WO2022146098A1 - 장단기 기억 네트워크 기반 스펙트럼 노이즈 감소 및 비선형 초음파 변조를 이용한 구조물의 피로 균열 검출 방법 및 이를 위한 시스템 - Google Patents
장단기 기억 네트워크 기반 스펙트럼 노이즈 감소 및 비선형 초음파 변조를 이용한 구조물의 피로 균열 검출 방법 및 이를 위한 시스템 Download PDFInfo
- Publication number
- WO2022146098A1 WO2022146098A1 PCT/KR2021/020341 KR2021020341W WO2022146098A1 WO 2022146098 A1 WO2022146098 A1 WO 2022146098A1 KR 2021020341 W KR2021020341 W KR 2021020341W WO 2022146098 A1 WO2022146098 A1 WO 2022146098A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- signal
- ultrasound
- ultrasonic
- reconstructed
- target structure
- Prior art date
Links
- 230000003595 spectral effect Effects 0.000 title claims abstract description 66
- 238000000034 method Methods 0.000 title claims abstract description 62
- 230000006403 short-term memory Effects 0.000 title claims abstract description 4
- 230000009467 reduction Effects 0.000 title claims description 36
- 238000005259 measurement Methods 0.000 claims abstract description 66
- 238000001514 detection method Methods 0.000 claims abstract description 58
- 230000006870 function Effects 0.000 claims abstract description 52
- 230000007787 long-term memory Effects 0.000 claims abstract 4
- 238000002604 ultrasonography Methods 0.000 claims description 111
- 230000004044 response Effects 0.000 claims description 58
- 238000012545 processing Methods 0.000 claims description 44
- 230000005284 excitation Effects 0.000 claims description 41
- 238000012549 training Methods 0.000 claims description 14
- 238000004590 computer program Methods 0.000 claims description 11
- 238000005336 cracking Methods 0.000 claims description 10
- 230000007774 longterm Effects 0.000 claims description 5
- 230000001131 transforming effect Effects 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 4
- 230000001965 increasing effect Effects 0.000 abstract description 6
- 210000004027 cell Anatomy 0.000 description 36
- 230000015654 memory Effects 0.000 description 15
- 238000001228 spectrum Methods 0.000 description 14
- 230000009021 linear effect Effects 0.000 description 12
- 229910052782 aluminium Inorganic materials 0.000 description 11
- XAGFODPZIPBFFR-UHFFFAOYSA-N aluminium Chemical compound [Al] XAGFODPZIPBFFR-UHFFFAOYSA-N 0.000 description 11
- 230000000694 effects Effects 0.000 description 8
- 238000012935 Averaging Methods 0.000 description 7
- 230000004913 activation Effects 0.000 description 4
- 238000005070 sampling Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 229910052751 metal Inorganic materials 0.000 description 2
- 239000002184 metal Substances 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 230000035945 sensitivity Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000001066 destructive effect Effects 0.000 description 1
- 230000035080 detection of muscle activity involved in regulation of muscle adaptation Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000006698 induction Effects 0.000 description 1
- 230000001939 inductive effect Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 239000006249 magnetic particle Substances 0.000 description 1
- 230000000737 periodic effect Effects 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
- 238000011410 subtraction method Methods 0.000 description 1
- 238000001931 thermography Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
- G01N29/44—Processing the detected response signal, e.g. electronic circuits specially adapted therefor
- G01N29/46—Processing the detected response signal, e.g. electronic circuits specially adapted therefor by spectral analysis, e.g. Fourier analysis or wavelet analysis
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
- G01N29/04—Analysing solids
- G01N29/043—Analysing solids in the interior, e.g. by shear waves
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
- G01N29/34—Generating the ultrasonic, sonic or infrasonic waves, e.g. electronic circuits specially adapted therefor
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
- G01N29/34—Generating the ultrasonic, sonic or infrasonic waves, e.g. electronic circuits specially adapted therefor
- G01N29/341—Generating the ultrasonic, sonic or infrasonic waves, e.g. electronic circuits specially adapted therefor with time characteristics
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
- G01N29/36—Detecting the response signal, e.g. electronic circuits specially adapted therefor
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
- G01N29/44—Processing the detected response signal, e.g. electronic circuits specially adapted therefor
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
- G01N29/44—Processing the detected response signal, e.g. electronic circuits specially adapted therefor
- G01N29/4445—Classification of defects
Definitions
- the present invention relates to the field of fatigue crack detection of structures, and more particularly, to a method and system for detecting fatigue cracks in structures using a long short-term memory (LSTM) network and nonlinear ultrasonic modulation technology.
- LSTM long short-term memory
- Fatigue cracking of metal structures is actually a very important problem. This is because fatigue cracks are a major cause of failures that cause breakage of metal structures and components. In general, fatigue cracks often become noticeable only after cracks have reached about 80% of the overall fatigue life of the structure. Early detection of fatigue cracks is important to avoid catastrophic failure.
- NDE non-destructive evaluation
- SHM structural state monitoring
- the nonlinear ultrasonic method is promising, the nonlinear corresponding response is quite weak, so the nonlinear characteristic is vulnerable to noise. Therefore, it is difficult to extract nonlinear characteristics in noisy conditions using the spectral density function. In particular, this is because noise overlaps with non-linear characteristics in the spectral domain. Therefore, it is important to reduce the spectral noise, and there is a great benefit in improving the fatigue crack detection performance based on the nonlinear characteristics.
- a spectral subtraction method was used to reduce spectral noise.
- the noise spectrum is usually estimated and updated in the absence of a signal.
- spectral subtraction can result in the power spectrum being assumed to be negative. Also, since there is no signal-free period in the ultrasonic signal, this method is not suitable for the ultrasonic signal.
- Zero-padding is a widely used method for spectral noise reduction. Zero padding is a way to extrapolate a signal to zero. However, since the zero extrapolated signal cannot retain the information of the signal, the non-linear characteristics of the signal are also lost.
- the response signal processing unit in the response signal processing unit, ultrasonic measurement measured from a structure that is simultaneously excited with two ultrasonic signals each having two distinct frequencies training the LSTM network using the signals for learning to secure a predictive model for the time domain signal;
- a target structure that is simultaneously excited with a first ultrasound signal of a first frequency ( ⁇ a ) and a second ultrasound signal of a second frequency ( ⁇ b ) (where ⁇ a ⁇ ⁇ b ) inputting the ultrasound measurement signal of the current time stage measured from the LSTM to the trained LSTM network to obtain an ultrasound prediction signal of the next time stage according to the set number of data points; obtaining a reconstructed signal using the obtained ultrasound prediction signal; Fourier transforming the reconstructed signal in the response signal processing unit; and detecting, in the response signal processing unit, a nonlinear modulation component based on a value of a spectral density function obtained using
- the method for detecting fatigue cracks in the structure includes: calculating, in the response signal processing unit, a damage index by substituting the spectral density function into a nonlinear modulation parameter expression; and detecting a fatigue crack of the structure using the calculated damage index.
- the spectral density function P xN ( ⁇ ) is can be obtained using
- X N ( ⁇ ) represents a Fourier transform signal of the reconstructed ultrasound signal x N (t)
- * represents a complex conjugate
- E[] represents an expectation operator.
- the damage index, ⁇ DN is can be saved with
- a vibrator in the method for detecting uniform fatigue of the structure, generates the first and second ultrasound signals and applies them to the first and second excitation devices attached to the target structure, respectively, to simultaneously apply the first and second ultrasound signals to the target structure. to with; and sensing the vibration of the target structure according to the excitation of the first and second ultrasound signals with a vibration detection device attached to the target structure and providing a corresponding ultrasound measurement signal to the response signal processing unit.
