A method and system for detecting fatigue cracks in structures using spectral noise reduction based on long short-term memory networks and nonlinear ultrasonic modulation.

Abstract

A method and system for detecting fatigue cracks in a structure are disclosed. A vibration detection element detects vibrations of the structure generated when the structure is excited at low and high frequencies, and generates a measurement signal in the time domain. Based on a trained Long Short-Term Memory (LSTM) network, a signal for the next time step is predicted based on the measurement signal at the current time step measured by the vibration detection element. The measurement signal and the predicted signal are combined to generate a new combined signal with increased length in the time domain. A Fourier transform is performed on the combined signal to reduce noise and enhance the information of the measurement signal in the frequency domain, and the Fourier-transformed signal is used to obtain the spectral density function and detect cracks in the structure.

Description

Technical Field

[0001] This invention relates to the technical field of detecting fatigue cracks in structures, and more particularly to a method and system for detecting fatigue cracks in structures based on long short-term memory (LSTM) networks and nonlinear ultrasonic modulation technology. Background Technology

[0002] Fatigue cracking in metal structures is a critical issue in practice. This is because fatigue cracking is a leading cause of failure in metal structures and components. Generally, fatigue cracks are not noticed until they have propagated to approximately 80% of the structure's overall fatigue life. Therefore, detecting fatigue cracks in their early stages is crucial to avoid catastrophic failures.

[0003] Various nondestructive testing (NDE) techniques and structural health monitoring (SHM) techniques are known, such as ultrasonic methods, thermal imaging methods, acoustic emission methods, magnetic particle testing methods, X-ray imaging methods, and eddy current methods. Among these, ultrasonic methods are particularly effective for continuous online monitoring of fatigue cracks, making them one of the most promising methods for fatigue crack detection. Furthermore, this technique has proven to effectively achieve a reasonable trade-off between resolution, detectability, and practicality.

[0004] Traditional linear ultrasonic methods utilize variations in amplitude attenuation, phase delay, and mode conversion of linear ultrasound waves transmitted through or reflected from fatigue cracks. However, because these changes in linear properties are only noticeable when structural damage is severe, traditional linear ultrasonic methods struggle to detect fatigue cracks that primarily occur in their early stages. Recent studies have shown that fatigue cracks are a source of nonlinear ultrasound, and that nonlinear ultrasonic methods are far more sensitive to fatigue cracks than traditional linear ultrasonic methods.

[0005] Nonlinear ultrasonic methods show promise, but their corresponding nonlinear responses are quite weak, making nonlinear characteristics susceptible to noise. Therefore, extracting nonlinear characteristics using the spectral density function is difficult under noisy conditions. Specifically, this is because noise overlaps with the nonlinear characteristics in the spectral domain. Therefore, reducing spectral noise is crucial and offers significant advantages for improving the performance of fatigue crack detection based on nonlinear characteristics.

[0006] Spectral subtraction has been employed to reduce spectral noise. This method reconstructs the power spectrum of the observed signal from additive noise by subtracting an estimate of the average noise spectrum from the noise signal spectrum. The noise spectrum is typically estimated and updated in the absence of a signal. However, due to the random variations in noise, spectral subtraction can lead to a negative estimate of the power spectrum. Furthermore, this method is unsuitable for ultrasonic signals because there are no signal-free periods in ultrasonic signals.

[0007] Zero-padding is a widely used method for reducing spectral noise. Zero-padding involves adding zero samples. However, since signals extrapolated with zeros cannot retain information about the signal, the nonlinear characteristics of the signal are also lost. Summary of the Invention

[0008] Technical goals

[0009] The objective of this invention is to provide a method that can improve the performance of detecting fatigue cracks in structures by using a data reduction effect without compromising the spectral density and amplitude of existing nonlinear ultrasonic modulation components, while using the predictive properties of an LSTM network for ultrasonic signals to reduce spectral noise.

[0010] Another objective of the present invention is to provide a system capable of performing the above-described methods.

[0011] The problems to be solved by the present invention are not limited to those described above, and can be extended in various ways without departing from the spirit and scope of the present invention.

[0012] Technical solution

[0013] A method for detecting fatigue cracks in a structure according to an embodiment for achieving the objectives of the present invention includes: in a response signal processing unit, training an LSTM network using ultrasonic measurement signals measured from a structure simultaneously excited by ultrasonic signals having two distinguishable frequencies to obtain a prediction model for time-domain signals; in the response signal processing unit, depending on the set number of data points, training an LSTM network using ultrasonic measurement signals measured from a structure simultaneously excited by ultrasonic signals having two distinguishable frequencies to obtain a prediction model for time-domain signals; a First ultrasonic signal and second frequency ω b The ultrasonic measurement signal of the target structure excited by the second ultrasonic signal at the current time step (where ω) a <ω b The signal is input into a trained LSTM network to obtain the ultrasonic prediction signal at the next time step; in the response signal processing unit, the obtained ultrasonic prediction signal is used to reconstruct the signal; in the response signal processing unit, the reconstructed signal is subjected to a Fourier transform; and in the response signal processing unit, whether a crack has appeared in the target structure is determined by detecting the nonlinear modulation component based on the value of the spectral density function obtained using the Fourier transform signal.

[0014] In one exemplary embodiment, a method for detecting fatigue cracks in a structure may include: calculating a damage index by substituting a spectral density function into a nonlinear modulation parameter equation in a response signal processing unit; and using the calculated damage index to detect fatigue cracks in the structure.

[0015] In one exemplary embodiment, an equation can be used. To calculate the spectral density function P xN (ω), where X N (ω) represents the reconstructed ultrasonic signal x N The Fourier transform signal of (t), where * denotes complex conjugate, and E[] denotes the expectation operator. Furthermore, the damage index β... DN Equations can be used

[0016] To obtain.

[0017] In one exemplary embodiment, the method for detecting fatigue cracks in a structure may further include: simultaneously exciting a target structure by generating first and second ultrasonic signals by an excitation unit and applying them respectively to a first excitation element and a second excitation element attached to the target structure; detecting vibration of the target structure based on the excitation of the first and second ultrasonic signals by a vibration detection element attached to the target structure, and providing the corresponding ultrasonic measurement signal to a response signal processing unit.

[0018] In one exemplary embodiment, the prediction model of a trained LSTM network can be designed and trained in such a way that the previous cell state is updated to a new cell state by partially adding new information provided by the input gate while forgetting some of the previous cell state information, in order to predict the ultrasonic signal at the next time step by learning the nonlinear modulation frequency component (which is the pattern caused by fatigue cracks on the full time series data of the ultrasonic measurement signal).

[0019] In one exemplary embodiment, the reconstructed signal may be a signal reconstructed using only the ultrasound prediction signal at the next time step.

[0020] In one exemplary embodiment, the number of data points of the reconstructed signal is determined by multiplying the number of data points of the ultrasonic measurement signal by a data reduction rate α having a predetermined size, and the data reduction rate α can be determined in the range of 0 to 1.

