Nonlinear ultrasonic guided wave crack location method based on relative value and data driven

Abstract

The present invention relates to a nonlinear ultrasonic guided wave crack location method based on relative value and data drive, comprising the following steps: collecting the response signal of the measured object under the excitation signal, performing wavelet transform on the response signal, extracting the corresponding nonlinear feature, collecting the location prediction model to obtain the corresponding relative value of the crack position, so as to locate the crack of the measured object; wherein the location prediction model adopts a one-dimensional convolutional neural network with an attention mechanism, and the training process comprises: establishing a finite element simulation model of a plate-like structure with crack damage in a polar coordinate system, obtaining the off-plane displacement data at the effective node, performing wavelet transform on the off-plane displacement data, and establishing a nonlinear feature data set; training the location prediction model based on the nonlinear feature data set to characterize the mapping relationship between the nonlinear feature and the relative value of the crack position. Compared with the prior art, the present invention can simply and conveniently establish a data set to achieve accurate location of cracks in a plate-like structure.

Description

Technical Field

[0001] The invention relates to the problem of crack positioning in a plate-like structure, and in particular to a nonlinear ultrasonic guided wave crack positioning method based on relative value and data driving. Background Art

[0002] Fatigue cracks are a common form of damage in plate structures. Their initiation and expansion stages account for more than 70% of the life of the structure. However, they are not easy to be detected before they expand into macro cracks, which can easily cause safety accidents. Ultrasonic guided wave detection is an important means of health detection of plate structures. It has been widely studied and explored in the detection of large-area thin plate conditions. However, traditional linear ultrasonic guided wave detection methods cannot detect fatigue cracks. In recent years, studies have found that the detection and characterization of material performance degradation are closely related to ultrasonic nonlinear effects. The development of nonlinear ultrasonic theory makes fatigue crack detection in plate structures possible. Due to the complex propagation mechanism of guided waves, weak characteristics of nonlinear guided wave signals, and multimodal and dispersion characteristics of guided wave signals, it is difficult to accurately locate cracks in plate structures. To solve the above problems, from the perspective of data, data-driven methods can be used to analyze guided wave signals with strong feature extraction and logical judgment in high-dimensional space. However, the existing research on damage location based on data-driven methods belongs to the category of linear guided waves, and there is a gap in the research on nonlinear guided wave location.

[0003] There is a problem in the study of nonlinear waveguide positioning of fatigue cracks based on data-driven methods: it is difficult to obtain sufficient data for neural network training. Sufficient data volume is the key to determining the prediction performance of neural networks. At present, there are two main ways to obtain data in linear waveguide positioning research: experimental method, processing damaged specimens at different positions, and conducting experimental measurements to obtain data. The amount of data obtained is related to the number of specimens processed. The other is the finite element simulation method. A finite element model similar to the actual working condition is established in the simulation software, and the absolute position of the damage in the model is constantly changed to obtain a large amount of simulation data. Because it takes a long time, is costly and difficult to process fatigue cracks in plate structures, it is not feasible to use experimental methods to obtain data. The traditional finite element simulation method is to establish a simulation model by changing the absolute position of the damage in the plate, and collect the off-plane displacement data at the fixed node. The amount of data depends on the number of models established. The training cycle of the finite element model is long and requires high computer computing performance. Therefore, establishing a training data set through the traditional finite element simulation method requires strong computer performance and high time cost. Summary of the invention

[0004] The purpose of the present invention is to overcome the difficulty in locating cracks in plate structures and the difficulty in obtaining sufficient data in data-driven methods, and to provide a nonlinear ultrasonic guided wave crack locating method based on relative values ​​and data-driven. The present invention can be used to simply and conveniently establish a data set, and based on the established nonlinear characteristic data set and neural network model, accurate positioning of cracks in plate structures can be achieved.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A nonlinear ultrasonic guided wave crack location method based on relative value and data drive comprises the following steps:

[0007] Collecting the response signal of the measured object under the excitation signal, performing wavelet transform on the response signal, and extracting the corresponding nonlinear characteristics;

[0008] Based on the nonlinear characteristics, a pre-trained positioning prediction model is collected to obtain a corresponding relative value of a crack position, so as to locate the crack of the object under test;

[0009] The positioning prediction model adopts a one-dimensional convolutional neural network with an attention mechanism, and the process of training the positioning prediction model includes:

[0010] A finite element simulation model of a plate structure with crack damage is established in a polar coordinate system, the off-plane displacement data at the effective nodes are obtained, the off-plane displacement data are subjected to wavelet transformation, and a nonlinear characteristic data set is established;

[0011] A positioning prediction model is trained based on a nonlinear feature data set, and the positioning prediction model represents the mapping relationship between the nonlinear feature and the relative value of the crack position.