- the predictive model of the trained LSTM network updates a previous cell state to a new cell state by forgetting a part of the previous cell state information and partially adding new information provided from an input gate.
- it may be designed and trained to obtain the ultrasound prediction signal of the next time step by learning the nonlinear modulation frequency component, which is a pattern induced by fatigue cracks, over the entire time series data of the measured ultrasound signal.
- the reconstructed signal may be a signal reconstructed using only the ultrasound prediction signals of the next time step.
- the number of data points of the reconstructed signal is determined by multiplying the number of data points of the ultrasound measurement signal by a data reduction rate ( ⁇ ) of a predetermined size, and the data reduction rate ( ⁇ ) is 0 to 1 It may be determined in the following range.
- the reconstructed signal may be a signal reconstructed by combining the ultrasound measurement signals of the current start stage and the ultrasound prediction signals of the next time stage.
- the Fourier transform signal of the reconstructed signal may be a signal of a frequency domain in which noise is reduced and information of the ultrasound measurement signal is enhanced.
- exemplary embodiments provide a fatigue crack detection system for a structure including a first excitation element, a second excitation element, an exciter, 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 is configured to excite the target structure by vibrating as a first ultrasound signal of a first frequency ⁇ a is input.
- the second excitation element is attached to the first region of the target structure, and vibrates as a second ultrasound signal of a second frequency ⁇ b (where ⁇ a ⁇ ⁇ b ) is input to excite the target structure.
- the vibrator is configured to generate the first ultrasound signal and the second ultrasound signal and simultaneously provide them to the first and second excitation devices, respectively.
- the vibration detection device is attached to a second area spaced apart from the first area of the target structure, and detects the vibration of the target structure according to the excitation of the first and second ultrasound signals to generate a corresponding ultrasound measurement signal.
- the response signal processing unit trains an LSTM network using ultrasonic measurement signals measured from a structure simultaneously excited by two ultrasonic signals having two distinct frequencies, respectively, for learning to secure a predictive model for time-domain signals function to; (ii) measured from a target structure that is simultaneously excited with a first ultrasound signal of a first frequency ( ⁇ a ) and a second ultrasound signal of a second frequency ( ⁇ b ) (provided that ⁇ a ⁇ ⁇ b ) a function of inputting the ultrasound measurement signal of the current time stage into the trained LSTM network and obtaining an ultrasound prediction signal of the next time stage according to the set number of data points; (iii) a function of obtaining a reconstructed signal using the obtained ultrasound prediction signal; and (iv) an arithmetic processing unit configured to detect a nonlinear modulation component based on a value of a spectral density function obtained using a Fourier-transformed signal to perform a function of determining whether a crack has occurred in the target structure
- the operation processing unit of the response signal processing unit includes a function of calculating a damage index by substituting the spectral density function into a nonlinear modulation parameter expression; And it may further include a function of detecting a fatigue crack of the structure using the calculated damage index.
- the response signal processing unit may further include a digitizing unit that converts the analog measurement signal measured by the vibration detection device into a digital measurement signal and provides it to the operation processing unit.
- the first and second vibrating elements and the vibration detecting element may be formed of a piezoelectric element.
- the predictive model of the trained LSTM network updates a previous cell state to a new cell state by forgetting a part of the previous cell state information and partially adding new information provided from an input gate.
- it may be designed and trained to obtain the ultrasound prediction signal of the next time step by learning the nonlinear modulation frequency component, which is a pattern induced by fatigue cracks, over the entire time series data of the measured ultrasound signal.
- the reconstructed signal may be a signal reconstructed using only the ultrasound prediction signals of the next time step.
- the number of data points of the reconstructed signal is determined by multiplying the number of data points of the ultrasound measurement signal by a data reduction rate ( ⁇ ) of a predetermined size, and the data reduction rate ( ⁇ ) is 0 to 1 It may be determined in the following range.
- the reconstructed signal may be a signal reconstructed by combining the ultrasound measurement signals of the current start stage and the ultrasound prediction signals of the next time stage.
- the present invention is designed and trained to predict the ultrasonic signal of the next time step by learning the nonlinear modulation frequency component, which is a pattern induced by fatigue cracks, over the entire time series data of the measured ultrasonic signal. It uses an LSTM network. In the signal reconstructed through the trained LSTM, the nonlinear modulation component required for fatigue crack detection is maintained and amplified at the original level, and the noise component is reduced. That is, using the trained LSTM network can significantly increase the SNR (up to 276%), effectively improving the fatigue crack detection performance.
- the trained LSTM network can generate a reconstructed signal using significantly fewer data points than the measured signal without reducing the modulation amplitude. Thereby, the amount of data to be processed can be reduced.
- the trained LSTM network can generate a reconstructed ultrasound signal using only 20% of the original data. At this time, the non-linear modulation amplitude obtained from the reconstructed signal is the same as the modulation amplitude of the original signal, so that the fatigue crack detection performance does not deteriorate despite the reduction in data amount.
- Fig. 1 (A) illustrates performing ultrasonic measurement of an intact structure using an ultrasound-based crack detection system according to an exemplary embodiment of the present invention, and (B) is a frequency response measured from the intact structure. indicates characteristics.
- FIG. 2 exemplifies nonlinear ultrasonic measurement of a damaged structure with cracks using the ultrasound-based crack detection system of FIG. 1, and (B) shows the frequency response characteristics measured from the damaged structure.
- FIG. 3 is a schematic diagram showing a structure (A) of an LSTM network and a structure (B) of a memory cell used in a fatigue crack detection method according to an exemplary embodiment of the present invention.
- FIG. 4 is a flowchart schematically illustrating a spectral noise and data reduction algorithm for fatigue crack detection of a structure according to an exemplary embodiment of the present invention.
- FIG. 5 illustrates a time-series ultrasonic measurement signal (A) and a signal (B) representing the signal in the frequency spectrum domain.
- FIG. 6 shows training an LSTM network prediction model according to an exemplary embodiment of the present invention.
- Fig. 7 schematically illustrates reconstruction of an ultrasound signal using a trained LSTM network according to an exemplary embodiment.
- FIG. 10 illustrates a result of comparing the spectral density values before (A) and after (B) applying the LSTM network-based spectral noise and data reduction method according to an exemplary embodiment to a 'fatigue crack-free' aluminum sheet specimen do.
- 11 and 12 show the data reduction effect obtained when the trained LSTM network-based spectral noise and data reduction method according to an exemplary embodiment of the present invention is applied to an aluminum sheet specimen with fatigue cracks.
- FIG. 13 shows the results of evaluating the performance of a trained LSTM network-based spectral noise and data reduction method according to an exemplary embodiment of the present invention using an aluminum sheet specimen.
- FIG. 1A illustrates performing ultrasonic measurement on an intact structure 50 using an ultrasound-based crack detection system 10 according to an exemplary embodiment of the present invention, and (B) illustrates the structure The frequency response characteristic measured from (50) is shown.
- FIG. 2A illustrates that nonlinear ultrasonic measurement of the damaged structure 60 with cracks is performed using the ultrasonic measurement system 10 of FIG. 1 , and (B) is measured from the damaged structure 60 shows the frequency response characteristics.
- the ultrasonic-based crack detection system 10 includes a vibrator 20 , a response signal processing unit 30 , first and second excitation elements 42 and 44 , and a vibration detection element 46 .
- a vibrator 20 As shown in FIG. 1 , the ultrasonic-based crack detection system 10 includes a vibrator 20 , a response signal processing unit 30 , first and second excitation elements 42 and 44 , and a vibration detection element 46 .
