging weight : Wherein the damage probability in the discrete coordinate point (x,y) P r (x,y) Estimate by superimposed calculation of all transmission paths N P Di r The damage index obtained by the convolutional neural network, W D (x,y) Is the excitation end imaging correction coefficient, W r (x,y) Is the linear attenuation imaging weight, the difference between the damaged signal and the normal signal is used to obtain the time of flight, and then the imaging path where the damage is located is determined.​​ 6. The method of claim 1, wherein, In step S4, combined with multi-sensor damage index, the damage state is evaluated by using scattering feature, and then structure image and damage positioning are obtained; array sensor is used to arrange damage monitoring platform, and according to the damage evaluation result and probability imaging technology of convolutional neural network on scattering signal, the structure damage is detected.

7. The ultrasonic guided wave damage identification method of migration of multi-structural data features of claim 1, wherein, In step S4, the combination of multi-sensor damage index is used to evaluate the damage state by using scattering feature, and then structure image and damage positioning are obtained, and the monitoring sample is a structure of different materials, specifically anisotropic material or composite material, and the target domain data is collected from different samples of the same material.

Images