[0021] In one exemplary embodiment, the reconstructed signal may be a signal reconstructed by combining the ultrasonic measurement signal at the current time step and the ultrasonic prediction signal at the next time step.

[0022] In one exemplary embodiment, the Fourier transform of the reconstructed signal can be a signal in the frequency domain, which reduces noise and enhances the information of the ultrasonic measurement signal.

[0023] In another aspect of the invention, an exemplary embodiment provides a system for detecting fatigue cracks in a structure, comprising a first excitation element, a second excitation element, an excitation unit, a vibration detection element, and a response signal processing unit. The first excitation element is attached to a first region of the target structure and configured to detect vibration cracks by inputting a first frequency ω. a The target structure is excited by vibration from a first ultrasonic signal. A second excitation element is attached to a first region of the target structure and configured to excite the target structure by vibration upon input of a second ultrasonic signal ω. b The second ultrasonic signal is used to excite the target structure through vibration, where ω a <ω b The excitation unit is configured to generate a first ultrasonic signal and a second ultrasonic signal, and simultaneously provide them to a first excitation element and a second excitation element, respectively. A vibration detection element is attached to a second region of the target structure spaced apart from the first region and is configured to detect vibrations of the target structure based on the excitation of the first and second ultrasonic signals to generate corresponding ultrasonic measurement signals. The response signal processing unit includes an operation processing unit configured to perform the following functions: training an LSTM network using ultrasonic measurement signals measured from a structure simultaneously excited by ultrasonic signals having two distinguishable frequencies to obtain a prediction model for time-domain signals; and, depending on the set number of data points, processing the vibrations measured from the structure simultaneously excited by the first frequency ω. a First ultrasonic signal and second frequency ω b The ultrasonic measurement signal of the target structure excited by the second ultrasonic signal at the current time step (where ω) a <ω b The signal is input into a trained LSTM network to obtain the ultrasonic prediction signal at the next time step; the signal is reconstructed using the obtained ultrasonic prediction signal; in the response signal processing unit, the reconstructed signal is subjected to a Fourier transform; and whether a crack has appeared in the target structure is determined by detecting the nonlinear modulation component based on the value of the spectral density function obtained using the Fourier transform signal.

[0024] In an exemplary embodiment, the operation processing unit of the response signal processing unit may be configured to further perform the following functions: calculate the damage index by substituting the spectral density function into the nonlinear modulation parameter equation; and use the calculated damage index to detect fatigue cracks in the structure.

[0025] In one exemplary embodiment, the response signal processing unit may further include a digitization unit that converts the analog measurement signal measured by the vibration detection element into a digital measurement signal and provides the converted digital measurement signal to the operation processing unit.

[0026] In one exemplary embodiment, the first excitation element, the second excitation element, and the vibration detection element may be composed of piezoelectric elements.

[0027] In one exemplary embodiment, the prediction model of a trained LSTM network can be designed and trained in such a way that the previous cell state is updated to a new cell state by partially adding new information provided by the input gate while forgetting some of the previous cell state information, in order to predict the ultrasonic signal at the next time step by learning the nonlinear modulation frequency component (which is the pattern caused by fatigue cracks on the full time series data of the ultrasonic measurement signal).

[0028] In one exemplary embodiment, the reconstructed signal may be a signal reconstructed using only the ultrasound prediction signal at the next time step.

[0029] In one exemplary embodiment, the number of data points of the reconstructed signal can be determined by multiplying the number of data points of the ultrasonic measurement signal by a data reduction rate α of a predetermined size, and the data reduction rate α can be determined in the range of 0 to 1.

[0030] In one exemplary embodiment, the reconstructed signal may be a signal reconstructed by combining the ultrasonic measurement signal at the current time step and the ultrasonic prediction signal at the next time step.

[0031] Invention Effects

[0032] According to an exemplary embodiment of the present invention, an LSTM network is used, which is designed and trained to predict the ultrasonic signal at the next time step by learning nonlinear modulation frequency components (which are fatigue crack-induced patterns on the entire time series data of the measured ultrasonic signal). In the signal reconstructed by the trained LSTM, the nonlinear modulation components required for fatigue crack detection are preserved and amplified at their original level, while noise components are reduced. That is, using a trained LSTM network can significantly improve the signal-to-noise ratio (SNR) (up to 276%), thereby effectively improving the performance of fatigue crack detection.

[0033] Furthermore, a trained LSTM network can generate a reconstructed signal using far fewer data points than the measured signal without reducing the modulation amplitude. In this way, the amount of data to be processed can be reduced. A trained LSTM network can generate a reconstructed ultrasonic signal using only 20% of the original data. In this case, the nonlinear modulation amplitude obtained from the reconstructed signal is equal to the modulation amplitude of the original signal; therefore, despite the reduced number of data points, the performance of fatigue crack detection is not degraded. Attached Figure Description

[0034] Figure 1(A) illustrates the performance of ultrasonic measurements on an intact structure using an ultrasonic modulation-based crack detection system according to an exemplary embodiment of the invention, and (B) represents the frequency response characteristics measured from the intact structure.

[0035] Figure 2 (A) shows the use of Figure 1 The ultrasonic modulation-based crack detection system performs nonlinear ultrasonic measurements on a damaged structure with cracks, and (B) shows the frequency response characteristics measured from the damaged structure.

[0036] Figure 3 The structure of (A) an LSTM network and (B) a storage cell for a fatigue crack detection method according to an exemplary embodiment of the present invention are schematically shown.

[0037] Figure 4 The flowchart illustrates a spectral noise and data reduction algorithm for fatigue crack detection in a structure according to an exemplary embodiment of the present invention.

[0038] Figure 5 The diagram shows (A) a time-series ultrasonic measurement signal and (B) a signal representing that signal in the spectral domain.

[0039] Figure 6 A training LSTM network prediction model according to an exemplary embodiment of the present invention is shown.

[0040] Figure 7 The illustration schematically depicts the use of a trained LSTM network to reconstruct ultrasonic signals according to an exemplary embodiment of the present invention.

[0041] Figure 8 and Figure 9 A comparison is provided of the spectral density function P of the signal measured before (A) and after (B) using a trained LSTM network to reduce spectral noise and data on an aluminum plate sample with “fatigue cracks”. x Example of (ω).

[0042] Figure 10 An example is provided comparing the spectral density values ​​before (A) and after (B) applying spectral noise and data reduction techniques using a trained LSTM network to an aluminum plate sample "without fatigue cracks".

[0043] Figure 11 and Figure 12 The data reduction effect obtained when a spectral noise and data reduction method based on a trained LSTM network according to an exemplary embodiment of the present invention is applied to an aluminum plate sample with fatigue cracks is demonstrated.