- the vibrator 20 may be configured to generate and provide an excitation signal for excitation of the structure 50 by making the first and second excitation elements 42 and 44 vibrate.
- the vibrator 20 may include a waveform generator capable of generating any periodic waveform having a predetermined frequency.
- the vibrator 20 may generate a low-frequency ultrasound signal LF( ⁇ a ) and a high-frequency ultrasound signal HF( ⁇ b ) as an excitation signal and provide it to the first and second excitation elements 42 and 44 , respectively.
- the first and second excitation elements 42 and 44 may be attached close to each other in the first region of the non-damaged structure 50 , and the vibration detection element 46 is attached to the second region of the structure 50 .
- the first region and the second region may be located at opposite ends of the crack detection target region.
- the vibrator 20 is connected to the first and second excitation elements 42 and 44 .
- the first and second excitation elements 42 and 44 may be formed of, for example, piezoelectric elements.
- the first and second excitation elements ( 42 , 44 can excite the structure 50 by oscillating at a low frequency ⁇ a and a high frequency ⁇ b . Accordingly, the structure 50 is subjected to ultrasonic vibrations of a low frequency ( ⁇ a ) and a high frequency ( ⁇ b ).
- the vibration detecting element 46 may also be formed of, for example, a piezoelectric element. While the structure 50 vibrates by the excitation of the first and second excitation elements 42 and 44 , the vibration may be transmitted to the vibration detection site 46 .
- the vibration detection device 46 may sense the vibration of the structure 50 and generate electrical signals corresponding thereto.
- the vibration detection device 46 may sense the vibration of the structure 50 and output low-frequency and high-frequency analog signals.
- the amplitude of the two analog response signals can be set to a peak-to-peak voltage of, for example, 16V.
- the response signal processing unit 30 may be connected to the vibration detection device 46 .
- the response signal processing unit 30 receives an analog response signal corresponding to the vibration of the structure 50 detected by the vibration detection device 46 , performs predetermined processing, and obtains a frequency response of the structure 50 to obtain a frequency response of the structure 50 .
- ) can be configured to calculate information on whether cracks have occurred.
- the response signal processing unit 30 may be configured to process an analog response signal corresponding to the vibration detected by the vibration detection device 46 to calculate a damage index of the structure 50 .
- the response signal processing unit 30 receives the digitizing unit 32 that converts the output analog signal of the vibration detection device 46 into a digital signal and the converted digital signal, performs a predetermined operation, and performs a frequency It may include an arithmetic processing unit 34 for calculating a response, a damage index of the structure 50, and the like.
- the digitizing unit 32 may convert, for example, at a sampling rate of 1 MHz for 0.1 second.
- the response signal processing unit 30 may be implemented as a hardware component, a software component, and/or a combination of a hardware component and a software component.
- the arithmetic processing unit 34 may be implemented as a computer program and a computing device capable of executing a computer program and performing a predetermined operation indicated by the computer program or providing a function.
- the arithmetic processing unit 34 can execute a computer program to train to predict the measurement signal of the next time step through LSTM network-based learning, which will be described later, and can convert the time domain signal into a frequency domain signal, , may be a computing device configured to calculate a damage index by executing a computer program implementing a spectral noise reduction algorithm to be described later.
- the computing device may be, for example, a processor, controller, arithmetic logic unit (ALU), digital signal processor, microcomputer, field programmable array (FPA), programmable logic unit (PLU), microprocessor, or instruction manual. It may be implemented using one or more general purpose computers or special purpose computers, such as any other device capable of executing and responding to instructions.
- the computing device may further include computing resources such as memory, data storage, input/output units, and the like.
- the computer program implemented to perform the function of the response signal processing unit 30, the LSTM network model, and its training data may be stored in one or more computer-readable recording media.
- the method according to the exemplary embodiment of the present invention may be implemented in the form of program instructions that can be executed by various computer means and recorded in a computer-readable medium.
- the computer-readable recording medium may include program instructions, data files, data structures, etc. alone or in combination.
- the program instructions recorded on the recording medium may be specially designed and configured for the embodiment, or may be known and used by those skilled in the art of computer software.
- the exciter 20 is an ultrasonic signal of two distinct frequencies, for example, a low frequency ( ⁇ a ) and a high frequency ( ⁇ b ) (provided that ⁇ a ⁇ ⁇ b )
- ⁇ a low frequency
- ⁇ b high frequency
- the signal obtained by measuring the vibration of the structure 50 by the vibration detecting device 46 may be an analog signal in the time domain.
- the response signal processing unit 30 processes the analog measurement signal provided through the vibration detection element 46 and sees it in the spectral domain, the two frequencies ⁇ a of the input signal that excite the structure 50 and Only ( ⁇ b ) is the response (A, B) measured. That is, ultrasonic input signals (LF( ⁇ a ) and HF( ⁇ b )) of two distinct frequencies ( ⁇ a and ⁇ b , where ⁇ a ⁇ ⁇ b ) are simultaneously applied to the undamaged structure 50 without cracks.
- the frequency response (A, B) due to the vibration of the intact structure 50 is observed only at the two input frequencies ( ⁇ a and ⁇ b ) (refer to (B) of FIG. 1 ).
- Nonlinear ultrasonic modulation is induced by Referring to FIG. 2 , as in the case of FIG.
- the vibrator 20 transmits two ultrasonic excitation signals LF( ⁇ a ) and HF( ⁇ b ) of a low frequency ( ⁇ a ) and a high frequency ( ⁇ b ) to the first and the second excitation elements 42 and 44, respectively, to simultaneously excite the damaged (non-linear) structure 60, and at the same time measure the vibration response of the damaged structure 60 through the vibration detection element 46 If it is, an analog response signal in the time domain as shown in FIG. 2B can be obtained. When the analog response signal is converted into the frequency domain and the response in the spectral domain is viewed, it appears different from the response of the intact structure 50 .
- the response signal component of the structure 60 in the spectral domain includes the response component A and B at two input frequencies ⁇ a and ⁇ b as well as its Modulation components (M o , M s ) at the sum frequency ( ⁇ b + ⁇ a ) and the difference frequency ( ⁇ b - ⁇ a ), which are the modulation frequencies of the two input frequencies, are also included. That is, the amplitude of the high-frequency ultrasonic input signal is modulated by the crack opening and closing mechanism due to the low-frequency ultrasonic input. Amplitude modulation produces additional frequency components (non-linear modulation components) at the sum and difference ( ⁇ a ⁇ b ) of the input frequencies.
- This phenomenon is called 'nonlinear ultrasonic modulation'. Since this phenomenon occurs only when a nonlinear feature exists, it can be viewed as an index of damage to the structure. Therefore, it is possible to identify the presence of fatigue cracks by finding and extracting modulation components in the frequency spectrum domain.
- an exemplary embodiment of the present invention proposes a method for effectively detecting damage to a structure by reducing spectral noise using an LSTM network and a Fourier transform.
- LSTM networks are effectively used for time-series prediction of sequential data such as time-domain signals.
- LSTM networks are a special kind of Recurrent Neural Network (RNN) architecture that can be applied to sequential data.
- RNN Recurrent Neural Network
- typical RNNs have problems with vanishing gradients, so it is often difficult to learn long-term dependencies of data that are temporally sequential.
- LSTM networks have been developed. LSTM networks use blocks of memory cells that can represent long-term dependencies in time series data. That is, the LSTM network can control how much data from the distant past is reflected (remembered) by adding a cell state gate (C) to the existing RNN.
