Metal crack damage guided wave identification method based on numerical simulation and depth migration

Abstract

The invention belongs to the technical field of aircraft strength experiments, and particularly relates to a metal crack damage guided wave identification method based on numerical simulation and depth migration. The method comprises the following steps: establishing a metal fatigue crack damage guided wave monitoring simulation model, and obtaining numerical crack damage guided wave monitoring data; expanding the numerical crack damage guided wave monitoring data to obtain simulated source domain crack damage guided wave monitoring data; acquiring real crack damage guided wave monitoring data of a plurality of real metal fatigue test pieces, and expanding the real crack damage guided wave monitoring data to obtain real target domain crack damage guided wave monitoring data; and establishing a deep transfer learning model based on distribution distance measurement and a domain adversarial neural network, and carrying out domain adversarial training, so that a classifier aiming at simulated source domain crack damage guided wave monitoring data in the deep transfer learning model can be directly migrated to real target domain crack damage guided wave monitoring data.

Description

Technical Field

[0001] This application belongs to the technical field of aircraft strength experiments, and particularly relates to a method for identifying metal crack damage guided waves based on numerical simulation and deep transfer. Background Technique

[0002] As an online and quasi-real-time structural damage monitoring means, structural health monitoring technology acquires signals related to the structural health state through various advanced sensor networks integrated on the structure, and conducts health state assessment, which is of great significance for ensuring structural integrity. Among them, the structural damage monitoring technology based on guided waves has received extensive attention from researchers due to the characteristics of long-distance propagation of guided waves in the structure, sensitivity to small damage, and large-area monitoring.

[0003] In recent years, with the rapid development of artificial intelligence and computer technology, the intelligent monitoring methods of guided wave damage based on machine learning and deep learning have developed rapidly. Among these methods, the guided wave damage monitoring methods based on machine learning models such as support vector machines, decision trees, and artificial neural networks establish the mapping relationship between artificial features in time domain, frequency domain, time-frequency domain, etc. and the structural health state, so as to realize the discrimination of the unknown structural damage state. The accuracy of the intelligent monitoring model of guided wave damage depends on multiple aspects, such as multi-dimensional and multi-domain artificial features based on expert knowledge, the network architecture of the machine learning model, etc.

[0004] Compared with machine learning models, deep learning models such as convolutional neural networks, deep encoders, and long short-term memory networks are new methods for intelligent monitoring of guided wave damage that have been actively explored in recent years due to their powerful automatic feature mining and non-linear mapping capabilities. However, all the structural damage intelligent monitoring methods based on guided wave data driving require a large amount of completely labeled training data, and the training data and test data must satisfy independent and identically distributed. However, in the actual engineering application process, due to the limitations of test time and test cost, usually only a large amount of non-damaged monitoring data can be obtained. And sufficient real damage monitoring data can be obtained through numerical simulation and fatigue tests of simple test pieces, but their spatial distribution is different from that of complex structure damage monitoring data, resulting in the difficulty of directly applying the intelligent monitoring model established for simple structure damage.

[0005] Therefore, it is hoped that there is a technical solution to overcome or at least mitigate at least one of the above defects of the existing technology. Summary of the Invention

[0006] The purpose of this application is to provide a method for identifying metal crack damage guided waves based on numerical simulation and deep transfer, so as to solve the problem that it is difficult to obtain a large amount of completely labeled damage samples in the engineering scenario, which restricts the application of the intelligent monitoring model.

[0007] The technical solution of this application is:

[0008] A method for identifying metal crack damage guided waves based on numerical simulation and deep transfer, comprising:

[0009] Step 1, establish a numerical simulation model for monitoring metal fatigue crack damage guided waves, and obtain numerical crack damage guided wave monitoring data;

[0010] Step 2, convert the numerical crack damage guided wave monitoring data into an analytical signal, and perform data augmentation on the analytical signal to obtain simulated source domain crack damage guided wave monitoring data;

[0011] Step 3, conduct fatigue tests on multiple real metal fatigue specimens to obtain real crack damage guided wave monitoring data, convert the real crack damage guided wave monitoring data into an analytical signal, and perform data augmentation on the analytical signal to obtain real target domain crack damage guided wave monitoring data;

[0012] Step 4, establish a deep transfer learning model based on distribution distance metric and domain adversarial neural network, use the simulated source domain crack damage guided wave monitoring data and the real target domain crack damage guided wave monitoring data as inputs, and perform domain adversarial training on the deep transfer learning model, so that the classifier for the simulated source domain crack damage guided wave monitoring data in the deep transfer learning model can be directly transferred to the real target domain crack damage guided wave monitoring data.