[0044] Figure 13 The results of a performance evaluation of a spectral noise and data reduction method based on a trained LSTM network according to an exemplary embodiment of the present invention are shown using an aluminum plate sample. Detailed Implementation

[0045] Preferred embodiments of the invention will now be described in more detail with reference to the accompanying drawings. In the drawings, the same reference numerals will be used for the same elements, and redundant descriptions of the same elements will be omitted.

[0046] Figure 1 (A) illustrates an ultrasonic measurement of an intact structure 50 using an ultrasonic-based crack detection system 10 according to an exemplary embodiment of the invention, and (B) illustrates the frequency response characteristics measured from the structure 50. Figure 2 (A) shows the use of Figure 1 The ultrasonic-based crack detection system 10 shown performs nonlinear ultrasonic measurements on a damaged structure 60 with cracks, and (B) shows the frequency response characteristics measured from the damaged structure 60.

[0047] like Figure 1 As shown, the ultrasonic crack detection system 10 may include an excitation unit 20, a response signal processing unit 30, a first excitation element 42 and a second excitation element 44, and a vibration detection element 46.

[0048] The excitation unit 20 can be configured to generate and provide excitation signals to excite the structure 50 by vibrating the first excitation element 42 and the second excitation element 44. In one embodiment, the excitation unit 20 may include a waveform generator capable of generating arbitrary periodic waveforms with a predetermined frequency. The excitation unit 20 may generate a low-frequency ultrasonic signal LF(ω) a ) and high-frequency ultrasonic signal HF(ω) b These are used as excitation signals and are provided to the first excitation element 42 and the second excitation element 44 respectively.

[0049] The first excitation element 42 and the second excitation element 44 can be attached close to each other in a first region of the complete structure 50, and the vibration detection element 46 can be attached to a second region of the structure 50. The first region and the second region can be located at opposite ends of the target region for crack detection. The excitation unit 20 is connected to the first excitation element 42 and the second excitation element 44.

[0050] In an exemplary embodiment, the first excitation element 42 and the second excitation element 44 may be composed of, for example, piezoelectric elements. When the excitation unit 20 simultaneously transmits a low-frequency ultrasonic signal LF(ω)... a ) and high-frequency ultrasonic signal HF(ω) bWhen the inputs are respectively given to the first excitation element 42 and the second excitation element 44, the first excitation element 42 and the second excitation element 44 can respectively be driven by low frequency ω a and high frequency ω b Vibration is used to excite structure 50. Therefore, structure 50 is excited at a low frequency ω. a and high frequency ω b It produces ultrasonic vibrations.

[0051] In one exemplary embodiment, the vibration detection element 46 may also be composed of, for example, a piezoelectric element. When the structure 50 vibrates under the excitation of the first excitation element 42 and the second excitation element 44, the vibration can be transmitted to the vibration detection device 46. The vibration detection element 46 can detect the vibration of the structure 50 and generate a corresponding electrical signal. For example, the vibration detection element 46 can detect the vibration of the structure 50 and output low-frequency and high-frequency analog signals. The amplitudes of the two analog response signals can be set to, for example, a peak-to-peak voltage of 16V.

[0052] In one exemplary embodiment, the response signal processing unit 30 may be connected to the vibration detection element 46. The response signal processing unit 30 may be configured to generate information about the presence of cracks in the structure 50 by receiving a simulated response signal corresponding to a vibration of the structure 50 detected by the vibration detection element 46 and performing specific processing to obtain the frequency response of the structure 50. In one exemplary embodiment, the response signal processing unit 30 may be configured to generate a damage index of the structure 50 by processing the simulated response signal corresponding to the vibration detected by the vibration detection element 46.

[0053] In one exemplary embodiment, the response signal processing unit 30 may include a digitization unit 32 that converts the output analog signal from the vibration detection element 46 into a digital signal; and an operation processing unit 34 that receives the converted digital signal and performs specific operations to obtain the frequency response, damage index, etc. of the structure 50. The digitization unit 32 may perform the conversion at, for example, a sampling rate of 1 MHz and 0.1 seconds. The response signal processing unit 30 may be implemented using hardware components, software components, and / or a combination of hardware and software components.

[0054] The operation processing unit 34 can be implemented as a computer program and a computing device capable of executing the computer program and performing predetermined tasks or providing functions as instructed by the computer program. The operation processing unit 34 can be a computing device configured to predict the measurement signal at the next time step by executing a computer program (described later) to convert the signal in the time domain to a signal in the frequency domain through learning from an LSTM network, and to calculate the impairment index by executing a computer program implementing a spectral denoising algorithm (described later). The computing device can be implemented using, for example, a processor, controller, arithmetic logic unit (ALU), digital signal processor, microcomputer, field-programmable array (FPA), programmable logic unit (PLU), microprocessor, or one or more general-purpose or special-purpose computers, just like any other device capable of executing and responding to instructions. The computing device may also include computing resources such as memory, data storage devices, input / output units, and other computing resources.

[0055] Computer programs implemented to perform the functions of the response signal processing unit 30, including LSTM network models and their learning data, can be stored in one or more computer-readable recording media. The method according to an exemplary embodiment of the present invention can be implemented in the form of program commands, which can be executed by various computer means and recorded on a computer-readable medium. The computer-readable recording medium can include program commands, data files, data structures, etc., individually or in combination. The program commands recorded on the recording medium can be specifically designed and configured for the exemplary embodiment and can be known and available to those skilled in the art of computer programming.

[0056] In the ultrasonic-based crack detection system 10 with this configuration, the excitation unit 20 can have two different frequencies, for example, a low frequency ω. a and high frequency ω b (where ω) a <ω b Two ultrasonic signals are applied to the first excitation element 42 and the second excitation element 44 respectively to simultaneously excite the complete structure 50. Therefore, the complete structure 50 vibrates, and simultaneously, the vibration detection element 46 can be used to measure the response corresponding to the vibration of the structure 50. The signal of the vibration of the structure 50 measured by the vibration detection element 46 can be an analog signal in the time domain. The response signal processing unit 30 processes the analog measurement signal received from the vibration detection element 46 to observe only the two frequencies ω of the input signal that excite the structure 50. a and ω b The responses A and B are as seen in the spectral domain. In other words, when there are two different frequencies ω... a and ω b (ω a<ω b The ultrasonic input signal LF(ω) a ) and HF(ω) b When simultaneously input to the complete structure 50, only at the two input frequencies ω a and ω b The frequency responses A and B caused by the vibration of the undamaged structure 50 are observed below (see Figure 1 (B)).