- C cell state gate
- 3A and 3B schematically show a structure of an LSTM network and a structure of a memory cell used in a fatigue crack detection method according to an exemplary embodiment of the present invention, respectively.
- the LSTM network 80 has a structure in which blocks of a plurality of memory cells 70 are sequentially connected.
- Each memory cell 70 may include an input gate (i t ), an output gate (o t ), a forgetting gate (f t ), and a self-circulating neuron. These gates i t , o t , f t control the interaction with adjacent memory cells.
- An LSTM network may include an input layer, a fully connected hidden layer, and an output layer.
- the hidden layer includes memory cells, associated gate portions, and a hidden portion, wherein the hidden portion provides an input to the gate portion and the memory cells.
- the input gate i t controls whether the input signal can modify the state of the memory cell 70 .
- the output gate o t controls whether the state of the other memory cell 70 can be modified.
- the forgetting gate f t may decide to forget or remember the previous state. That is, how much the previous memory cell state (C t-1 ) is reflected is controlled by the forgetting gate (f t ).
- An LSTM memory cell can be calculated as:
- 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 ) are the input weights, respectively. , bias, and recurrent weight, respectively.
- the superscripts f, i, o, and c denote forget gate, input gate, output gate, and self-recurrent neuron, respectively.
- X t is the sequential input data at the t-th time step.
- S t , C t , , ⁇ and ⁇ are the hidden state, the cell state, a new candidate value of that cell state, a sigmoid activation function and a hyperbolic tangent (tanh) activation function, respectively.
- Wow are the pointwise multiplication and pointwise addition operators, respectively.
- the first step in the LSTM network 80 is to select the information to be discarded from the cell state. This determination is performed by the forgetting gate (f t ) in equation (1).
- the second step is to select new information to be stored in the cell state.
- the input gate determines which input value should be updated.
- a new candidate value that can be added to the cell state using the hyperbolic tangent activation function ( ⁇ ) in equation (3) ) are obtained sequentially.
- the third step combines them, i.e., forgetting some of the previous cell state information (forgetting information we decided to forget at the previous time step) and partially adding new information provided from the input gate to the previous cell state (C t ). -1 ) to the new cell state (C t ).
- a key characteristic of the LSTM network 80 is the cell state C t passing through the top of the memory cell block 70 and linearly interacting with the gates, as shown in FIG. 3 .
- the LSTM network 80 may remove past information from the cell state C t or add new information to the cell state C t as it passes through these gates.
- Gates provide a way for information to selectively pass through the cell state.
- the sigmoid activation function ⁇ used for the gate in the memory cell 70 may adjust the amount of information passing through the sigmoid function while varying an output value between 0 and 1. When the output value is 0, it means that no information is passed, and when the output value is 1, it means that all information is passed.
- a spectral noise and data reduction technique may use signal prediction (reconstruction) characteristics of an LSTM network.
- LSTM networks have the advantage of learning long-term patterns in data. Since many data points in the time domain are obtained with a high sampling frequency, the ultrasound signal can be treated as long-term time-series data.
- an LSTM network can be built and trained to learn fatigue crack-induced patterns (nonlinear modulation frequency components) from the measured ultrasonic signals.
- fatigue crack induction patterns can be extracted across the entire time series data.
- the noise component has a random distribution in the data and the LSTM network does not learn the noise component.
- the noise component can be removed using the LSTM network, and the reconstructed signal can only contain meaningful components such as fatigue crack-inducing nonlinear modulation components.
- the amplitude of the nonlinear modulation component can be amplified, and the noise floor level in the spectral region can be lowered.
- the Fourier transform transforms data from the time domain to the frequency domain. That is, the Fourier transform can be used to analyze the measured time domain signal x(t) in the frequency domain.
- the spectral coefficient of the measurement signal x(t) is:
- a Fourier transform is derived by a spectral coefficient representing the weight value of a specific frequency for a signal when expressed as a Fourier series, and the formula for the Fourier transform is given as:
- the Fourier transform represents the information necessary to describe the measurement signal x(t) as a linear combination of sinusoidal signals at different frequencies. That is, the signal length T is related to the averaging when transforming x(t) to X( ⁇ ) through a Fourier transform as in equations (7 - 9). Therefore, if the length T of the measurement signal x(t) can be increased while maintaining the information of the measurement signal x(t) as it is, the spectral noise of the signal X( ⁇ ) in the frequency domain obtained by performing the Fourier transform will have a noise averaging effect ( This can be reduced through the effect of noise averaging.
- the Fourier transform signal of the signal provides an effect of increasing the number of averaging. Since the noise is randomly distributed, the noise decreases as the number of averaging increases.
- the LSTM network 80' trained through the learning of a large amount of training data is used in the time domain of the next time interval of the measurement signal of the current time interval. signal can be predicted. In building the LSTM network model, noise is not learned, but information included in the measurement signal is learned. Therefore, noise is not reflected in the prediction signal using the LSTM network, and only the learned information is reflected.
- the number of averaging increases and noise is reduced, but the information contained in the signal can be strengthened. Information on cracks can also be enhanced. Since the noise is reduced while the crack information contained in the signal is enhanced, the SNR can be increased to improve the crack detection capability.
- the spectral density function (power function) is calculated to extract the modulation component and is written as
- first-order modulations at the sum frequency and the difference frequency, ⁇ a ⁇ ⁇ b are considered as the damage index.
- the nonlinear modulation parameter ( ⁇ D ) is defined as
- This nonlinear modulation parameter ( ⁇ D ) can be used as a damage index because of its sensitivity to fatigue cracking.
- spectral noise can be reduced by using an LSTM network and a Fourier transform.
- An LSTM network can be trained to predict the signal x'(t) of the next time step of the measured signal of the current time step by learning the signal x(t) measured at the current time step. Through the training, a model can be obtained for the LSTM network that can predict the signal of the next time step.
- the signal length T of the measured signal x(t) is extended while only the information of the signal is maintained. Therefore, spectral noise is reduced when performing Fourier transform based on the effect of noise averaging.
- the signal-to-noise ratio may be improved due to the relationship between the signal length in the time domain and the spectral signal.
- the flowchart of Fig. 4 schematically shows a spectral noise and data reduction algorithm for fatigue crack detection of a structure according to an exemplary embodiment of the present invention.
- a spectral noise and data reduction technique for nonlinear ultrasound-based fatigue crack detection may use the LSTM network 80 .
- an ultrasound signal can be measured from the target structure 60 while excitation of the structure 60 using the excitation signals of low frequency LF ( ⁇ a ) and high frequency ( ⁇ b ) at the current time step ( S10).
- the measured ultrasound signal x(t m ) is a signal in the time domain and has a length of T (0 ⁇ t m ⁇ T), which is a time interval of the current time step.
- Fig. 5 (A) exemplifies time series ultrasonic data, and (B) exemplifies a signal in which this signal is expressed in the frequency spectrum domain.
- the ultrasonic measurement signal x(t) has N m data points.
- the ultrasonic measurement signal x(t) may include a vibration response component (ie, an input frequency component) of the target structure 60 and a random noise component, as illustrated in FIG. , since a fatigue crack is generated in the target structure 60, it may further include a nonlinear modulation component caused by the fatigue crack.
- the vibrator 20 transmits two excitation signals of a low frequency LF ( ⁇ a ) and a high frequency HF ( ⁇ b ) to the first and second excitation devices attached to the target structure 60 ( 42 and 44 are respectively applied to generate vibration to simultaneously excite the target structure 60 .
