A method for ultrasonic guided wave defect detection and localization based on singular value decomposition

Abstract

The present invention relates to a method for ultrasonic guided wave defect detection and location based on singular value decomposition, belonging to the field of ultrasonic guided wave detection technology. The method comprises the following steps: S1: acquiring a waveguide signal set matrix; S2: obtaining the waveguide signal defect information index TRDI; S3: building a convolutional neural network (CNN) model, training the CNN model using the waveguide signal defect information index TRDI, and establishing a nonlinear function mapping relationship between the waveguide signal defect information index TRDI and the defect location; and S4: using the trained CNN model to detect and locate defects in the entire composite material. By incorporating the waveguide signal defect information index TRDI into the neural network calculation, the method eliminates the effects of noise and temperature on the detection results.

Description

Technical Field

[0001] The present invention belongs to the technical field of ultrasonic guided wave detection, and in particular relates to a method for ultrasonic guided wave defect detection and positioning based on singular value decomposition. Background Art

[0002] Composite materials, with their high specific strength and stiffness, are widely used in various advanced engineering fields. Many aerospace and engineering equipment operate in harsh environments, subjecting them to uncontrollable environmental factors and causing structural defects such as wear and fatigue, which can be very harmful to the equipment. Therefore, research into non-destructive inspection and assessment techniques for composite materials to predict potential hazards is crucial for ensuring the economical and reliable operation of aerospace and engineering equipment.

[0003] The practical application environment of ultrasonic nondestructive testing (NDT) technology is often highly variable. Performing NDT on large, in-service engineering equipment in environments subject to temperature fluctuations and high noise levels is extremely challenging. Environmental factors such as temperature and noise are significant factors influencing the practical application of NDT. These factors can obscure useful information in ultrasonic signals, rendering NDT results unreliable or even erroneous.

[0004] Traditionally, temperature compensation techniques have been used to address temperature effects. These techniques use a set of waveguide signals at different temperatures as a baseline signal and then use signal processing to compensate for temperature-induced waveform variations. However, these temperature compensation methods only approximate temperature effects and fail to eliminate the impact of noise, making them unsuitable for the unpredictable and changing environments of real-world applications. Summary of the Invention

[0005] In order to overcome the deficiencies in the prior art, the present invention proposes a real-time detection system for composite material defects using ultrasonic guided waves, which aims to solve the technical problem of how to eliminate the influence of noise and temperature on the detection results.

[0006] To achieve the above objectives, the present invention provides a method for ultrasonic guided wave defect detection and positioning based on singular value decomposition, comprising the following steps: S1: collecting a waveguide signal set matrix; S2: obtaining a waveguide signal defect information index TRDI; S3: building a convolutional neural network (CNN) model, training the convolutional neural network (CNN) model using the waveguide signal defect information index TRDI, and establishing a nonlinear function mapping relationship between the waveguide signal defect information index TRDI and the defect position; and S4: using the convolutional neural network (CNN) model to perform defect detection and positioning on the entire composite material.

[0007] Furthermore, in S2, a rectangular window function is used to slide and divide the waveguide signal set matrix into short time domain segments. The width of the rectangular window is set to be the same as the length of the excitation signal generated by the waveform generator. The rectangular window is slid along the entire signal matrix to decompose the waveguide signal into different short time domain segments.

[0008] Furthermore, in step S2, a rectangular window function with the same width as the excitation signal is selected to decompose the waveguide signal set matrix:

[0009]

[0010] X W d represents the decomposed short time domain segment, w(d) is the rectangular window function, Δt represents the width of the rectangular window function, x N*D represents the waveguide signal set matrix, where:

[0011] x N*D =[x1, x2, ..., x N ] T , (2)

[0012] D represents the signal sampling length, N represents the number of waveguide time domain signals in the waveguide signal set, x represents a single signal, and T represents the transposed matrix;

[0013] For the decomposed short time domain segments, perform singular value decomposition SVD:

[0014]

[0015] cov represents covariance processing, the superscript H represents conjugate transpose, U and ∑ are the eigenvector matrix and eigenvalue matrix after performing singular value decomposition (SVD), respectively;

[0016] Perform signal processing on the diagonal singular values of the matrix ∑:

[0017]

[0018] λ d represents the processed output value, λ n represents the singular value corresponding to the matrix ∑, λ1 represents the first singular value of the matrix ∑, and all output values form a specific time series related to the defect information, namely the waveguide signal defect information index TRDI.

