CN110879254A - Steel rail crack acoustic emission signal detection method based on improved least square generation type countermeasure network - Google Patents

Steel rail crack acoustic emission signal detection method based on improved least square generation type countermeasure network Download PDF

Info

Publication number
CN110879254A
CN110879254A CN201811029104.XA CN201811029104A CN110879254A CN 110879254 A CN110879254 A CN 110879254A CN 201811029104 A CN201811029104 A CN 201811029104A CN 110879254 A CN110879254 A CN 110879254A
Authority
CN
China
Prior art keywords
network
noise
signal
discriminator
sample
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201811029104.XA
Other languages
Chinese (zh)
Other versions
CN110879254B (en
Inventor
王康伟
章欣
王艳
韩瑞东
沈毅
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Institute of Technology
Original Assignee
Harbin Institute of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Harbin Institute of Technology filed Critical Harbin Institute of Technology
Priority to CN201811029104.XA priority Critical patent/CN110879254B/en
Publication of CN110879254A publication Critical patent/CN110879254A/en
Application granted granted Critical
Publication of CN110879254B publication Critical patent/CN110879254B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • G01N29/14—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 using acoustic emission techniques
    • 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
    • G01N29/4409—Processing the detected response signal, e.g. electronic circuits specially adapted therefor by comparison
    • G01N29/4418—Processing the detected response signal, e.g. electronic circuits specially adapted therefor by comparison with a model, e.g. best-fit, regression 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
    • G01N29/44—Processing the detected response signal, e.g. electronic circuits specially adapted therefor
    • G01N29/4481—Neural networks

Landscapes

  • Physics & Mathematics (AREA)
  • Analytical Chemistry (AREA)
  • Chemical & Material Sciences (AREA)
  • Pathology (AREA)
  • Immunology (AREA)
  • Biochemistry (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Signal Processing (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Acoustics & Sound (AREA)
  • Investigating Or Analyzing Materials By The Use Of Ultrasonic Waves (AREA)
  • Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)

Abstract

The invention discloses a steel rail crack acoustic emission signal detection method based on an improved least square generation type countermeasure network, which solves the problem of automatically filtering complex noise by training a noise model through an countermeasure neural network in different noise environments. The method comprises the following steps: firstly, synthesizing a noisy signal sample library under different noise backgrounds and normalizing. And secondly, initializing the generative countermeasure network. And thirdly, inputting the denoised sample and the reference sample into a discriminator network together and updating the weight of the discriminator network according to the discrimination error. And fourthly, updating and generating a network weight according to the loss value of the discriminator and the denoised mean square error. And fifthly, training the discrimination network and the generation network alternately, and finally realizing effective removal of noise by using a generator forward network. Compared with the prior art, the invention has the following advantages: 1) the same structure can be used for simultaneously suppressing multiple noises; 2) the noise model is automatically learned without any prior knowledge; 3) the crack acoustic emission signal can still be detected under the background of high speed and strong noise.