- the response corresponding to the excitation is measured using the vibration detection element 46 attached to the target structure 60 , and the measurement signal x(t m ) is provided to the response signal processing unit 30 .
- the measurement signal x(t m ) may be an ultrasound signal.
- a predictive model 80 by constructing an LSTM network and training that LSTM network using the original ultrasound signal x(N m f s ) to learn the basic pattern of the measured ultrasound signal without random noise patterns. It can be done (step S20).
- the input and output of an LSTM network can be set to adjacent data points in time series data. That is, an LSTM network can be designed and trained to predict the next time step using data from the current time step. 6 illustrates training the LSTM network prediction model 80' in this way.
- the signal of the next time step of the measurement signal x(t m ) of the current time step that is, the time T f (provided that T ⁇ t f
- the signal x'(t f ) up to ⁇ T f ) can be predicted.
- a trained LSTM network model can be built to predict the signal of the next time step when the measurement signal of the current time step is input.
- a new time-series ultrasound signal can be reconstructed by generating a prediction signal of an ultrasound signal in the time domain using the trained LSTM network model.
- 7 schematically illustrates reconstruction of an ultrasound signal using a trained LSTM network 80' according to an exemplary embodiment.
- the number of data points N f by inputting the ultrasound measurement signal x(t m ) of the current time step into the trained LSTM network model 80'.
- a prediction signal x'(t f ) of the next time step can be generated.
- the number of data points N f may be preset as an optimal value, and the optimal value may be obtained through testing.
- the following reconstructed ultrasound signal x N (t) is generated by combining the measurement signals of the current time step (x(t m )) and the prediction signals of the next time step (x′(t f )) It can be done (step S30).
- N is an indicator indicating a reconstructed ultrasound signal.
- the prediction signals x' It is also possible to obtain the reconstructed ultrasound signal (x N (t)) using only t f )).
- the length of the reconstructed ultrasound signal is as short as the length T of the measurement signal (x(t m )) compared to the previous embodiment. Since the measurement signal contains a lot of noise, the signal reconstructed only from the prediction signal without noise component can reduce noise more effectively than the reconstructed signal combining the measurement signal and the prediction signal.
- the reconstructed ultrasound signal x N (t) is Fourier transformed and analyzed in the frequency spectrum domain to find the nonlinear modulation frequency component caused by fatigue cracking (step S40).
- the noise component is effectively reduced in the frequency spectrum domain of the reconstructed ultrasound signal x N (t), so that the components corresponding to the two input frequencies and these frequencies Only the components corresponding to the modulation frequency of
- Equation (14) a Fourier transform may be performed on the reconstruction signal x N (t) using Equation (9), and a spectral density function may be calculated (S40).
- the Fourier transform equation for the reconstruction signal x N (t) is as shown in Equation (14) below.
- Equation (16) By substituting the spectral density function equation (15) of the combined signal x N (t) into the nonlinear modulation parameter equation (12), the damage index ( ⁇ DN ) equation expressed in equation (16) can be obtained.
- the damage index ( ⁇ DN ) of the structure 60 can be calculated using Equation (16).
- the obtained damage index ( ⁇ DN ) equation can be used to detect fatigue cracks.
- a signal-to-noise ratio may be calculated using the following equation.
- NF mean noise floor
- the effect of data reduction can be evaluated by estimating the number of data points (N f ) required to achieve a modulation amplitude level equal to the modulation amplitude level of the original measurement signal.
- the number of data points (N f ) can be obtained by the following equation.
- ⁇ represents a data reduction rate, and its value may be determined in the range of 0 or more and 1 or less.
- an LSTM network having a single hidden layer may be configured to avoid overfitting problems.
- An adaptive moment estimation (ADAM) optimizer may be used for training the LSTM network.
- the slope attenuation and square slope attenuation coefficients may be set, for example, to 0.9 and 0.999, respectively.
- the epsilon value and the initial learning rate for preventing division by zero may be set to 1.0e -8 and 1.5e -3 , respectively.
- the learning rate drop factor was set to 0.1 after half the training epoch.
- a Root Mean Square Error (RMSE) function can be used as a cost function.
- the construction, training and testing of the LSTM network can be performed in the MATLAB (R2019a) environment of a computer equipped with a processor (eg GPU) and RAM.
- a processor eg GPU
- 90% of the time domain ultrasound measurement signal can be used as a training data set, and the remaining 10% can be used as a validation data set.
- the spectral noise and data reduction algorithm can be implemented as a computer program and , may be stored in a computer-readable recording medium. Also, the computer program recorded on such a recording medium can be executed by the computing device of the response signal processing unit 30 .
- the performance of the LSTM network-based spectral noise and data reduction technique was experimentally verified.
- the first and second piezoelectric (PZT) transducers (corresponding to the first and second excitation elements 42 and 44) for excitation of ultrasonic signals to the aluminum plate specimen and the third piezoelectric transducer for detecting ultrasonic vibrations (vibration detection elements ( 46)) was installed on an aluminum plate specimen.
- a fatigue crack of 9 mm in length and 20 ⁇ m in width was generated by applying a fatigue load of 28,000 times to the aluminum plate specimen.
- the duration and peak-to-peak amplitude were set at 0.1 s and 12V.
- the input frequency was selected in consideration of the local resonance characteristics of the specimen and the superposition of the high-order harmonic component of the LF input and the nonlinear modulation component.
- the corresponding ultrasonic response detected through the third piezo transducer was obtained by a digitizer digitizing at a sampling rate of 1 MHz for 0.1 sec.
- the ultrasound response was averaged 5 times in the time domain to improve the SNR.
- FIG. 8 and 9 show the measurements before (A) and after (B) the spectral noise and data reduction method using the trained LSTM network according to an exemplary embodiment of the present invention is applied to the 'fatigue cracked' aluminum sheet specimen.
- the result of comparing the spectral density function (P x ( ⁇ )) of the signal is illustrated.
- Fig. 8 shows that N m is 100K (N f is set equal to N m to 100K for fair comparison), and the data acquisition time ( duration) T is 0.1 sec , f s is set to 1 MHz
- the variables are the same as in Fig. 8.
- the amplitudes P x ( ⁇ b + ⁇ a ) and P x ( ⁇ b ) of the spectral density functions at the modulation frequencies ( ⁇ b + ⁇ a ) and ( ⁇ b - ⁇ a ) - ⁇ a ) is denoted by A Ms and A Md .
- the values of the nonlinear modulation frequency components A Ms and A Md are 6.71e -7 and 8.03e -7 , respectively.
- 8B is a spectral density function of the signal reconstructed by the trained LSTM network.
- the values of A Ms and A Md were estimated to be 3.27e -6 and 2.99e -6 , respectively.
- the amplitude of the modulation frequency components in Fig. (B) was significantly strengthened compared to Fig. (A), and the NF value clearly decreased as Figs (A) and (B) were calculated as 4.6e -8 and 3.0e -9 respectively.
- P x ( ⁇ ) in the case of Fig. (A) and the case of Fig. (B) The SNRs were 34 dBW and 76 dBW, respectively. Therefore, the SNR was improved by 224%.
- the NF level is lowered in the figure (B) compared to the figure (A), but the frequency modulation component is not observed in both figures. That is, before the method of the present invention is applied, it is not easy to determine whether or not the modulation frequency component is included because the spectral density value contains a lot of noise components. It can be clearly seen that this does not exist.
- the LSTM network 80' model constructed according to the exemplary method of the present invention greatly reduces the noise component included in the spectral density value of the reconstructed ultrasound signal, and the nonlinear frequency modulation component is It can be seen that by strengthening the SNR, it is possible to clearly detect the modulation frequency components appearing on both sides of the input high frequency component.