[0019] Furthermore, in the S3, the three convolutional layers LReLU of the convolutional neural network CNN model are respectively set with convolution kernels of sizes 3×3, 5×5, and 7×7, and the numbers are 16, 32, and 64 respectively. The specific function processing feature of each layer is that the waveguide signal defect information index TRDI passes through the convolutional layer LReLU to perform vector convolution operation and convolution processing of the linear rectification function, and each feature map is nonlinearly processed by the linear rectification function.

[0020] Furthermore, in S4, the softmax function is used to perform defect detection and positioning operations, the cross entropy function is selected as the loss function of the convolutional neural network CNN model, the Adam algorithm is used to optimize the convolutional neural network CNN model with an adaptive learning rate, and the convolutional neural network CNN model is optimized by small-batch stochastic gradient descent based on the first-order gradient. The hyperparameters of the Adam algorithm are set to: α = 0.001, β1 = 0.9, β2 = 0.999. The Adam algorithm updates the learnable parameters of the entire convolutional neural network CNN model to obtain the optimal parameter settings.

[0021] Beneficial Effect: The extracted defect-related wave components of the waveguide signal are all scattered from the same damage source, resulting in strong correlation between the corresponding defect-related wave components. Signal variations caused by noise or temperature have little correlation. In other words, signal variations caused by noise or temperature have little impact on the defects of the waveguide signal. Therefore, the waveguide signal defect information index (TRDI) obtained using a specific formula retains useful time-domain defect-related information, eliminates the adverse effects of temperature and noise on the signal, and reduces the error rate of defect information. Note that the corresponding waveguide signal defect information index (TRDI) cannot be obtained using other formulas.

[0022] The guided wave signal defect information index (TRDI) helps speed up the training and fitting of convolutional neural network (CNN) models, completing an input and output test in milliseconds. This method improves the guided wave signal's immunity to temperature and noise interference, while also enhancing defect detection accuracy, enabling more accurate determination of defect locations and broadening the scope of ultrasonic guided wave nondestructive testing. This approach addresses the difficulty in extracting effective defect information from guided wave signals contaminated by temperature and heavy noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of ultrasonic guided wave defect detection and positioning based on singular value decomposition (SVD) in the present invention;

[0024] Figure 2 Schematic diagram of extracting the defect information index TRDI of the waveguide signal in the present invention;

[0025] Figure 3Schematic diagram of the convolutional neural network (CNN) model constructed in the present invention;

[0026] Figure 4 This is the convolutional neural network operation flow chart. DETAILED DESCRIPTION

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0028] The method for ultrasonic guided wave defect detection and location based on singular value decomposition in this embodiment, as shown in the figure, includes the following steps:

[0029] Step 1, collecting the waveguide signal set matrix;

[0030] Specifically, multiple piezoelectric transducers are arranged on a composite material board in a one-transmitter, one-receiver configuration. A waveform generator generates an excitation signal, which is then applied to each piezoelectric transducer element via a power amplifier. An oscilloscope then collects the response signals from the piezoelectric transducer elements. These response signals form a waveguide signal matrix.

[0031] Step 2: Obtain the waveguide signal defect information index TRDI.

[0032] Specifically, a rectangular window function is used to partition the waveguide signal matrix into short time-domain segments. The width of the rectangular window is set to the same length as the excitation signal generated by the waveform generator. The rectangular window is then slid across the entire signal matrix to decompose the waveguide signal into different short time-domain segments. Singular value decomposition is performed on these decomposed short time-domain segments to obtain the eigenvalue matrix, or singular values, of the waveguide signal matrix. Signal processing is performed on the extracted singular values to obtain the waveguide signal defect information index (TRDI).