[0012] Furthermore, the finite element simulation model includes an excitation array element and a crack damage unit. The excitation array element is used as the polar coordinate origin. Under the condition that the distance between the crack damage unit and the origin remains unchanged, the polar angle value of the crack damage unit is continuously changed to establish multiple simulation models, thereby obtaining off-surface displacement data at effective nodes.

[0013] Further, the valid nodes are determined based on the data set establishment requirements.

[0014] Furthermore, the effective nodes are arranged in a square array, the number of array elements and the array element spacing in the square array are adjustable variables, and off-plane displacement data sets with different numbers of array elements and different array element spacings are constructed.

[0015] Furthermore, when training the positioning prediction model based on the nonlinear feature data set, the order of samples in the data set is disrupted and then divided into a training set, a test set, and a validation set.

[0016] Furthermore, the one-dimensional convolutional neural network includes a preliminary feature extraction module and a feature enhancement extraction module, both of which have an attention mechanism.

[0017] Furthermore, the relative value of the crack position includes a relative length value and a relative angle value.

[0018] Furthermore, the excitation signal is a sine wave time domain pulse signal modulated by a 10-cycle Hanning window.

[0019] Furthermore, the wavelet transform is implemented using Morlet wavelet basis function.

[0020] The present invention also provides a computer-readable storage medium, comprising one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs include instructions for executing the nonlinear ultrasonic guided wave crack location method based on relative values ​​and data-driven as described above.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1. The present invention proposes for the first time to use a data-driven method and nonlinear ultrasonic guided wave signal characteristics to realize crack location in a plate structure, which can simply and conveniently establish a data set and accurately locate the crack position;

[0023] 2. Compared with the modeling method of changing the absolute position of the damage, the introduction of the relative value concept of the present invention greatly reduces the cost of creating a data set. At the same time, under the relative value concept, the characteristics of the signals received by each node in different models are highly similar and can be used interchangeably. That is, after the introduction of the relative value concept, the data has wide applicability, thereby improving the wide application of the neural network model trained by using simulation data;

[0024] 3. The present invention adopts a one-dimensional convolutional neural network with an attention mechanism to construct a positioning prediction model, which can enhance the extraction of non-sensitive features by the neural network and improve the accuracy of crack position prediction;

[0025] 4. Provide a new research idea for solving the damage location problem of large-area plate structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the process of the present invention;

[0027] Figure 2 is a schematic diagram of the finite element model;

[0028] Figure 3 Schematic diagram of 1D-Attention-CNN model;

[0029] Figure 4This is a comparison chart of signal nonlinear characteristics under the same relative value label conditions of different models;

[0030] Figure 5 This is a schematic diagram of crack location results;

[0031] Figure 6 is the prediction result of the test set. DETAILED DESCRIPTION

[0032] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0033] The present invention provides a nonlinear ultrasonic guided wave crack location method based on relative value and data drive, comprising the following steps: collecting the response signal of the measured object under the excitation signal, performing wavelet transform on the response signal, and extracting the corresponding nonlinear features; based on the nonlinear features, collecting the pre-trained location prediction model to obtain the corresponding relative value of the crack position to locate the crack of the measured object. Among them, the location prediction model adopts a one-dimensional convolutional neural network with an attention mechanism, and the process of training the location prediction model includes: establishing a finite element simulation model of a plate structure with crack damage in a polar coordinate system, obtaining the off-plane displacement data at the effective node, performing wavelet transform on the off-plane displacement data, and establishing a nonlinear feature data set; training the location prediction model based on the nonlinear feature data set, and the location prediction model characterizes the mapping relationship between the nonlinear feature and the relative value of the crack position.

[0034] The above method uses a data-driven method and nonlinear ultrasonic guided wave signal characteristics to realize crack location in a plate structure, which can simply and conveniently establish a data set and accurately locate the crack position. At the same time, the present invention introduces the concept of relative value. When two models are different, but the damage "relative value" label is the same, it can be considered that the damage of the two models is in the same position, and the characteristics of the received signals of each node in the model are highly similar and can be used interchangeably. That is, after the introduction of the concept of relative value, the data has wide applicability, which greatly reduces the cost of creating a data set.