- FIGS. 11 and 12 show the data reduction effect obtained when the trained LSTM network-based spectral noise and data reduction method according to an exemplary embodiment of the present invention is applied to an aluminum sheet specimen with fatigue cracks.
- the SNRs of the measured ultrasound signal and the reconstructed ultrasound signal were estimated to be 49dBW and 76dBW, respectively.
- FIG. 12 shows the result of performing a test similar to that of FIG. 11 using the 'measured ultrasound signal' with 145k data points.
- the SNRs of the measured ultrasound signal and the reconstructed ultrasound signal were estimated to be 38dBW and 56dBW, respectively.
- 11 and 12 show that a method according to an exemplary embodiment of the present invention can achieve similar frequency modulation amplitudes with reduced data points and noise.
- FIG. 13 shows the results of evaluating the performance of the trained LSTM network-based spectral noise and data reduction method using an aluminum plate specimen according to an exemplary embodiment of the present invention.
- Figure (A) shows the spectral noise reduction performance. For 60k and 80k data points, SNR improved by 238% and 276%, respectively.
- Figure (B) shows the data reduction performance. For 220k and 270k data points, the data reduction rate ( ⁇ ) was 0.18 and 0.19, respectively.
- the embodiments of the present invention it is possible to significantly improve the fatigue crack detection performance by significantly reducing the spectral noise through the spectral noise and data reduction method using the trained LSTM network.
- the signal reconstructed from the trained LSTM can reduce the noise component (SNR is increased by up to 276%) while maintaining and amplifying the nonlinear modulation component required for fatigue crack detection.
- the trained LSTM network can generate a reconstructed signal using significantly fewer data points than the measured signal without reducing the modulation amplitude, thus reducing the amount of data to be processed (using only 20% of the original data). to generate a reconstructed signal).
- the present invention can be utilized to detect cracks occurring in physical structures, structures, and the like.
Landscapes
- Physics & Mathematics (AREA)
- Immunology (AREA)
- Biochemistry (AREA)
- Pathology (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Analytical Chemistry (AREA)
- Chemical & Material Sciences (AREA)
- Engineering & Computer Science (AREA)
- Signal Processing (AREA)
- Mathematical Physics (AREA)
- Spectroscopy & Molecular Physics (AREA)
- Acoustics & Sound (AREA)
- Investigating Or Analyzing Materials By The Use Of Ultrasonic Waves (AREA)
Abstract
Description
Claims (20)
- 응답신호처리부에서, 서로 구별되는 두 가지 주파수를 각각 갖는 두 초음파신호로 동시에 가진되는 구조물로부터 계측된 초음파 계측신호들을 학습용으로 이용하여 장단기 기억(LSTM) 네트워크를 훈련시켜 시간영역 신호에 대한 예측 모델을 확보하는 단계;상기 응답신호처리부에서, 제1 주파수(ωa)의 제1 초음파신호와 제2 주파수(ωb)(단, ωa < ωb)의 제2 초음파 신호로 동시에 가진되는 대상 구조물(target structure)로부터 계측된 현재시간단계의 초음파 계측신호를 훈련된 LSTM 네트워크에 입력하여 설정된 데이터 포인트 수에 따라 다음시간단계의 초음파 예측신호를 구하는 단계;상기 응답신호처리부에서, 구해진 초음파 예측신호를 사용하여 재구성된 신호를 구하는 단계;상기 응답신호처리부에서, 상기 재구성된 신호를 푸리에 변환하는 단계; 및상기 응답신호처리부에서, 푸리에 변환된 신호를 이용하여 구한 스펙트럼 밀도함수의 값에 기초하여 비선형 변조 성분을 검출하여 상기 대상 구조물의 균열 발생 여부를 판별하는 단계를 포함하는 것을 특징으로 하는 구조물의 피로 균열 검출방법.
- 제1항에 있어서, 상기 응답신호처리부에서, 상기 스펙트럼 밀도함수를 비선형 변조 매개변수 식에 대입하여 손상지수를 산출하는 단계; 및 산출된 손상지수를 이용하여 상기 구조물의 피로 균열을 검출하는 단계를 더 포함하는 것을 특징으로 하는 구조물의 피로 균열 검출방법.
- 제1항에 있어서, 두 가진기가 상기 제1 및 제2 초음파신호를 생성하여 상기 대상 구조물에 부착된 제1 및 제2 가진소자에 각각 인가함으로써 상기 대상 구조물을 동시에 가진하는 단계; 및 상기 제1 및 제2 초음파신호의 가진에 따른 상기 대상 구조물의 진동을 상기 대상 구조물에 부착된 진동검출소자로 감지하여 대응하는 초음파 계측신호를 상기 응답신호처리부에 제공하는 단계를 더 포함하는 것을 특징으로 하는 것을 특징으로 하는 구조물의 피로 균열 검출방법.
- 제1항에 있어서, 상기 훈련된 LSTM 네트워크의 상기 예측 모델은, 이전 셀 상태 정보의 일부를 망각하고 입력 게이트로부터 제공되는 새로운 정보를 부분적으로 부가하여 이전 셀 상태를 새로운 셀 상태로 업데이트하는 방식으로, 계측 초음파 신호의 전체 시계열 데이터에 걸쳐 피로 균열로 유발된 패턴인 비선형 변조주파수 성분을 학습하는 것을 통해 다음시간단계의 초음파 예측신호를 구할 수 있도록 설계되고 훈련된 것을 특징으로 하는 구조물의 피로 균열 검출방법.
- 제1항에 있어서, 상기 재구성된 신호는 상기 다음시간단계의 초음파 예측신호들만으로 재구성한 신호인 것을 특징으로 하는 구조물의 피로 균열 검출방법.
- 제7항에 있어서, 상기 재구성된 신호의 데이터 포인트 개수는 상기 초음파 계측신호의 데이터 포인트 개수에 소정 크기의 데이터 감축률(α)을 곱한 값으로 정해지고, 상기 데이터 감축률(α)은 0이상 1이하의 범위에서 정해지는 것을 특징으로 하는 구조물의 피로 균열 검출방법.
- 제1항에 있어서, 상기 재구성된 신호는 현재시단단계의 초음파 계측신호들과 다음시간단계의 초음파 예측신호들을 결합하여 재구성한 신호인 것을 특징으로 하는 구조물의 피로 균열 검출방법.
- 제1항에 있어서, 상기 재구성된 신호의 푸리에 변환신호는 노이즈가 감소되고 초음파 계측신호의 정보는 강화된 주파수 영역의 신호인 것을 특징으로 하는 구조물의 피로 균열 검출방법.