[0057] However, unlike the above situation, when two ultrasonic input signals LF(ω) a ) and HF(ω) b When applied to a damaged (nonlinear) structure 60 with features such as cracks, the nonlinear ultrasonic modulation is caused by the crack opening and closing mechanism. (Reference) Figure 2 , such as in Figure 1 In this case, the excitation unit 20 will simultaneously have a low frequency ω a and high frequency ω b Two ultrasonic excitation signals LF(ω) a ) and HF(ω) b These are respectively applied to the first excitation element 42 and the second excitation element 44 to excite the damaged (nonlinear) structure 60. Simultaneously, if the vibration response of the damaged structure 60 is measured by the vibration detection element 46, a simulated response signal in the time domain can be obtained, such as... Figure 2 As shown in (B). By transforming the analog response signal to the frequency domain, the difference in the response in the spectrum compared to the complete structure 50 can be examined. When structure 60 operates nonlinearly due to the crack, the response from structure 60 in the frequency domain can be included not only at the two input frequencies ω. a and ω b The response components A and B at the given frequency also include those at the following frequencies (ω). b +ω a ) and difference frequency (ω) b -ω a Modulation component M under ) o and M s This is the modulation frequency of the input frequency. That is, due to the low-frequency ultrasonic input signal, the amplitude of the high-frequency ultrasonic input signal is modulated by the crack opening and closing mechanism. Amplitude modulation occurs at the sum and difference frequencies (ω) of the input frequency. a ±ω b Additional frequency components (nonlinear modulation components) are created under these conditions. This phenomenon is called "nonlinear ultrasonic modulation." Since this phenomenon only occurs when nonlinear characteristics are present, it can be considered an index of damage in the structure. Therefore, the presence of fatigue cracks can be identified by searching for and extracting the modulation components in the spectral domain.

[0058] However, the amplitude of the nonlinear ultrasonic modulation component caused by structural nonlinearity is usually small (M O M S This makes it difficult to distinguish from noise. To address this problem of being highly susceptible to noise, signal processing is required to reduce noise and differentiate the modulation frequency response, i.e., the amplitude of the modulation component from the noise. To this end, an exemplary embodiment of the present invention proposes a method for selectively detecting damage to a structure by using an LSTM network and Fourier transform to reduce spectral noise. This will be described in detail below.

[0059] LSTM networks are effectively used to predict sequential data, such as time-domain signals. An LSTM network is a special type of recurrent neural network (RNN) architecture that can be applied to sequential data. However, traditional RNNs often struggle to learn the long-term dependencies of time-series data due to gradient explosion and vanishing gradient problems during training. To overcome this problem, LSTM networks have been developed. LSTM networks use blocks of storage cells that can represent the long-term dependencies of time-series data. Specifically, LSTM networks can control the degree to which data is remembered (reflected) from the distant past by adding cell state gates (C) to an existing RNN.

[0060] Figure 3 (A) and (B) schematically illustrate the architecture of the LSTM network and storage unit in a method for detecting fatigue cracks according to an exemplary embodiment of the present invention.

[0061] refer to Figure 3 (A) The LSTM network 80 has an architecture in which blocks of multiple memory cells 70 are sequentially connected. Each memory cell 70 may include an input gate i t Output gate o t Forgotten Gate t and self-circulating neurons. These gates i t o t and f t Controls interaction with adjacent memory cells. An LSTM network can include an input layer, a fully connected hidden layer, and an output layer. The hidden layer includes memory cells, associated gate cells, and hidden cells, which provide input to the gate cells and memory cells. Input gate i t It can control whether the input signal can modify the state of the storage unit 70. On the other hand, the output gate o t It can control whether the state of other storage units 70 can be modified. Forget gate f t It can be decided whether to forget or remember the previous state. That is, the previous storage unit state C. t-1 The degree to which something can be reflected is determined by the forgetting gate f. tControlled. Computation within an LSTM memory cell block can be represented as follows:

[0062] f t =σ(X) t U J +S t-1 W J +b f )......(1)

[0063] i t =σ(X) t U i +S t-1 W i +b i )......(2)

[0064]

[0065]

[0066] O t =σ(X) t U o +S t-1 W o +b o )......(5)

[0067]

[0068] Here, U(U) f U i U o U c b(b) f b i b o b c ) and W(W f W i W o W c These represent the input weights, bias weights, and recurrent weights, respectively. The superscripts f, i, o, and c represent the forget gate, input gate, output gate, and recurrent neuron, respectively. X t The data is input sequentially at time step t. t C t , σ and τ represent the hidden state, the cell state, the new candidate value for the cell state, the sigmoid activation function, and the hyperbolic tangent (tanh) activation function, respectively. Finally, and These are the dot multiplication and addition operators, respectively.

[0069] The first step in an LSTM network is to select which information to discard from the cell state. This decision is made by the forget gate f in equation (1). t The second step is to select the new information to be stored in the cell state. The input gate in equation (2) determines which input values ​​to update. The hyperbolic tangent (tanh) activation function τ sequentially creates new candidate values ​​that can be added to the cell state in equation (3). The third step is to restore the old cell state C by forgetting a portion of the previous cell state information (forgetting the information that was decided to be forgotten at the previous time step) and partially adding new information provided from the input gate. t-1 Update to the new unit state C t In this regard, as shown in equation (4), the previous cell state C will be changed. t-1 dot product of f t The value obtained With new candidate values dot multiplication by input gate information i t The value obtained Add them together. Finally, use output gate O. t Update the hidden cell state, and update the cell state C. t As shown in equations (5) and (6).

[0070] The key feature of LSTM networks is the cell state C. t It passes horizontally through the top of the storage cell block 70 and interacts linearly with the gate, as... Figure 3 As shown. The LSTM network 80 can remove past information or add new information to the cell state through gates. Gates provide a way to selectively pass information through the cell state. For example, the sigmoid activation function σ in the gates used for storage cell 70 controls the amount of information passing through the sigmoid function by changing its output value between 0 and 1. If the output value is 0, it means that no information passes through. If it is 1, it means that all information passes through.

[0071] The spectral noise reduction and data reduction techniques according to exemplary embodiments utilize the signal prediction (reconstruction) properties of LSTM networks. LSTM networks have advantages in learning long-term patterns in data. Ultrasonic signals can be considered long-term time-series data because many data points in the time domain are obtained at high sampling frequencies. Considering this, LSTM networks can be constructed and trained to learn fatigue crack-induced patterns (nonlinear modulation frequency components) in the measured ultrasonic signals. Here, fatigue crack-induced patterns can be extracted along the entire time-series data. However, noise components have a random distribution in the data, and LSTM networks do not learn noise components. Therefore, LSTM networks can be used to remove noise components, and the reconstructed signal retains only meaningful components, such as the nonlinear modulation components caused by fatigue cracks. Thus, the amplitude of the nonlinear modulation components can be amplified, and the noise floor level can be reduced in the spectral domain.

[0072] The Fourier transform converts data from the time domain to the frequency domain. That is, the Fourier transform can be used to analyze a measured time-domain signal x(t) in the frequency domain.

[0073] The spectral coefficients of the measured signal x(t) are given by the following equation.

[0074]

[0075] The coefficient a(0) is a constant component of x(t) and is calculated by equation (7), where k = 0.

[0076]

[0077] Here, it is simply the average value of x(t) over one period. When the signal is represented as a Fourier series, the Fourier transform is derived from the weights used to represent the specific frequencies of the signal, and the formula for the Fourier transform is as follows.