[0033] See Figure 2 The specific extraction stage of the waveguide signal defect information index TRDI is as follows:

[0034] Select a rectangular window function with the same width as the excitation signal to decompose the waveguide signal set matrix:

[0035]

[0036] X W d represents the decomposed short time domain segment, w(d) is the rectangular window function, Δt represents the width of the rectangular window function, x N*D represents the waveguide signal set matrix, where:

[0037] X N*D =[x1, x2, ..., x N ] T , (2)

[0038] D represents the signal sampling length, N represents the number of waveguide time domain signals in the waveguide signal set, x represents a single signal, and T represents the transposed matrix;

[0039] For the decomposed short time domain segments, perform singular value decomposition SVD:

[0040]

[0041] cov represents covariance processing, the superscript H represents conjugate transpose, and U and ∑ are the eigenvector matrix and eigenvalue matrix after performing singular value decomposition (SVD), respectively.

[0042] Signal processing is performed on the singular values of the diagonal items of the matrix ∑.

[0043]

[0044] λ d represents the processed output value, λ n represents the singular value corresponding to the matrix ∑, λ1 represents the first singular value of the matrix ∑, and all output values form a specific time series related to the defect information, namely the waveguide signal defect information index TRDI.

[0045] Step 3: Build a convolutional neural network (CNN) model, train the convolutional neural network (CNN) model using the waveguide signal defect information index (TRDI), and establish a nonlinear function mapping relationship between the waveguide signal defect information index (TRDI) and the defect position.

[0046] In step 4, the trained convolutional neural network (CNN) model can be used to determine defects in the short time domain segments of all collected guided wave signals, and to detect and locate defects in the entire composite material.

[0047] See Figure 3 and Figure 4Specifically, the three convolutional layers LReLU of the convolutional neural network CNN model set convolution kernels of size 3×3, 5×5 and 7×7 respectively, and the number is 16, 32 and 64 respectively. The specific function processing feature of each layer is that the waveguide signal defect information index TRDI passes through the convolution layer LReLU, and performs vector convolution operation and convolution processing of the linear rectification function. Each feature map is nonlinearly processed by the linear rectification function. The convolution layer LReLU obtains the defect judgment condition after several convolutions of the waveguide signal defect information index TRDI, and makes a preliminary prediction on whether there are defects. After passing through the convolution layer LReLU, the waveguide signal defect information index TRDI enters the maximum pooling layer Max pooling for downsampling operation, outputs the feature Flatten, inputs the linear layer FCL, and the waveguide signal defect information index TRDI enters the last layer. After the softmax function operation, the input waveguide signal defect information index TRDI is calculated to determine whether it has defects and the defect location distribution. The specific deep learning structure is as follows: Figure 3 shown.

[0048] The cross-entropy function was selected as the loss function for the CNN model. The Adam algorithm was used to optimize the CNN model with an adaptive learning rate and mini-batch stochastic gradient descent based on first-order gradients. The hyperparameters of the Adam algorithm were set to: α = 0.001, β1 = 0.9, and β2 = 0.999.

[0049] Traditional detection technology uses a set of guided wave signals at different temperatures as a baseline signal and uses signal processing to compensate for temperature-induced waveform changes. However, no measures are taken to address noise, making it impossible to eliminate the effects of noise.

[0050] A typical one-dimensional convolutional neural network (CNN) model in the existing technology was used for training. The ambient temperature difference was 20℃-40℃, and there were several different levels of noise, such as 0dB, 5dB, 10dB, 15dB, and 20dB. After training, the model's verification accuracy was 34.72%, the test accuracy was 30.54%, the detection time was 391.38 seconds, and the number of iterations was 50.

[0051] Compared to the prior art, a convolutional neural network (CNN) model using singular value decomposition was trained, achieving 100% validation accuracy, 100% test accuracy, a detection time of 61.29 seconds, and 44 iterations. In this embodiment, the extracted defect-related wave components of the waveguide signal are all scattered from the same damage source, resulting in strong correlation between the corresponding defect-related wave components. Signal changes caused by noise or temperature have little correlation, meaning that signal changes caused by noise or temperature have little impact on the defects of the waveguide signal. Therefore, using a specific formula, the waveguide signal set matrix is decomposed into many short time-domain segments. The width of the rectangular window used in the decomposition is the same as the excitation signal length. Therefore, each short time-domain segment can be considered to cover most of the wave components, and each segment can only contain a single damage scattered signal from a single damage source. The extracted defect-related wave components of the waveguide signal are all scattered from the same damage source, resulting in strong correlation between the corresponding defect-related wave components. The resulting waveguide signal defect information index (TRDI) retains useful time-domain defect-related information, eliminates the adverse effects of temperature and noise on the signal, and reduces the error rate of defect information. Note that the corresponding waveguide signal defect information index TRDI cannot be obtained by using other formulas.