Description

Steel rail crack acoustic emission signal detection method based on improved least square generation type countermeasure network
Technical Field
The invention relates to a method in the field of high-speed railway steel rail crack signal detection, in particular to a steel rail crack acoustic emission signal detection method based on an improved least square generation type countermeasure network.
Background
At present, the four-longitudinal four-transverse high-speed rail nets in China enter a comprehensive collecting and managing stage, and eight-longitudinal eight-transverse high-speed rail net construction is already carried out. When the high-speed railway is developed rapidly, the high-speed and stable running of trains is related to the safety of people's lives and national property, and the safety of high-speed railways is more and more highly valued by people. The occurrence of delamination, pitting, cracking, breaking and other rail damage affecting the performance of rails in railways during long term use is increasing. In a high-speed railway system, the effects of long-term collision, extrusion and the like generated by a high-speed train are more prominent, the probability of crack occurrence and the crack propagation speed are improved, if the detection is not timely performed and safety measures are taken, the crack is easy to propagate under the external force of subsequent continuous action, and thus rail breakage, even train derailment and other major accidents are caused.
A large amount of acoustic emission phenomena can be generated when the steel rail cracks move, strain energy is released in the form of elastic waves and is propagated in the steel rail, the elastic waves can be collected by an acoustic emission sensor and recorded in an acoustic emission signal, and therefore a large amount of information directly related to the dynamic cracks can be read from the acoustic emission signal. The acoustic emission detection technology can be used for dynamic crack characteristic research, can detect crack initiation and expansion conditions by passive real-time monitoring of signals in the steel rail under the driving load, and is very suitable for evaluating the steel rail including the damage severity level and even the service life stage of the steel rail, so that corresponding maintenance measures can be taken as soon as possible.
However, since the acoustic emission detection technology has sensitive and passive characteristics, the acoustic emission collected signals are easily affected by noise. The signals acquired on the railway site usually have more complex noise components, and background noise is mainly caused by friction and abrasion accompanied by the mechanical interaction between the wheel and the rail, so that the more obvious stationarity and time sequence exist; there are also random noise anomalies caused by some unknown factors. The invention provides a steel rail crack acoustic emission signal detection method based on an improved least square generation type countermeasure network, which realizes the establishment of a complex noise model and the elimination of noise components based on a generation type countermeasure model (GAN) in deep learning, wherein the model comprises two parts: generator network (Generator, G) and Discriminator network (Discriminator, D). The method trains time sequence models aiming at different types of noise and forms a filter aiming at the same type of noise in a generator network, so as to play a role in filtering an original noise-containing signal, and judges whether the processed signal still contains a noise component or not through a discriminator network, so that the processed signal and the processed signal are alternately optimized in an iteration mode in opposition, and finally, the composite filtering effect on complex noise components is realized, and the filtering precision is further improved.
Disclosure of Invention
The invention provides a steel rail crack acoustic emission signal detection method based on an improved least square generation type countermeasure network. Under the condition that the crack signal is completely submerged by wheel track noise, the method can be combined with the proposed method, the purposes of inhibiting complex noise components in a railway field and detecting the occurrence of the crack acoustic emission signal can be realized, and further guidance is provided for the characteristic extraction and classification of the rail crack damage.
The invention is realized by the following technical scheme:
a rail crack acoustic emission signal detection method based on an improved least square generation type countermeasure network comprises the following steps: collecting specific wheel-track noise signals generated when a railway on-site train runs, keeping the initial amplitude value superposed with the reference crack signals, synthesizing noise-containing crack signals, normalizing the noise-containing signals to obtain normalized noise-containing signals z, giving out labels corresponding to the crack signals and mapping the labels into attribute labels x in One-hot coding formatcE.g. elastic or plastic phases correspond to [0,1]]Or [1,0 ]]Will label xcConcatenating after a normalized noisy signal z to obtain training samples (z, x)c) (ii) a Setting structural parameters, iteration times and initial values of a generator network (G) and a discriminator network (D); a batch of noisy signal samples (z, x)c) Inputting the filtered signal and a reference signal without noise into a discriminator network (D), calculating Mean Square Error (MSE) between the filtered signal and the reference signal without noise and the random gradient of a counterloss function, and updating weight matrix parameters of the discriminator (D) according to the Mean Square Error and the random gradient descending direction of the loss function; after the discriminator optimizes the k steps, the mean square error is usedUpdating the weight matrix parameter of a generator (G) in the random gradient descending direction of the loss value obtained by the discriminator for the noise signal identification in the countermeasure loss function; and repeating the steps, obtaining an optimal filtering structure or stopping after reaching the maximum iteration times through the alternating iteration and optimization of the optimal filtering structure and the optimal filtering structure, and for the trained GAN network, directly processing the test sample by adopting the forward network of the generator in the test process and then judging the denoising effect such as the signal-to-noise ratio after the processing.
The flow chart of the invention is shown in figure 1, and is divided into five steps, and the concrete steps are as follows:
the method comprises the following steps: and synthesizing and normalizing a noisy signal sample library under different noise backgrounds.
Collecting specific wheel-rail noise signal x generated during train running on railway site(i) noiseAnd maintaining the initial amplitude of the reference crack signal x for the same length of time(i)Superposing, synthesizing noise-containing crack signals, and preprocessing the noise-containing crack signals to obtain normalized noise-containing signals z(i)，z(i)＝(x(i)+x(i) noise)/max(abs(x(i)+x(i) noise))；
Where max (-) represents the maximum value of the vector, and abs (-) represents the absolute value of the vector element. i is the sequence number of the signal sample in the training set. Giving out the label corresponding to the crack signal and mapping the label to the attribute label x of One-hot coding format(i) cE.g. elastic or plastic phase corresponding to a label of [0,1]]Or [1,0 ]]Will label x(i) cSpliced to a normalized noisy signal z(i)Then n training samples (z, x) are obtainedc)n×lIn the same way, m test samples (z) are obtainedt，xct)m×lAnd l is the number of sampling points contained in the standard noise-containing signal.