[0035] like Figure 1 As shown, in this method, the construction and training process of the positioning prediction model is as follows.

[0036] 1. Finite element model construction

[0037] (1) Model structure parameter setting

[0038] The aluminum plate (300×200×2mm) created using Abaqus / EXPLICIT software 3 , density: 2750kg / m3 , Young's modulus: 74919.65Mpa, Poisson's ratio: 0.331) finite element model, as shown in the schematic diagram Figure 2 As shown in the figure, material nonlinearity and motion nonlinearity are not considered in the simulation process. A crack damage unit (10×15mm 2 ), which is designed according to the working conditions where cracks are prone to occur at bolted connections. It consists of a circular through hole with a diameter of 5 mm and a crack with a length of 7.5 mm, in which the crack surface is set to be hard contact, frictionless and openable / closed.

[0039] (2) Grid parameter settings

[0040] Taking into account the calculation accuracy and efficiency of the model, the unit size is set to 1 mm during meshing, the mesh of the crack damage unit is refined, the crack is locally meshed, and the crack is equally divided into 60 nodes.

[0041] (3) Excitation signal parameter setting

[0042] In this simulation, the excitation array element is single and fixed in position. The off-plane displacement signal is symmetrically loaded at the position of the excitation array element. The excitation signal is a sinusoidal wave time domain pulse signal modulated by a 10-cycle Hanning window. The displacement signal amplitude is 1e-6m and the excitation frequency is set to 280kHz.

[0043] (4) Complete data acquisition

[0044] The number of excitation array elements in the model is unique and their positions are fixed, and they are set as the origin of the polar coordinate system. The polar coordinates of the crack damage unit are (m, θ). In the simulation, the absolute distance between the crack damage unit and the excitation array element remains unchanged (40 mm), that is, m = 40. At this time, θ becomes the only variable that changes the position of the crack damage unit. In order to obtain a complete data set, it is only necessary to establish a simulation model by changing θ (θ∈[25°,80°], with a change interval of 5°), and a total of 12 simulation models are established. The data set obtained in this way is similar to the data set obtained in the Cartesian coordinate system, with the array element position fixed and the absolute position of the crack damage unit changed.

[0045] This embodiment collects response signals by receiving array elements. The relative value concept requires that the simulation model be established in a polar coordinate system, and the position of the excitation array elements is fixed and the number is unique. Taking the excitation array element as the coordinate origin (pole), while ensuring that the distance between the crack damage unit and the coordinate origin remains unchanged, the polar angle value of the crack damage unit is continuously changed, multiple simulation models are established, and the off-plane displacement data at the effective nodes (i.e., effective array elements) are obtained. The off-plane displacement data at the effective array element refers to the off-plane displacement data of all nodes collected in a single simulation model. The effective nodes are subsequently determined according to the requirements for data set establishment, and off-plane displacement data sets with different numbers of array elements and different array element spacings are constructed. The requirements for data set establishment include two parts: square array mode requirements and array element number requirements. In the square array mode, the array element spacing is equal, and the number of array elements can be limited to 8, 12, 16, 20 and other situations.

[0046] 2. Data processing

[0047] (1) Nonlinear feature extraction

[0048] In order to obtain the second harmonic characteristic components in the response signal, the wavelet transform is used to perform time-frequency analysis on the response signal, and the wavelet coefficients near the double frequency (560kHz) are obtained for subsequent neural network model training.

[0049] Assume that the signal to be analyzed f(t)∈L 2 (R), the wavelet sequence can be obtained by scaling and translating the wavelet basis function ψ(t). The wavelet transform of f(t) is defined as:

[0050]

[0051] Among them, a is the scaling factor, b is the translation factor, and WT(a,b) represents the similarity between f(t) and the wavelet basis function ψ(t) in different scale spaces.

[0052] The excitation signal is a sine wave modulated by a 10-cycle Hanning window. Compared with the commonly used wavelet basis functions, the Morlet wavelet basis function is selected to extract the nonlinear characteristics of the waveguide signal.

[0053] (2) Creation of nonlinear feature dataset

[0054] This embodiment uses a square array of 16 array elements to obtain the off-plane displacement data of the node where the receiving array element is located at equal intervals. The array element spacing L increases from 5mm to 56mm (1mm step), and finally constructs 52 kinds of array element spacings. The original waveguide time domain signal data set is constructed. On this basis, the wavelet coefficients at the double frequency of the time domain signal are calculated by wavelet transform, and the nonlinear feature data set corresponding to the original waveguide time domain signal data set is constructed and used as the input of the neural network.