[0013] In at least one embodiment of the present application, in Step 1, numerical crack damages with different lengths and orientation angles are introduced into the numerical simulation model for monitoring metal fatigue crack damage guided waves, and the dynamic display analysis algorithm is used to obtain numerical crack damage guided wave monitoring data.

[0014] In at least one embodiment of the present application, numerical crack damages with different lengths from 0 mm to 8 mm at a step size of 0.5 mm, different orientation angles from 80° to 100° at a step size of 5°, and a width of 0.05 mm are introduced into the numerical simulation model for monitoring metal fatigue crack damage guided waves.

[0015] In at least one embodiment of the present application, in Step 2, converting the numerical crack damage guided wave monitoring data into an analytical signal and performing data augmentation on the analytical signal includes:

[0016] Converting the numerical crack damage guided wave monitoring data into an analytical signal:

[0017] z(t) = A(t)e jφ(t)

[0018] Performing data augmentation on the analytical signal:

[0019] z1(t) = δA(t)ejβφ(t)

[0020] Among them, z(t) is the analytic signal of the numerical crack damage guided wave monitoring data, A(t) is the amplitude envelope of the analytic signal, Φ(t) is the instantaneous phase of the analytic signal, z1(t) is the analytic signal after data enhancement, δ is the amplitude envelope scaling coefficient of the numerical crack damage guided wave monitoring data, and β is the instantaneous phase scaling coefficient of the numerical crack damage guided wave monitoring data.

[0021] In at least one embodiment of the present application, in step three, the material of the real metal fatigue specimen is 7050 aluminum alloy, the size is 400mm×168mm×3mm, and a circular through hole with a diameter of 25mm is provided at the center position.

[0022] In at least one embodiment of the present application, in step four, the deep transfer learning model is:

[0023]

[0024] Among them, D s is the training set of the simulated source domain crack damage guided wave monitoring data, D t is the training set of the real target domain crack damage guided wave monitoring data, L f is the total loss function of the feature extractor, L MMD is the distribution distance metric between the source domain and the target domain, α is the hyperparameter of the distribution metric loss, λ is the weight of the domain discriminator loss L d , and L y is the loss function of the classifier.

[0025] In at least one embodiment of the present application, the weight λ of the domain discriminator loss L d is:

[0026] λ = 2 / (1 + e -10p ) - 1

[0027] Among them, p is the model training process.

[0028] The invention has at least the following beneficial technical effects:

[0029] The method for identifying metal crack damage guided waves based on numerical simulation and deep transfer in the present application uses numerical simulation to obtain the simulated crack guided wave response signal to solve the problem of insufficient damage annotation data in the engineering process; uses the domain adversarial neural network to minimize the domain distribution difference between the simulated source domain data and the real target data, extracts deep features with damage classification and domain invariance, realizes the domain adaptive transfer of numerical damage guided wave diagnosis knowledge to real damage, improves the reliability and application scope of the guided wave damage intelligent monitoring method, and provides accurate information input for the formulation of the metal structure maintenance guarantee plan. Description of the Drawings

[0030] Figure 1 is a flowchart of a method for identifying metal crack damage guided waves based on numerical simulation and deep transfer in an embodiment of the present application;

[0031] Figure 2 is a schematic diagram of the architecture of a deep transfer learning model based on distribution distance metric and domain adversarial neural network in an embodiment of the present application;

[0032] Figure 3 is a schematic diagram of a real metal fatigue specimen in an embodiment of the present application;

[0033] Figure 4 is a schematic diagram of the damage classification and identification result in an embodiment of the present application. Detailed Embodiment

[0034] To make the purpose, technical solutions and advantages of the implementation of the present application clearer, the technical solutions in the embodiments of the present application will be described in more detail below with reference to the accompanying drawings in the embodiments of the present application. In the drawings, the same or similar reference numerals denote the same or similar elements or elements with the same or similar functions throughout. The described embodiments are some, but not all, of the embodiments of the present application. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present application and should not be construed as limiting the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the scope of protection of the present application. The embodiments of the present application will be described in detail below with reference to the drawings.