- 대상 구조물의 제1 영역에 부착되고, 제1 주파수(ωa)의 제1 초음파신호가 입력됨에 따라 진동함으로써 상기 대상 구조물을 가진하도록 구성된 제1 가진소자;상기 대상 구조물의 제1 영역에 부착되고, 제2 주파수(ωb)(단, ωa < ωb)의 제2 초음파 신호가 입력됨에 따라 진동함으로써 상기 대상 구조물을 가진하도록 구성된 제2 가진소자;상기 제1 초음파신호와 상기 제2 초음파 신호를 생성하여 상기 제1 및 제2 가진소자에 각각 동시에 제공하도록 구성된 가진기;상기 대상 구조물의 제1 영역과 이격된 제2 영역에 부착되고, 상기 제1 및 제2 초음파신호의 가진에 따른 상기 대상 구조물의 진동을 감지하여 대응하는 초음파 계측신호를 생성하도록 구성된 진동검출소자; 및상기 진동검출소자가 생성한 초음파 계측신호를 처리하여 상기 대상 구조물의 균열 발생 여부에 관한 정보를 산출하도록 구성된 응답신호 처리부를 구비하고,상기 응답신호 처리부는 (i) 서로 구별되는 두 가지 주파수를 각각 갖는 두 초음파신호로 동시에 가진되는 구조물로부터 계측된 초음파 계측신호들을 학습용으로 이용하여 장단기 기억(LSTM) 네트워크를 훈련시켜 시간영역 신호에 대한 예측 모델을 확보하는 기능; (ii) 제1 주파수(ωa)의 제1 초음파신호와 제2 주파수(ωb)(단, ωa < ωb)의 제2 초음파 신호로 동시에 가진되는 대상 구조물(target structure)로부터 계측된 현재시간단계의 초음파 계측신호를 훈련된 LSTM 네트워크에 입력하여 설정된 데이터 포인트 수에 따라 다음시간단계의 초음파 예측신호를 구하는 기능; (iii) 구해진 초음파 예측신호를 사용하여 재구성된 신호를 구하는 기능; 및 (iv) 푸리에 변환된 신호를 이용하여 구한 스펙트럼 밀도함수의 값에 기초하여 비선형 변조 성분을 검출하여 상기 대상 구조물의 균열 발생 여부를 판별하는 기능을 수행하도록 구성된 연산처리부를 포함하는 것을 특징으로 하는 구조물의 피로 균열 검출 시스템.
- 제11항에 있어서, 상기 응답신호처리부의 상기 연산처리부는, 상기 스펙트럼 밀도함수를 비선형 변조 매개변수 식에 대입하여 손상지수를 산출하는 기능; 및 산출된 손상지수를 이용하여 상기 구조물의 피로 균열을 검출하는 기능을 더 포함하는 것을 특징으로 하는 구조물의 피로 균열 검출 시스템.
- 제11항에 있어서, 상기 응답신호처리부는 상기 진동검출소자가 계측한 아날로그 계측신호를 디지털 계측신호로 변환하여 상기 연산처리부로 제공하는 디지타이징부를 더 포함하는 것을 특징으로 하는 구조물의 피로 균열 검출 시스템.
- 제11항에 있어서, 상기 제1 및 제2 가진소자, 그리고 상기 진동검출소자는 압전소자로 구성된 것을 특징으로 하는 구조물의 피로 균열 검출 시스템.
- 제11항에 있어서, 상기 훈련된 LSTM 네트워크의 상기 예측 모델은, 이전 셀 상태 정보의 일부를 망각하고 입력 게이트로부터 제공되는 새로운 정보를 부분적으로 부가하여 이전 셀 상태를 새로운 셀 상태로 업데이트하는 방식으로, 계측 초음파 신호의 전체 시계열 데이터에 걸쳐 피로 균열로 유발된 패턴인 비선형 변조주파수 성분을 학습하는 것을 통해 다음시간단계의 초음파 예측신호를 구할 수 있도록 설계되고 훈련된 것을 특징으로 하는 구조물의 피로 균열 검출 시스템.
- 제11항에 있어서, 상기 재구성된 신호는 상기 다음시간단계의 초음파 예측신호들만으로 재구성한 신호인 것을 특징으로 하는 구조물의 피로 균열 검출 시스템.
- 제16항에 있어서, 상기 재구성된 신호의 데이터 포인트 개수는 상기 초음파 계측신호의 데이터 포인트 개수에 소정 크기의 데이터 감축률(α)을 곱한 값으로 정해지고, 상기 데이터 감축률(α)은 0이상 1이하의 범위에서 정해지는 것을 특징으로 하는 구조물의 피로 균열 검출 시스템.
- 제11항에 있어서, 상기 재구성된 신호는 현재시단단계의 초음파 계측신호들과 다음시간단계의 초음파 예측신호들을 결합하여 재구성한 신호인 것을 특징으로 하는 구조물의 피로 균열 검출 시스템.
- 컴퓨터 프로그램이 저장된 컴퓨터 판독 가능한 저장 매체로서,상기 컴퓨터 프로그램은, 컴퓨터 장치의 프로세서에 의해 실행될 때, 컴퓨터 장치로 하여금 대상 구조물의 초음파 가진에 따른 진동의 감지를 통해 생성되어 입력으로 제공되는 초음파 계측신호를 처리하여 상기 대상 구조물의 균열 발생 여부에 관한 정보를 산출하도록 구성되며, 상기 컴퓨터 프로그램은,(i) 서로 구별되는 두 가지 주파수를 각각 갖는 두 초음파신호로 동시에 가진되는 구조물로부터 계측된 초음파 계측신호들을 학습용으로 이용하여 장단기 기억(LSTM) 네트워크를 훈련시켜 시간영역 신호에 대한 예측 모델을 확보하는 기능; (ii) 제1 주파수(ωa)의 제1 초음파신호와 제2 주파수(ωb)(단, ωa < ωb)의 제2 초음파 신호로 동시에 가진되는 대상 구조물(target structure)로부터 계측된 현재시간단계의 초음파 계측신호를 훈련된 LSTM 네트워크에 입력하여 설정된 데이터 포인트 수에 따라 다음시간단계의 초음파 예측신호를 구하는 기능; (iii) 구해진 초음파 예측신호를 사용하여 재구성된 신호를 구하는 기능; (iv) 상기 재구성된 신호를 푸리에 변환하는 기능; 및 (v) 푸리에 변환된 신호를 이용하여 구한 스펙트럼 밀도함수의 값에 기초하여 비선형 변조 성분을 검출하여 상기 대상 구조물의 균열 발생 여부를 판별하는 기능을 포함하는 것을 특징으로 하는 컴퓨터 판독가능한 저장매체.