[0055] (3) Relative value label creation

[0056] The present invention transforms the crack location problem into a regression problem. Different from the existing research, the present invention establishes "relative value" labels: angle relative value labels (i.e. angle relative value, angle relative)) and distance relative value labels (i.e. length relative value, length relative). If the location of the excitation array element is the origin of the polar coordinate system, and the polar coordinates of the crack damage unit are (θ, m), then the angle relative value label is θ / 90°, and the length relative value label is m / M, where m represents the absolute distance between the crack damage unit and the excitation array element (coordinate origin), and M represents the side length of the square detection area.

[0057] The present invention proposes that: when two models are different but the damage "relative value" labels are the same, it can be considered that the damage of the two models is in the same position, and the characteristics of the received signals of each node in the model are highly similar and can be used interchangeably, that is, after the introduction of the relative value concept, the data has wide applicability.

[0058] To prove the above point, when the angle relative value label and the distance relative value label are the same, the detection area (i.e., the side length of the square array) is changed, and two simulation models A and B are established. The polar coordinates of the crack damage unit are set to (40°, 40) and (40°, 80), respectively, and the detection area of ​​model A is set to 40×40mm 2 , the detection area of ​​model B is set to 80×80mm 2 , that is, the relative value labels of the two models are consistent. Figure 4 (a) and (b) in the figure represent the nonlinear characteristics of the received signals of the two array elements under model A. Figure 4 (c) and (d) in the figure represent the nonlinear characteristics of the received signals of the two array elements under model B respectively.

[0059] In horizontal comparison, Point-A1 and Point-A2 are at different receiving positions, and the nonlinear characteristics of the two signals are different. Therefore, the nonlinear characteristics of the signals received at different positions under the same model are different. In vertical comparison, although Point-A1 and Point-B1 are at different positions, they belong to the same type of receiving position in the relative value category, and the characteristics of the two signals are similar, with only differences in amplitude. The comparison results between Point-A2 and Point-B2 are also the same. Therefore, when the relative label value is the same, the nonlinear characteristics of the signals received at the same position under different models are similar. Therefore, after the introduction of the "relative value" label, if you want to obtain a large amount of data, you only need to change the polar coordinates θ of the crack damage unit of the simulation model in the polar coordinate system, and the obtained data can be used for crack location analysis when the detection area is different. In addition, in actual detection, the area of ​​the detection area is different from the location of the crack damage unit. When using conventional simulation methods and the "absolute value" label, the simulation data cannot cover the actual situation, which further illustrates the necessity of introducing the "relative value" label.

[0060] (4) Dataset Division

[0061] In order to improve the accuracy of the model and the convergence speed of model training, the data is standardized and converted into dimensionless values ​​that conform to the N(0,1) normal distribution. In addition, because the data fitting ability of the neural network is too strong, if samples with the same arrangement pattern appear repeatedly in a batch, the neural network will learn the arrangement pattern as a feature, which is prone to overfitting and reduces the generalization ability of the model. Therefore, it is necessary to disrupt the order of samples in the data set. Finally, the data set is divided into two parts: 90% is the training set and 10% is the test set. In addition, 20% of the training set is used as a validation set.

[0062] 3. 1D-Attention-CNN model training

[0063] The 1D-Attention-CNN structure proposed by the present invention for the one-dimensional waveguide signal feature extraction in the plate-like structure is as follows Figure 3 As shown in the figure, it has a preliminary feature extraction module and a feature enhancement extraction module, both of which have an attention mechanism, and is mainly composed of a data input layer, a one-dimensional convolution layer, a pooling layer, an attention mechanism layer, an FC layer, an output layer, etc. The model takes the waveguide nonlinear feature data as input and the relative value of the crack position as output.

[0064] First, the wavelet coefficients near the second frequency of the waveguide signal are used as the input of the neural network. After two convolution and pooling calculations, the attention mechanism layer is introduced to establish the connection between each feature space and each feature extraction channel to complete the preliminary feature extraction. Among them, the attention mechanism layer includes channel attention units and spatial attention units. Then, the convolution layer, pooling layer and attention mechanism layer are used to extract features again to further enhance the extraction of non-sensitive features by the neural network. The extracted features are calculated in turn by three alternating FC layers (64, 32, 16) and Dropout layers (output), and finally the polar angle value of the crack and the relative value of the distance are output to determine the crack location. During the research process, the relevant parameters were adjusted through repeated experiments, and the final model parameters are as follows Figure 3 In addition, the ReLU activation function is used to realize the nonlinear mapping of features. The RMSProp optimization algorithm is selected in the network training process, the learning rate is set to 0.001, the loss function is selected as MSE, and the number of iterations is set to 100.