[0035] In the description of the present application, it should be understood that the orientation or positional relationship indicated by the terms "center", "longitudinal", "lateral", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as limiting the scope of protection of the present application.

[0036] The following combines the attached Figures 1 to 4 to make a further detailed description of the present application.

[0037] The present application provides a method for identifying metal crack damage guided waves based on numerical simulation and deep transfer, as Figure 1 shown, including the following steps:

[0038] Step 1: Establish a simulation model for monitoring guided waves of metal fatigue crack damage and obtain numerical crack damage guided wave monitoring data;

[0039] Step 2: Convert the numerical crack damage guided wave monitoring data into an analytical signal, and perform data augmentation on the analytical signal to obtain the simulated source domain crack damage guided wave monitoring data;

[0040] Step 3: Conduct fatigue tests on multiple real metal fatigue specimens to obtain real crack damage guided wave monitoring data, convert the real crack damage guided wave monitoring data into an analytical signal, and perform data augmentation on the analytical signal to obtain the real target domain crack damage guided wave monitoring data;

[0041] Step 4: Establish a deep transfer learning model based on distribution distance metric and domain adversarial neural network. Use the simulated source domain crack damage guided wave monitoring data and the real target domain crack damage guided wave monitoring data as inputs, and perform domain adversarial training on the deep transfer learning model, so that the classifier for the simulated source domain crack damage guided wave monitoring data in the deep transfer learning model can be directly transferred to the real target domain crack damage guided wave monitoring data.

[0042] For the method for identifying metal crack damage guided waves based on numerical simulation and deep transfer in this application, first in Step 1, establish a simulation model for monitoring metal fatigue crack damage guided waves. By introducing numerical crack damages with different lengths and orientation angles in the simulation model for monitoring metal fatigue crack damage guided waves, use the dynamic display analysis algorithm to obtain the numerical crack damage guided wave monitoring data.

[0043] In Step 2, convert the numerical crack damage guided wave monitoring data into the form of an analytical signal (as shown in Equation 1), and perform data augmentation on the analytical signal to obtain the enhanced analytical signal. Linearly change its amplitude envelope and instantaneous phase (as shown in Equation 2) to simulate the signal fluctuations introduced by uncertain factors such as the material of the real specimen, manufacturing process, sensor installation process, and environmental temperature changes.

[0044] z(t) = A(t)e jφ(t) (1)

[0045] z1(t) = δA(t)e jβφ(t) (2)

[0046] Among them, z(t) is the analytical signal of the numerical crack damage guided wave monitoring data, A(t) is the amplitude envelope of the analytical signal, Φ(t) is the instantaneous phase of the analytical signal, z1(t) is the analytical signal after data augmentation, δ is the amplitude envelope scaling coefficient of the numerical crack damage guided wave monitoring data, and β is the instantaneous phase scaling coefficient of the numerical crack damage guided wave monitoring data.

[0047] Furthermore, in Step 3, fatigue tests are carried out on multiple real metal fatigue specimens to obtain real crack damage guided wave monitoring data, which are converted into the form of analytic signals in the manner of Step 2, and their amplitude envelopes and instantaneous phases are linearly changed. Data augmentation techniques are used to increase and balance the fatigue experiment monitoring data samples to improve the classification performance of the damage identification model.