[0078]

[0079] The Fourier transform represents the information needed to describe a measured signal x(t) as a linear combination of sinusoidal signals at different frequencies. That is, when x(t) is transformed into X(ω) via the Fourier transform as shown in equation (7-9), the signal length T is related to averaging. Therefore, if the length T of the measured signal x(t) can be extended while preserving its information, the spectral noise of the signal X(ω) obtained by performing the Fourier transform can be reduced in the frequency domain through noise averaging.

[0080] More specifically, as the length of the signal in the time domain increases, the Fourier transform of the signal provides an effect that enhances the number of averaging iterations. Since noise is randomly distributed, it decreases with increasing averaging iterations. Therefore, to increase the signal length in the time domain, the signal for the next time interval of the measured signal at the current time interval can be predicted using an LSTM network 80' trained on a large amount of training data. In building the LSTM network model, noise is not learned, but information included in the measured signal is. Therefore, noise is not reflected in the signal predicted using the LSTM network, but only the learned information. If the measured and predicted signals are combined into a new signal in the time domain and a Fourier transform is performed on the new signal, noise decreases with increasing averaging iterations, but the information included in the signal is enhanced. Information about cracks can also be enriched. While reducing noise, the crack information included in the signal is also enhanced, thus increasing the SNR and improving crack detection capability.

[0081] The modulation component can be extracted by calculating the spectral density function (power function), and the function is given below.

[0082] P x (ω)=E[X(ω)X * (ω)]......(10)

[0083] Here, * denotes complex conjugation, and E[] denotes the expectation operator. In an exemplary embodiment, the damage exponent is considered at the sum frequency and difference frequency ω. a ±ω b The next first-order modulation. Nonlinear modulation parameter β D It can be defined as follows.

[0084]

[0085] Here, k a and k b These are the corresponding wave numbers. Substituting equation (10) into equation (11), we obtain the following equation.

[0086]

[0087] Due to its sensitivity to fatigue cracks, this nonlinear modulation parameter β D It can be used as a damage index.

[0088] As explained above, spectral noise reduction is developed using LSTM networks and Fourier transforms. An LSTM network can be trained to predict the signal x′(t) at the next time step of the measured signal by learning the signal x(t) at the current time step. Through training, a model of the LSTM network capable of predicting the signal at the next time step is obtained. The signal length T of the signal x(t) measured by the LSTM network can be extended using the preserved signal information. Therefore, when performing a Fourier transform based on the noise averaging effect, spectral noise is reduced. Conversely, since the characteristics of the signal, i.e., the signal information, are learned by the trained LSTM network and reflected in the signal at the predicted time step, the signal information is not reduced even after performing a Fourier transform. Therefore, the signal-to-noise ratio (SNR) can be improved due to the relationship between the signal length in the time domain and the spectral signal.

[0089] Based on this concept, it is possible to reduce spectral noise and data used for fatigue crack detection in structures. Figure 4 The flowchart schematically illustrates a spectral noise and data reduction algorithm for detecting fatigue cracks in a structure according to an exemplary embodiment of the present invention. An ultrasonic-based spectral noise and data reduction technique for nonlinear fatigue crack detection can utilize an LSTM network 80.

[0090] refer to Figure 4 It is possible to measure ultrasonic signals from target structure 60 while using low-frequency LFω at the current time step. a and high frequency ω b The excitation signal is applied to the excitation structure 60 (S10). The measured ultrasonic signal x(t) m () is a signal in the time domain and has a length T (0 ≤ t) m ≤T), which is the time interval of the current time step. Figure 5 (A) shows the time-series ultrasonic data, and (B) shows the signal representing this signal in the spectral domain. The ultrasonic measurement signal x(t) has N m There are several data points. Here, in the ultrasonic measurement signal x(t), the spectral density may include the vibration response component of the target structure 60 (i.e., the input frequency component) and random noise components, and because the target structure 60 has fatigue cracks, it may even include nonlinear modulation components caused by fatigue cracks.

[0091] For ultrasonic signal measurement, as described above, the excitation unit 20 can transmit low-frequency LF(ω) signals. a ) and high frequency HF(ω) bTwo excitation signals are transmitted to the first excitation element 42 and the second excitation element 44 attached to the target structure 60, respectively, to simultaneously vibrate the target structure 60, thereby exciting the target structure 60. Simultaneously, the vibration detection element 46 attached to the target structure 60 can be used to measure the response corresponding to the excitation, and the measured signal x(t) can be transmitted to the first excitation element 42 and the second excitation element 44. m The measured signal x(t) is provided to the response signal processing unit 30. m It can be an ultrasonic signal.

[0092] As a next step, an LSTM network can be constructed and the original ultrasonic signal x(N) can be used. m f s The LSTM network is trained to build a prediction model 80 to learn the basic pattern of the measured ultrasound signal, rather than random noise patterns (step S20). The input and output of the LSTM network can be set as neighboring data points in the time series data. That is, the LSTM network can be designed and trained to predict the data at the next time step using the data at the current time step. Figure 6 The training of the LSTM network prediction model 80' in this manner is shown.

[0093] Specifically, in the response signal processing unit 30, the measurement signal x(t) at the current time step can be predicted based on a trained LSTM network model. m The signal for the next time step, i.e., until time T. f (where T≤t) f ≤T f The signal x'(t) f Therefore, as mentioned above, when the measurement signal at the current time step is input, the trained LSTM network model can predict the signal at the next time step by learning a large amount of measurement signals in advance.

[0094] New time-series ultrasonic signals can be reconstructed by using a trained LSTM network model to generate a predicted signal of the ultrasonic signal in the time domain. Figure 7 The use of a trained LSTM network 80' to reconstruct an ultrasonic signal is illustrated schematically according to an exemplary embodiment.

[0095] In one exemplary embodiment, such as Figure 7 As shown in (A), the ultrasonic measurement signal x(t) at the current time step m ) can be input into the trained network model 80', according to the number of data points N f To generate the prediction signal x'(t) for the next time step f The number of data points N fIt can be preset to an optimal value, and the optimal value can be obtained through testing. Then, it can be obtained by combining the measured signal x(t) at the current time step. m ) and the predicted signal x′(t) for the next time step f To generate the following reconstructed ultrasonic signal x N (t)(Step S30). Here, N is an index representing the reconstructed ultrasonic signal.

[0096] x N (t)=x(t m )+x′(t f ), 0≤t≤T f ......(13-1)

[0097] Unlike the previous embodiments, in obtaining the reconstructed ultrasonic signal x N In (t), such as Figure 7 As shown in (B), only the predicted signal x′(t) can be used. f ) but does not include the measurement signal x(t) m To reconstruct the ultrasonic signal x N (t).