- 대상 구조물의 초음파 가진에 따른 진동의 감지를 통해 생성되어 입력으로 제공되는 초음파 계측신호를 처리하여 대상 구조물의 균열 발생 여부에 관한 정보를 산출하기 위해 컴퓨터 판독가능 저장 매체에 저장된 컴퓨터 실행가능 프로그램으로서,상기 컴퓨터 실행가능 프로그램은,(i) 서로 구별되는 두 가지 주파수를 각각 갖는 두 초음파신호로 동시에 가진되는 구조물로부터 계측된 초음파 계측신호들을 학습용으로 이용하여 장단기 기억(LSTM) 네트워크를 훈련시켜 시간영역 신호에 대한 예측 모델을 확보하는 기능; (ii) 제1 주파수(ωa)의 제1 초음파신호와 제2 주파수(ωb)(단, ωa < ωb)의 제2 초음파 신호로 동시에 가진되는 대상 구조물(target structure)로부터 계측된 현재시간단계의 초음파 계측신호를 훈련된 LSTM 네트워크에 입력하여 설정된 데이터 포인트 수에 따라 다음시간단계의 초음파 예측신호를 구하는 기능; (iii)구해진 초음파 예측신호를 사용하여 재구성된 신호를 구하는 기능; (iv) 상기 재구성된 신호를 푸리에 변환하는 기능; 및 (v) 푸리에 변환된 신호를 이용하여 구한 스펙트럼 밀도함수의 값에 기초하여 비선형 변조 성분을 검출하여 상기 대상 구조물의 균열 발생 여부를 판별하는 기능을 포함하는 것을 특징으로 하는 컴퓨터 실행가능 프로그램.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202180083951.8A CN116583747A (zh) | 2020-12-31 | 2021-12-31 | 使用基于长短期记忆网络的频谱降噪和非线性超声波调制检测结构中疲劳裂纹的方法及系统 |
US18/250,261 US20230384268A1 (en) | 2020-12-31 | 2021-12-31 | Method for detecting fatigue crack in structure using long short-term memory network-based spectral noise reduction and nonlinear ultrasonic modulation, and system therefor |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
KR20200189448 | 2020-12-31 | ||
KR10-2020-0189448 | 2020-12-31 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022146098A1 true WO2022146098A1 (ko) | 2022-07-07 |
Family
ID=82259534
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/KR2021/020341 WO2022146098A1 (ko) | 2020-12-31 | 2021-12-31 | 장단기 기억 네트워크 기반 스펙트럼 노이즈 감소 및 비선형 초음파 변조를 이용한 구조물의 피로 균열 검출 방법 및 이를 위한 시스템 |
Country Status (4)
Country | Link |
---|---|
US (1) | US20230384268A1 (ko) |
KR (1) | KR20220097296A (ko) |
CN (1) | CN116583747A (ko) |
WO (1) | WO2022146098A1 (ko) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116609442A (zh) * | 2023-07-17 | 2023-08-18 | 南京工业大学 | 基于非线性超声导波和深度学习的管道裂纹评估定位方法 |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR101693710B1 (ko) * | 2015-08-24 | 2017-01-17 | 세종대학교산학협력단 | 멤스 마이크로폰 어레이를 이용한 유도 초음파 영상화 장치 및 방법 |
KR101716877B1 (ko) * | 2016-06-09 | 2017-03-15 | 세종대학교산학협력단 | 자가 압전센싱 기반의 비선형 초음파 피로균열 검사 장치 및 방법 |
KR20190106773A (ko) * | 2018-03-06 | 2019-09-18 | 주식회사 엘지화학 | 배터리 팩의 균열을 진단하기 위한 장치와, 그것을 포함하는 배터리 팩 및 자동차 |
KR102061616B1 (ko) * | 2018-01-25 | 2020-01-02 | 한국과학기술원 | 비선형 스펙트럼 상관관계를 이용한 구조물 진단 방법 및 구조물 진단 시스템 |
KR102157903B1 (ko) * | 2019-05-10 | 2020-09-18 | 한국과학기술원 | 비선형 초음파 변조 기반 피로균열 파손 경고 시스템 및 그 방법과, 이를 이용한 구조물의 잔여 피로수명 추정 시스템 및 그 방법 |
-
2021
- 2021-12-29 KR KR1020210190639A patent/KR20220097296A/ko active IP Right Grant
- 2021-12-31 US US18/250,261 patent/US20230384268A1/en active Pending
- 2021-12-31 CN CN202180083951.8A patent/CN116583747A/zh active Pending
- 2021-12-31 WO PCT/KR2021/020341 patent/WO2022146098A1/ko active Application Filing
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR101693710B1 (ko) * | 2015-08-24 | 2017-01-17 | 세종대학교산학협력단 | 멤스 마이크로폰 어레이를 이용한 유도 초음파 영상화 장치 및 방법 |
KR101716877B1 (ko) * | 2016-06-09 | 2017-03-15 | 세종대학교산학협력단 | 자가 압전센싱 기반의 비선형 초음파 피로균열 검사 장치 및 방법 |
KR102061616B1 (ko) * | 2018-01-25 | 2020-01-02 | 한국과학기술원 | 비선형 스펙트럼 상관관계를 이용한 구조물 진단 방법 및 구조물 진단 시스템 |
KR20190106773A (ko) * | 2018-03-06 | 2019-09-18 | 주식회사 엘지화학 | 배터리 팩의 균열을 진단하기 위한 장치와, 그것을 포함하는 배터리 팩 및 자동차 |
KR102157903B1 (ko) * | 2019-05-10 | 2020-09-18 | 한국과학기술원 | 비선형 초음파 변조 기반 피로균열 파손 경고 시스템 및 그 방법과, 이를 이용한 구조물의 잔여 피로수명 추정 시스템 및 그 방법 |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116609442A (zh) * | 2023-07-17 | 2023-08-18 | 南京工业大学 | 基于非线性超声导波和深度学习的管道裂纹评估定位方法 |
CN116609442B (zh) * | 2023-07-17 | 2023-10-13 | 南京工业大学 | 基于非线性超声导波和深度学习的管道裂纹评估定位方法 |
Also Published As
Publication number | Publication date |
---|---|
US20230384268A1 (en) | 2023-11-30 |
KR20220097296A (ko) | 2022-07-07 |
CN116583747A (zh) | 2023-08-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Schubert | Numerical time-domain modeling of linear and nonlinear ultrasonic wave propagation using finite integration techniques––theory and applications | |
Samman et al. | Vibration testing for nondestructive evaluation of bridges. I: Theory | |
Loutridis et al. | Forced vibration behaviour and crack detection of cracked beams using instantaneous frequency | |
Koo et al. | The development of reactor coolant pump vibration monitoring and a diagnostic system in the nuclear power plant | |
Park et al. | An outlier analysis framework for impedance-based structural health monitoring | |
Saravanan et al. | Damage detection in structural element through propagating waves using radially weighted and factored RMS | |
WO2022146098A1 (ko) | 장단기 기억 네트워크 기반 스펙트럼 노이즈 감소 및 비선형 초음파 변조를 이용한 구조물의 피로 균열 검출 방법 및 이를 위한 시스템 | |
JP3958538B2 (ja) | コンクリート打音検査方法、及びコンクリート打音検査装置 | |
Joglekar et al. | Nonlinear analysis of flexural wave propagation through 1D waveguides with a breathing crack | |
KR101716877B1 (ko) | 자가 압전센싱 기반의 비선형 초음파 피로균열 검사 장치 및 방법 | |
Koh et al. | Effects of local stiffness changes and delamination on Lamb wave transmission using surface-mounted piezoelectric transducers | |
Dahlen et al. | Damage identification in concrete using impact non-linear reverberation spectroscopy | |
Li et al. | System identification-based frequency domain feature extraction for defect detection and characterization | |
JP2024052904A (ja) | 信号処理 | |
Gelman et al. | Novel vibration structural health monitoring technology for deep foundation piles by non‐stationary higher order frequency response function | |
Jin et al. | Single-impact nonlinear resonant acoustic spectroscopy for monitoring the progressive alkali–silica reaction in concrete | |
KR102157903B1 (ko) | 비선형 초음파 변조 기반 피로균열 파손 경고 시스템 및 그 방법과, 이를 이용한 구조물의 잔여 피로수명 추정 시스템 및 그 방법 | |
Niu et al. | Uncertainty quantification method for elastic wave tomography of concrete structure using interval analysis | |
Liu et al. | Deep learning-based solvability of underdetermined inverse problems in nonlinear ultrasonic characterization of micro damages | |
JP2000262523A (ja) | 消化器系活性モニタ | |
JPH08114638A (ja) | 機器異常診断装置 | |
KR101557270B1 (ko) | 스마트 구조물의 유지 관리를 위한 est 기반 단일 계측 시스템 | |
Zhou et al. | Damage identification for plate structures using physics-informed neural networks | |
Yang et al. | Vibration propagation identification of rotor-bearing-casing system using spatiotemporal graphical modeling | |
Spall et al. | Neural network approach to locating acoustic emission sources in nondestructive evaluation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 21915895 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 18250261 Country of ref document: US |
|
WWE | Wipo information: entry into national phase |
Ref document number: 202180083951.8 Country of ref document: CN |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21915895 Country of ref document: EP Kind code of ref document: A1 |