[0048] Finally, in Step 4, a deep transfer learning model architecture based on distribution distance metric and domain adversarial neural network is built, as Figure 2 shown, taking the simulated source domain crack damage guided wave monitoring data and the real target domain crack damage guided wave monitoring data as inputs, integrating the distribution distance metric loss into the adversarial training of the domain adversarial neural network (as shown in Equation 3), iteratively optimizing the network architectures and parameters of the feature extractor, classifier and domain discriminator of the deep transfer model, driving the feature extractor to learn discriminative and domain-invariant high-dimensional features, so that the damage classification model for the simulated source domain crack damage guided wave monitoring data can be directly used for the real target domain.

[0049]

[0050] where D s is the training set of the simulated source domain crack damage guided wave monitoring data, D t is the training set of the real target domain crack damage guided wave monitoring data, L f is the total loss function of the feature extractor, L MMD is the distribution distance metric between the source domain and the target domain, α is the hyperparameter of the distribution metric loss, λ is the weight of the domain discriminator loss L d and L y is the loss function of the classifier.

[0051] In an embodiment of the present application, guided wave damage identification is carried out on the fatigue crack at the hole edge of an aluminum alloy center plate hole test piece. The material of the aluminum alloy center plate hole test piece is 7050 aluminum alloy, with dimensions of 400mm×168mm×3mm and a circular through hole with a diameter of 25mm at the center position. An MTS type fatigue testing machine is used to carry out the fatigue test. The test load spectrum is a sine wave, the maximum load is 40kN, the stress ratio is 0.1, and the loading frequency is 8Hz. During the test, 2 P-51 piezoelectric sensors are glued to the surface of the test piece with 401 glue. The diameter of the sensor is 8mm and the thickness is 0.45mm, forming 2 scanning channels. The A1-S1 monitoring channel is used to monitor the crack propagation on the left side, and the A2-S2 monitoring channel is used to monitor the crack propagation on the right side. The excitation signal of the guided wave is a five-peak sine excitation signal with a center frequency of 230kHz, the sampling frequency of the signal is 10MHz, and the sampling length is 4000 data points. The metal open-hole test piece and the sensor monitoring scheme are as Figure 3as shown

[0052] For crack damage identification using the method for identifying metal crack damage guided waves based on numerical simulation and deep transfer of this application, the specific process is as follows:

[0053] First, numerical crack damage with different lengths ranging from 0 mm to 8 mm in steps of 0.5 mm, different direction angles ranging from 80° to 100° in steps of 5°, and a width of 0.05 mm is introduced into the metal open-hole finite element simulation model to obtain numerical crack damage guided wave monitoring data under various numerical damage states;

[0054] Furthermore, a data augmentation method that converts to an analytic signal is used to augment the numerical crack damage guided wave monitoring data, and a numerical damage guided wave monitoring source domain dataset is established.

[0055] Then, for the experimental crack damage guided wave monitoring data of multiple center-hole metal specimens with fatigue cracks, the data augmentation technique is also used to increase the amplitude and phase fluctuations of the response signal to expand the real crack damage guided wave monitoring target domain dataset.

[0056] Finally, the hyperparameter α of the distribution metric loss is selected as 5, and λ is a variable parameter that changes with the model training process p (p linearly changes from 0 to 1 according to the model iterative training). A deep transfer learning model of the distribution distance metric and domain adversarial neural network is established for model iterative training, and the damage classification and identification results of the deep transfer learning model on the real damage guided wave monitoring target domain test dataset are obtained, as -10p shown Figure 4 as shown

[0057] The method for identifying metal crack damage guided waves based on numerical simulation and deep transfer of this application is mainly used to monitor the presence or absence of metal fatigue crack damage, and has the following beneficial effects:

[0058] (1) Based on numerical simulation means, a large number of simulation damage guided wave monitoring signals under different health states can be obtained, effectively solving the problems of long acquisition cycle and high test cost of a large number of well-labeled data in the actual engineering application process;

[0059] (2) A deep transfer learning model based on the distribution distance metric and domain adversarial neural network is established, effectively reducing the distribution difference between the simulation source domain dataset and the real target domain dataset, enabling the classifier model trained on the labeled simulation source domain dataset to be directly transferred to the unlabeled real target domain dataset;

[0060] (3) The proposed structural damage identification model based on deep transfer learning has high classification accuracy, low false alarm rate and missed alarm rate, and can be used as an unsupervised deep transfer identification method for the automated and intelligent monitoring of metal structure fatigue cracks.