[0098] x N (t)=x′(t f ), T≤t≤T f ......(13-2)

[0099] In this case, compared to the previous embodiment, the reconstructed ultrasonic signal and the measured signal x(t) m The length T is the same. Since the measurement signal contains a lot of noise, the reconstructed signal obtained from the prediction signal alone, which has no noise components, may be more effective in reducing noise than the reconstructed signal obtained by combining the measurement signal and the prediction signal.

[0100] Next, the reconstructed ultrasonic signal x N (t) Perform a Fourier transform and analyze it in the spectral domain to identify the nonlinear modulation frequency components caused by fatigue cracks (step S40). When the reconstructed ultrasonic signal x obtained as described above... N (t) When performing analysis in the spectral domain, the reconstructed ultrasonic signal x can be obtained. N The noise component is effectively reduced in the spectral domain of (t), and therefore only the component corresponding to the two input frequencies and the component corresponding to its modulation frequency appear.

[0101] Specifically, equation (9) can be used to reconstruct the signal x. N (t) Performs a Fourier transform and can compute the spectral density function (S40). Used to reconstruct the signal x NThe Fourier transform equation of (t) is given by the following equation (14).

[0102]

[0103] Therefore, the spectral noise is reduced, and the spectral density function after noise reduction can be represented by the following equation (10).

[0104]

[0105] By combining the signal x N Substituting equation (15) of the spectral density function of (t) into equation (12) of the nonlinear modulation parameters, we can obtain the damage index β expressed by equation (16). DN Equation. Damage index β of structure 60. DN It can be calculated using equation (16).

[0106]

[0107] The damage index β obtained in this way DN The equation can be used to detect fatigue cracks.

[0108] To evaluate the noise reduction performance of the spectral noise and data reduction techniques according to the exemplary embodiments, the signal-to-noise ratio (SNR) can be calculated using the following equation.

[0109]

[0110] Here, P xN (ω b -w a ) and P xN (ω b +w a ) respectively represent the modulation frequency ω b ±ω a The amplitude of the power density function under the given conditions. In equation (17), the denominator is... It is the average noise floor (NF). The average NF is the average value of the spectral power density, excluding the main components of the signal, such as the input component and the nonlinear modulation component.

[0111] Finally, the number of data points N required to achieve a modulation amplitude level equal to the original measured signal can be estimated. f To evaluate the effectiveness of data reduction. The number of data points N f It can be obtained through the following equation.

[0112] N f =αN m ......(18)

[0113] Here, α represents the data reduction rate, and its value can be determined in the range of 0 or more to 1 or less.

[0114] Simultaneously, when constructing the LSTM network 80' model, a single-hidden-layer LSTM network can be built to avoid overfitting. The Adaptive Moments Estimation (ADAM) optimizer can be used for LSTM network training. Regarding the hyperparameters of the ADAM optimizer, the gradient decay and squared gradient decay factors can be set, for example, to 0.9 and 0.999, respectively. The ε value used to prevent division by zero and the initial learning rate can be set to 1.0e, respectively. -8 and 1.5e -3 After half of the training epochs, the learning rate reduction factor can be set to 0.1. The root mean square error (RMSE) function can be used as the cost function. The construction, training, and testing of the LSTM network can be performed in a MATLAB (R2019a) environment on a computer equipped with a processor (e.g., GPU) and RAM. For training, 90% of the measured time-domain ultrasonic signals can be used as the training dataset, and the remaining 10% of the measurement data can be used as the validation dataset. As mentioned above, the spectral noise and data reduction algorithms can be implemented as a computer program and can be stored on a computer-readable recording medium. Furthermore, the computer program recorded on such a recording medium can be executed by the computing device of the response signal processing unit 30.

[0115] The performance of the LSTM network-based spectral noise reduction and data reduction techniques according to the exemplary embodiment was experimentally verified. A first piezoelectric (PZT) transducer and a second piezoelectric (PZT) transducer (corresponding to the first excitation element 42 and the second excitation element 44) for ultrasonic signal excitation of an aluminum plate sample, and a third piezoelectric transducer (corresponding to the vibration detection element 46) for detecting ultrasonic vibrations were mounted on the aluminum plate sample. A fatigue load of 28,000 times was applied to the aluminum plate sample to generate a fatigue crack with a length of 9 mm and a width of 20 μm.

[0116] Two waveform generators (corresponding to excitation unit 20) are used to generate waveforms with unique frequencies (ω). a =48kHz and ω bSinusoidal ultrasonic input signals of LF and HF (203 kHz) were applied to the first and second piezoelectric transducers for excitation. The duration and peak-to-peak amplitude were set to 0.1 seconds and 12 V, respectively. The aforementioned input frequencies were selected considering the local resonance characteristics of the sample and the overlap of higher-order harmonic components and nonlinear modulation components of the LF input. The corresponding ultrasonic response detected by the third piezoelectric transducer was obtained by digitizing the ultrasonic response at a sampling rate of 1 MHz within 0.1 seconds using a digitizer to improve the SNR of the ultrasonic response, and the ultrasonic response was averaged five times in the time domain.

[0117] Figure 8 and Figure 9 The spectral density function P of the measured signal before (A) and after (B) applying spectral noise and data reduction techniques using a trained LSTM network according to an exemplary embodiment of the present invention to an aluminum plate sample with "fatigue cracks" is shown. x The comparison results of (ω). Figure 8 N is one of them m 100K(N) f It is set to 100K, which is equal to N. m (for fair comparison), and the data acquisition duration T is 0.1 seconds, and f s When it is set to 1MHz. Figure 9 N is one of them m 30K(N) f Also 30K), the data acquisition time T is set to 0.03 seconds, f s Set to 1MHz and other test parameters are the same as Figure 8 The same situation as in [the previous text]. Below, by A... Ms and A Md This indicates that at the modulation frequency (ω) b +ω a ) and (ω b -ω a The amplitude P of the spectral density function under ) x (ω b +ω a ) and P x (ω b -ω a ).

[0118] Figure 8 (A) Displays the nonlinear modulation frequency component A Ms and A Md The values ​​are 6.71e -7 and 8.03e -7 . Figure 8 (B) shows the spectral density function of the signal reconstructed by the trained LSTM network. And A Ms and A Md The estimated values ​​are 3.27e. -6 and 2.99e -6 .and Figure 8 (A) Compared to, Figure 8 The amplitude of the modulation frequency component in (B) is significantly amplified, and the calculated NF values ​​are 4.6e -8 and 3.0e -9 Its value decreased significantly, as shown in (A) and (B). Furthermore, P... x (ω) and The SNR values ​​were 34 dBW and 76 dBW, respectively. Therefore, the SNR was improved by 224%.

[0119] Figure 9 (A) indicates that due to its high noise level (A) Md +A Ms =4.4e -7 Furthermore, with an SNR of 21 dBW, the nonlinear modulation frequency component cannot be well distinguished from noise, making it difficult to detect. On the other hand, Figure 9 (B) shows the improvement in SNR (A) Md +A Ms =2.0e -6 With SNR = 56 dBW, the nonlinear modulation frequency components are clearly distinguished from noise, and they are easy to detect.