[0065] 4. Model prediction

[0066] After the neural network model training is completed, the crack model test set data with θ = 40° and L = 30 mm is predicted. The results are as follows Figure 5 As shown in Figure 2, the predicted position of the crack tip is close to the actual position. Figure 6 The predicted results of the angle and array element spacing of the test set data. The predicted angle values ​​and the predicted array element spacing values ​​are orderly distributed on both sides of the diagonal, which shows that there is a high consistency between the predicted values ​​and the true values, and the prediction results are accurate.

[0067] After obtaining a positioning prediction model that has been trained and can be used, the object under test that needs crack location can be stimulated, and the corresponding response signal can be collected. Nonlinear features are extracted from the response signal and input into the positioning prediction model. Crack location can be achieved based on the mapping relationship between the nonlinear feature data and the relative value of the crack position.

[0068] If the above method is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0069] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A nonlinear ultrasonic guided wave crack location method based on relative value and data drive, characterized in that: The following steps are involved: Collecting the response signal of the measured object under the excitation signal, performing wavelet transform on the response signal, and extracting the corresponding nonlinear characteristics; Based on the nonlinear characteristics, a pre-trained positioning prediction model is collected to obtain a corresponding relative value of a crack position, so as to locate the crack of the object under test; The positioning prediction model adopts a one-dimensional convolutional neural network with an attention mechanism, and the process of training the positioning prediction model includes: A finite element simulation model of a plate structure with crack damage is established in a polar coordinate system, the off-plane displacement data at the effective nodes are obtained, the off-plane displacement data are subjected to wavelet transformation, and a nonlinear characteristic data set is established; A positioning prediction model is trained based on a nonlinear feature data set, and the positioning prediction model represents the mapping relationship between the nonlinear feature and the relative value of the crack position.

2. The nonlinear ultrasonic guided wave crack location method based on relative value and data drive according to claim 1 is characterized in that: The finite element simulation model includes an excitation array element and a crack damage unit. The excitation array element is used as the polar coordinate origin. Under the condition that the distance between the crack damage unit and the origin remains unchanged, the polar angle value of the crack damage unit is continuously changed to establish multiple simulation models, thereby obtaining off-plane displacement data at effective nodes.

3. The nonlinear ultrasonic guided wave crack location method based on relative value and data drive according to claim 2 is characterized in that: The valid nodes are determined based on the data set establishment requirements.

4. The nonlinear ultrasonic guided wave crack location method based on relative value and data drive according to claim 2 is characterized in that: The effective nodes are arranged in a square array, the number of array elements and the array element spacing in the square array are adjustable variables, and off-plane displacement data sets with different numbers of array elements and different array element spacings are constructed.

5. The nonlinear ultrasonic guided wave crack location method based on relative value and data drive according to claim 1, characterized in that: When training a positioning prediction model based on a nonlinear feature data set, the order of samples in the data set is disrupted and then divided into a training set, a test set, and a validation set.

6. The nonlinear ultrasonic guided wave crack location method based on relative value and data drive according to claim 1 is characterized in that: The one-dimensional convolutional neural network includes a preliminary feature extraction module and a feature enhancement extraction module, both of which have an attention mechanism.

7. The nonlinear ultrasonic guided wave crack location method based on relative value and data drive according to claim 1 is characterized in that: The relative value of the crack position includes a relative length value and a relative angle value.

8. The nonlinear ultrasonic guided wave crack location method based on relative value and data drive according to claim 1, characterized in that: The excitation signal is a 10-cycle Hanning window modulated sinusoidal wave time domain pulse signal.

9. The nonlinear ultrasonic guided wave crack location method based on relative value and data drive according to claim 1, characterized in that: The wavelet transform is implemented using Morlet wavelet basis function.

10. A computer-readable storage medium, characterized in that: It comprises one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs comprise instructions for executing the nonlinear ultrasonic guided wave crack location method based on relative value and data-driven as claimed in any one of claims 1 to 9.

Images