[0061] As described above, it is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in this application should be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claimed rights.

Claims

1. A metal crack damage guided wave identification method based on numerical simulation and depth migration, characterized in that: include: Step 1: Establish a metal fatigue crack damage guided wave monitoring simulation model to obtain numerical crack damage guided wave monitoring data; Step 2: converting the numerical crack damage guided wave monitoring data into an analytical signal, and performing data enhancement on the analytical signal to obtain simulated source domain crack damage guided wave monitoring data; Step 3: Conduct fatigue tests on multiple real metal fatigue specimens to obtain real crack damage guided wave monitoring data, convert the real crack damage guided wave monitoring data into analytical signals, and perform data enhancement on the analytical signals to obtain real target domain crack damage guided wave monitoring data; Step 4: Establish a deep transfer learning model based on distribution distance metric and domain adversarial neural network, take the simulated source domain crack damage waveguide monitoring data and the real target domain crack damage waveguide monitoring data as input, and perform domain adversarial training on the deep transfer learning model, so that the classifier in the deep transfer learning model for the simulated source domain crack damage waveguide monitoring data can be directly migrated to the real target domain crack damage waveguide monitoring data.

2. The metal crack damage guided wave identification method based on numerical simulation and depth migration according to claim 1 is characterized in that: In step one, numerical crack damages of different lengths and orientation angles are introduced into the metal fatigue crack damage guided wave monitoring simulation model, and a dynamic display analysis algorithm is used to obtain numerical crack damage guided wave monitoring data.

3. The metal crack damage guided wave identification method based on numerical simulation and depth migration according to claim 2 is characterized in that: In the metal fatigue crack damage guided wave monitoring simulation model, numerical crack damage with different lengths varying from 0 mm to 8 mm in steps of 0.5 mm, different direction angles varying from 80° to 100° in steps of 5°, and a width of 0.05 mm is introduced.

4. The metal crack damage guided wave identification method based on numerical simulation and depth migration according to claim 3 is characterized in that: In step 2, the numerical crack damage guided wave monitoring data is converted into an analytical signal, and data enhancement is performed on the analytical signal, including: The numerical crack damage guided wave monitoring data is converted into analytical signals: z(t)=A(t)e jφ(t) Perform data enhancement on the analytical signal: z1(t)=δA(t)e jβφ(t) Among them, z(t) is the analytical signal of the numerical crack damage guided wave monitoring data, A(t) is the amplitude envelope of the analytical signal, Φ(t) is the instantaneous phase of the analytical signal, z1(t) is the analytical signal after data enhancement, δ is the amplitude envelope scaling factor of the numerical crack damage guided wave monitoring data, and β is the instantaneous phase scaling factor of the numerical crack damage guided wave monitoring data.

5. The metal crack damage guided wave identification method based on numerical simulation and depth migration according to claim 4 is characterized in that: In step three, the material of the real metal fatigue test piece is 7050 aluminum alloy, the size is 400mm×168mm×3mm, and a circular through hole with a diameter of 25mm is opened at the center.

6. The metal crack damage guided wave identification method based on numerical simulation and depth migration according to claim 5 is characterized in that: In step 4, the deep transfer learning model is: Among them, D s D is the training set of guided wave monitoring data for crack damage in the simulation source domain. t is the real target domain crack damage guided wave monitoring data training set, L f is the total loss function of the feature extractor, L MMD is the distribution distance metric between the source domain and the target domain, α is the hyperparameter of the distribution metric loss, and λ is the domain discriminator loss L d The weight, L y is the loss function of the classifier.

7. The metal crack damage guided wave identification method based on numerical simulation and depth migration according to claim 6 is characterized in that: Domain discriminator loss L d The weight λ is: λ=2 / (1+e -10p )-1 Among them, p is the model training process.

Images