[0120] Figure 10 The image shows a comparison of the spectral density values ​​before (A) and after (B) the application of an LSTM-based sample noise and data reduction method to an aluminum plate sample without "fatigue cracks" (data acquisition time T = 0.1 seconds, and N...). m =N f =100,000).

[0121] refer to Figure 10 The NF level decreases even lower in Figure (B) than in Figure (A), but no frequency modulation component is observed in either figure. In other words, before applying the method according to the invention, it is not easy to determine whether the modulation frequency component is included in the spectral density value due to the presence of a large number of noise components. However, after applying the method according to the invention, it can be clearly confirmed that there is no modulation frequency component in the spectral density value.

[0122] Summarize Figures 8 to 10The results shown demonstrate that the LSTM network 80' model established by the exemplary method according to the present invention significantly reduces the noise component included in the spectral density value of the reconstructed ultrasonic signal, while enhancing the nonlinear frequency modulation component, thereby improving the SNR and enabling the clear detection of the modulation frequency components appearing on both sides of the input high-frequency component.

[0123] at the same time, Figure 11 and Figure 12 The data reduction effect is shown when the spectral noise and data reduction method based on a trained LSTM network according to an exemplary embodiment of the present invention is applied to an aluminum plate sample with fatigue cracks.

[0124] refer to Figure 11 The nonlinear modulation frequency component A obtained from the "measured ultrasonic signal" with 480k data points Ms +A Md The amplitude is 6.3e -6 (See Figure (A)). The figure shows that even with only 100,000 data points (α = 0.21), the same level of nonlinear modulation frequency component A can be obtained from the "reconstructed ultrasonic signal" obtained using the trained LSTM network 80'. Ms +A Md The estimated SNR values ​​for the measured and reconstructed ultrasonic signals are 49 dBW and 76 dBW, respectively.

[0125] Figure 12 The demonstration shows an execution similar to using "measured ultrasonic signals" with 145,000 data points. Figure 11 The test results. The nonlinear modulation frequency component A was obtained from the "measured ultrasonic signal" with 145,000 data points. Ms +A Md The amplitude is 2.0e -6 Even with the number of data points reduced to 30,000 (α = 0.21), the nonlinear modulation frequency component A can be obtained from the "reconstructed ultrasonic signal" using the trained LSTM network 80'. Ms +A Md The same level of value 2.0e -6 The estimated SNR of the measured and reconstructed ultrasonic signals were 38 dBW and 56 dBW, respectively.

[0126] Figure 11 and Figure 12 The test results shown indicate that the method according to the exemplary method of the present invention can achieve a similar frequency modulation amplitude with reduced data points and noise.

[0127] Figure 13 The results of performance evaluation of the spectral noise and data reduction method based on a trained LSTM network according to an exemplary embodiment of the present invention are shown using aluminum plate samples. Figure (A) shows the spectral noise reduction performance, where the SNR is improved by 238% and 276% for 60k and 80k data points, respectively. Figure (B) shows the data reduction performance, where the data reduction rate α is 0.18 and 0.19 for 220k and 270k data points, respectively.

[0128] As described above, according to an exemplary embodiment of the present invention, the performance of fatigue crack detection can be significantly improved by using a trained LSTM network to significantly reduce spectral noise and data reduction methods. This is because the reconstructed signal from the trained LSTM can reduce noise components (SNR can be increased by up to 276%) while preserving and amplifying the nonlinear modulation components necessary for fatigue crack detection. Furthermore, the trained LSTM network can generate the reconstructed signal using far fewer data points than the measured signal without reducing the modulation amplitude, thereby reducing the amount of data to be processed (the reconstructed signal can be generated using only 20% of the original data).

[0129] Industrial applicability

[0130] This invention can be used to detect cracks that appear in physical structures, structural bodies, etc.

[0131] As described above, although embodiments have been depicted with limited accompanying drawings, it will be understood that those skilled in the art can make various modifications and changes to the invention without departing from the spirit and scope of the invention as set forth in the following claims. For example, suitable results may still be achieved even if the described techniques are performed in a different order than the methods described, and / or components of the systems, structures, devices, circuits, etc., are combined or assembled in a different form than described, or replaced or substituted by other elements or equivalents. Therefore, other embodiments, other examples, and equivalents of the claims fall within the scope of the claims.

Claims

1. A method for detecting fatigue cracks in a structure, comprising: In the response signal processing unit, an LSTM network is trained using ultrasonic measurement signals measured from a structure simultaneously excited by ultrasonic signals with two distinguishable frequencies to obtain a prediction model for time-domain signals. In the response signal processing unit, depending on the set number of data points, the signal will be processed simultaneously by the first frequency ω. a First ultrasonic signal and second frequency ω b The ultrasonic measurement signal at the current time step, obtained from the target structure excited by the second ultrasonic signal, is input into the trained LSTM network to obtain the ultrasonic prediction signal at the next time step, where ω a <ω b ; In the response signal processing unit, the obtained ultrasonic prediction signal is used to reconstruct the signal; In the response signal processing unit, the reconstructed signal is subjected to a Fourier transform; In the response signal processing unit, whether a crack has appeared in the target structure is determined by detecting the nonlinear modulation frequency component based on the value of the spectral density function obtained using the Fourier transform signal; as well as In the response signal processing unit, the damage index is calculated by substituting the spectral density function into the nonlinear modulation parameter equation; And the calculated damage index is used to detect fatigue cracks in the structure; The damage index β DN It uses equations To calculate; Where X N (ω) represents the reconstructed ultrasonic signal x N (t) The signal after Fourier transform, * denotes complex conjugate, and E[] denotes the expectation operator; ω a +ω b It is the sum and frequency, ω b -ω a It is the difference frequency.

2. The method of claim 1, wherein an equation is used. To calculate the spectral density function P xN (ω).

3. The method according to claim 1, further comprising: The target structure is simultaneously excited by generating the first and second ultrasonic signals by the excitation unit and applying them to the first and second excitation elements attached to the target structure, respectively; and the vibration of the target structure is detected by excitation of the first and second ultrasonic signals by the vibration detection element attached to the target structure, and the corresponding ultrasonic measurement signal is provided to the response signal processing unit.

4. The method of claim 1, wherein the prediction model of the trained LSTM network is designed and trained to predict the ultrasonic signal at the next time step by learning the nonlinear modulation frequency component, which is a fatigue crack-induced pattern on the full time series data of the ultrasonic measurement signal, in a manner that updates the previous cell state to a new cell state by partially adding new information provided by the input gate while forgetting some of the previous cell state information.

5. The method of claim 1, wherein the reconstructed signal is a signal reconstructed using only the ultrasonic prediction signal at the next time step.

6. The method of claim 5, wherein the number of data points of the reconstructed signal is determined by multiplying the number of data points of the ultrasonic measurement signal by a data reduction rate α having a predetermined size, and the data reduction rate α is determined in the range of 0 to 1.