[0052] The guided wave signal defect information index (TRDI) helps speed up the training and fitting of convolutional neural network (CNN) models, completing an input and output test in milliseconds. This method improves the guided wave signal's immunity to temperature and noise interference, while also enhancing defect detection accuracy, enabling more accurate determination of defect locations and broadening the scope of ultrasonic guided wave nondestructive testing. This approach addresses the difficulty in extracting effective defect information from guided wave signals contaminated by temperature and heavy noise.

[0053] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. A method for ultrasonic guided wave defect detection and location based on singular value decomposition, characterized in that: The following steps are included: S1: Acquires the waveguide signal set matrix; S2: Get the waveguide signal defect information index TRDI; In S2, a rectangular window function with the same width as the excitation signal is selected to decompose the waveguide signal set matrix: X W d represents the decomposed short time domain segment, w(d) is the rectangular window function, Δt represents the width of the rectangular window function, X N*D represents the waveguide signal set matrix, where: X N*D =[x1,x2,…,x N ] T , (2) D represents the signal sampling length, N represents the number of waveguide time domain signals in the waveguide signal set, x represents a single signal, and T represents the transposed matrix; For the decomposed short time domain segments, perform singular value decomposition SVD: cov represents covariance processing, the superscript H represents conjugate transpose, U and ∑ are the eigenvector matrix and eigenvalue matrix after performing singular value decomposition (SVD), respectively; Perform signal processing on the diagonal singular values of the matrix ∑: λ d represents the processed output value, λ n represents the singular value corresponding to the matrix ∑, λ1 represents the first singular value of the matrix ∑, and all output values form a specific time series related to the defect information, namely the waveguide signal defect information index TRDI; S3: Build a convolutional neural network (CNN) model, train the CNN model using the waveguide signal defect information index (TRDI), and establish a nonlinear function mapping relationship between the waveguide signal defect information index (TRDI) and the defect location. S4: Use the convolutional neural network (CNN) model to detect and locate defects in the entire composite material.

2. The method for ultrasonic guided wave defect detection and positioning based on singular value decomposition according to claim 1, characterized in that: In S2, a rectangular window function is used to slide and divide the waveguide signal set matrix into short time domain segments. The width of the rectangular window is set to be the same as the length of the excitation signal generated by the waveform generator. The rectangular window is slid along the entire signal matrix to decompose the waveguide signal into different short time domain segments.

3. The method for ultrasonic guided wave defect detection and location based on singular value decomposition according to claim 1, characterized in that: In S3, the three convolutional layers LReLU of the convolutional neural network CNN model are respectively set with convolution kernels of sizes 3×3, 5×5, and 7×7, and the numbers are 16, 32, and 64 respectively. The specific function processing feature of each layer is that the waveguide signal defect information index TRDI passes through the convolutional layer LReLU to perform vector convolution operation and convolution processing of the linear rectification function, and each feature map is nonlinearly processed by the linear rectification function.

4. The method for ultrasonic guided wave defect detection and location based on singular value decomposition according to claim 1, characterized in that: In S4, the softmax function is used to perform defect detection and positioning operations, the cross entropy function is selected as the loss function of the convolutional neural network (CNN) model, the Adam algorithm is used to optimize the convolutional neural network (CNN) model with an adaptive learning rate, and the convolutional neural network (CNN) model is optimized by small-batch stochastic gradient descent based on the first-order gradient. The hyperparameters of the Adam algorithm are set to: α = 0.001, β1 = 0.9, β2 = 0.

999. The Adam algorithm updates the learnable parameters of the entire convolutional neural network (CNN) model to obtain the optimal parameter settings.

Images