7. The method of claim 1, wherein the reconstructed signal is a signal reconstructed by combining the ultrasonic measurement signal at the current time step and the ultrasonic prediction signal at the next time step.

8. The method of claim 1, wherein the Fourier transform of the reconstructed signal is a signal in the frequency domain, in which noise is reduced and the information of the ultrasonic measurement signal is enhanced.

9. A system for detecting fatigue cracks in a structure, comprising: A first excitation element is attached to a first region of the target structure and configured to input a first frequency ω. a The target structure is excited by vibration when the first ultrasonic signal is received. A second excitation element is attached to the first region of the target structure and configured to input a second frequency ω. b The target structure is excited by vibration when the second ultrasonic signal is received, where ω a <ω b ; An excitation unit is configured to generate the first ultrasonic signal and the second ultrasonic signal, and simultaneously provide them to the first excitation element and the second excitation element, respectively. A vibration detection element, attached to a second region of the target structure spaced apart from the first region and configured to detect vibration of the target structure based on excitation by a first ultrasonic signal and a second ultrasonic signal for generating corresponding ultrasonic measurement signals; and A response signal processing unit, comprising an operation processing unit configured to perform the following functions: training an LSTM network using ultrasonic measurement signals measured from a structure simultaneously excited by ultrasonic signals having two distinguishable frequencies, to obtain a prediction model for time-domain signals; and, depending on the set number of data points, .... a First ultrasonic signal and second frequency ω b The ultrasonic measurement signal measured at the current time step from the target structure excited by the second ultrasonic signal is input into a trained LSTM network to obtain the ultrasonic prediction signal at the next time step, where ω a <ω b The signal is reconstructed using the obtained ultrasonic prediction signal; in the response signal processing unit, a Fourier transform is performed on the reconstructed signal; and whether a crack has occurred in the target structure is determined by detecting a nonlinear modulation frequency component based on the value of the spectral density function obtained using the Fourier transform signal; wherein the response signal processing unit is further configured to calculate a damage index by substituting the spectral density function into a nonlinear modulation parameter equation. And the calculated damage index is used to detect fatigue cracks in the structure; The damage index β DN It uses equations To calculate; Where X N (ω) represents the reconstructed ultrasonic signal x N (t) The signal after Fourier transform, * denotes complex conjugate, and E[] denotes the expectation operator; ω a +ω b It is the sum and frequency, ω b -ω a It is the difference frequency.

10. The system of claim 9, wherein the response signal processing unit further comprises a digitization unit that converts the analog measurement signal measured by the vibration detection element into a digital measurement signal and provides the converted digital measurement signal to the operation processing unit.

11. The system according to claim 9, wherein the first excitation element, the second excitation element, and the vibration detection element are composed of piezoelectric elements.

12. The system of claim 9, wherein the prediction model of the trained LSTM network is designed and trained to predict the ultrasonic signal at the next time step by learning the nonlinear modulation frequency component, which is a fatigue crack-induced pattern on the entire time series data of the ultrasonic measurement signal, in a manner that updates the previous cell state to a new cell state by partially adding new information provided by the input gate while forgetting some of the previous cell state information.

13. The system of claim 9, wherein the reconstructed signal is a signal reconstructed using only the ultrasonic prediction signal at the next time step.

14. The system of claim 13, wherein the number of data points of the reconstructed signal is determined by multiplying the number of data points of the ultrasonic measurement signal by a data reduction rate α having a predetermined size, and the data reduction rate α is determined in the range of 0 to 1.

15. The system of claim 9, wherein the reconstructed signal is a signal reconstructed by combining the ultrasonic measurement signal at the current time step and the ultrasonic prediction signal at the next time step.

16. A computer-readable storage medium storing a computer program configured to, when executed by a processor of a computer device, cause the processor of the computer device to calculate information about whether a crack has occurred in a target structure by processing an ultrasonic measurement signal, the ultrasonic measurement signal being generated by detecting vibrations caused by ultrasonic excitation of the target structure and provided as input. The computer program includes the following functions: training an LSTM network using ultrasonic measurement signals measured from a structure simultaneously excited by ultrasonic signals having two distinguishable frequencies to obtain a prediction model for time-domain signals; and, depending on the set number of data points, training an LSTM network using ultrasonic measurement signals measured from a structure simultaneously excited by ultrasonic signals having two distinguishable frequencies ω. a First ultrasonic signal and second frequency ω b The ultrasonic measurement signal measured at the current time step from the target structure excited by the second ultrasonic signal is input into the trained LSTM network to obtain the ultrasonic prediction signal at the next time step, where ω a <ω b The obtained ultrasonic prediction signal is used to reconstruct the signal; in the response signal processing unit, the reconstructed signal is subjected to Fourier transform; whether a crack has appeared in the target structure is determined by detecting the nonlinear modulation frequency component based on the value of the spectral density function obtained using the Fourier transform signal; the damage index is calculated by substituting the spectral density function into the nonlinear modulation parameter equation. And the calculated damage index is used to detect fatigue cracks in the structure; The damage index β DN It uses equations To calculate; Where X N (ω) represents the reconstructed ultrasonic signal x N (t) The signal after Fourier transform, * denotes complex conjugate, and E[] denotes the expectation operator; ω a +ω b It is the sum and frequency, ω b -ω a It is the difference frequency.

17. A computer-executable program stored in a computer-readable storage medium to calculate information about whether a crack has appeared in a target structure by processing an ultrasonic measurement signal, the ultrasonic measurement signal being generated by detecting vibrations caused by ultrasonic excitation of the target structure and provided as input. The computer program includes the following functions: training an LSTM network using ultrasonic measurement signals measured from a structure simultaneously excited by ultrasonic signals having two distinguishable frequencies to obtain a prediction model for time-domain signals; and, depending on the set number of data points, training an LSTM network using ultrasonic measurement signals measured from a structure simultaneously excited by ultrasonic signals having two distinguishable frequencies ω. a First ultrasonic signal and second frequency ω b The ultrasonic measurement signal measured at the current time step from the target structure excited by the second ultrasonic signal is input into the trained LSTM network to obtain the ultrasonic prediction signal at the next time step, where ω a <ω b The signal is reconstructed using the obtained ultrasonic prediction signal; in the response signal processing unit, the reconstructed signal is subjected to a Fourier transform; whether a crack has appeared in the target structure is determined by detecting the nonlinear modulation frequency component based on the value of the spectral density function obtained using the Fourier transform signal; and the damage index is calculated by substituting the spectral density function into the nonlinear modulation parameter equation. And the calculated damage index is used to detect fatigue cracks in the structure; The damage index β DN It uses equations To calculate; Where X N (ω) represents the reconstructed ultrasonic signal x N (t) The signal after Fourier transform, * denotes complex conjugate, and E[] denotes the expectation operator; ω a +ω b It is the sum and frequency, ω b -ω a It is the difference frequency.

Images