Step two: an initialization process for a generative countermeasure network, comprising: setting initial values and structural parameters of a generator network (G) and a discriminator network (D): including the number of network layers, the number of nodes and the initial value of weight, training the minimum batch nbatchIteration number N, learning rate α, regularization weight lambda and the likeThe forward networks of the bright G and the D all adopt full connection layers, and the interlayer calculation formula is as follows:
Figure BDA0001789260810000031
wherein the content of the first and second substances,
Figure BDA0001789260810000032
is the output of the l-level j cell,
Figure BDA0001789260810000033
is the output of the l +1 level i cell,
Figure BDA0001789260810000034
the weight value of the connection l layer j unit and l +1 layer i unit in the G or D network,
Figure BDA0001789260810000035
is the bias of the l +1 layer i cells,
Figure BDA0001789260810000036
for the output of the i +1 layer i unit, f (·) represents an activation function, and a sigmoid function or a hyperbolic tangent function and the like can be selected, and the corresponding expression is as follows:
Figure BDA0001789260810000037
in addition, in the selection of the activation function, since the generator network is used for filtering, a structure of a noise reduction auto encoder (DAE) is adopted and the activation function of the output layer should select a hyperbolic tangent function to simulate the fluctuation of the output signal in a positive and negative range; and the discriminator network is used for judging whether the filtered signal still contains noise components, and a sigmoid function is needed to be adopted as an output layer function, so that the output value is easily compressed in the range of [0,1] and an approximate probability result of two classifications is given. The input-output relationship between the generator network (G) and the discriminator network (D) is as follows:
Figure BDA0001789260810000038
Figure BDA0001789260810000039
wherein D (-) is a mapping operator formed by the neural network of the discriminator, thetadThe weight matrix parameter of the discriminator network (D) is used for discriminating the de-noised signal and the reference signal and giving out a discrimination result
Figure BDA00017892608100000310
For back propagation and parameter updating; g (-) is a mapping operator formed by the neural network of the generator, θgTo generate weight matrix parameters for the network (G) of generators, which act on noisy inputs z(i)Predicting and filtering to obtain corresponding de-noising signal
Figure BDA00017892608100000311
The objective function in the iterative optimization process of the discriminator network (D) and the generator network (G) is as follows:
Figure BDA00017892608100000312
Figure BDA00017892608100000313
step three: and inputting the denoised sample and the reference sample into a discriminator network together and updating the weight of the discriminator network according to the discrimination error. Sampling a batch of noisy signals
Figure BDA00017892608100000314
Inputting the filtered sample and a reference sample without noise into a forward network of the generator, substituting the filtered sample and the reference sample without noise into a discriminator network (D), calculating Mean Square Error (MSE) between the filtered sample and the reference sample and the random gradient of a countercheck loss function, and updating a weight matrix parameter theta of the discriminator (D) according to the descending direction of the random gradient of the loss functiond：
Figure BDA00017892608100000315
Wherein, D (-) is the mapping operator of the discriminator neural network, and G (-) is the mapping operator of the generator neural network.
Step four: and updating and generating a network weight according to the loss value of the discriminator and the denoised mean square error.
After the discriminator optimizes the k steps, updating a generator (G) weight matrix parameter theta in the descending direction of the random gradient according to the loss value obtained by the discriminator identifying the noise signal in the mean square error and the antagonistic loss functiong：
Figure BDA0001789260810000041
Wherein λ is1To combat the regularizing weight of the loss function, λ2Is a regularized weight of mean square error, λ1+λ21. The mean square error is calculated from the vector two norm:
Figure BDA0001789260810000042
wherein
Figure BDA0001789260810000043
x(i)(k) Representing the ith reference crack signal vector x(i)The (k) th element of (a),
Figure BDA0001789260810000044
representing the ith denoised sample vector
Figure BDA0001789260810000045
The kth element of (1).
Step five: and training the discrimination network and the generation network alternately, and finally realizing effective removal of noise by using a generator forward network.
Repeating the third step and the fourth step, inputting all training set data repeatedly, and stopping after the maximum iteration number N is reached through the alternating iteration and optimization of the training set data and the training set data; for the trained GAN structure, the testing process directly adopts a forward network of a generator to process a test sample, then judges the Peak Signal to Noise Ratio (PSNR) after the processing and checks whether the precision requirement is met:
Figure BDA0001789260810000046
wherein the content of the first and second substances,
Figure BDA0001789260810000047
for the k-th element, x, in the filtered i-th signal(i)(k) Is the ith kth element of the reference crack signal.
Compared with the prior art, the invention has the following advantages:
1. the method can be used for filtering various types of noise after being trained by the same filtering structure, is suitable for actual railway field application with a more complex noise environment, does not need any priori knowledge, and only needs to ensure that a training process gives sufficient reference crack signals and different types of noise signals;
2. the denoising method provided by the invention can still achieve the purposes of suppressing noise and detecting crack signals under the condition that the crack signals of the steel rail are completely submerged in the noise signals.
Drawings
FIG. 1 is a flow chart of the present invention.
FIG. 2 is a waveform diagram of a synthesized noisy signal and a reference crack signal in the context of wheel track noise according to the present invention.
FIG. 3 is a waveform diagram of a synthesized noisy signal and a reference flaw signal under the background of white Gaussian noise in the present invention.
Fig. 4 is a schematic structural diagram of a generative countermeasure network employed in the present invention.
FIG. 5 shows a denoised signal and a corresponding reference crack signal after the wheel-track noise is eliminated.
FIG. 6 shows a denoised signal after eliminating Gaussian white noise and a corresponding reference crack signal.
Detailed Description
The following describes a specific implementation manner of the present invention with reference to an embodiment and the accompanying drawings, wherein the specific implementation process of the method for detecting a rail crack acoustic emission signal based on an improved least square generation type countermeasure network is as follows:
executing the step one: synthesizing and normalizing a noise-containing signal sample library under different noise backgrounds, and acquiring a specific wheel-rail noise signal x generated when a train runs on a railway field(i) noise1With white Gaussian noise x obtained by simulation(i) noise2And maintaining the initial amplitude of the reference crack signal x for the same length of time(i)Superposing, synthesizing noise-containing crack signals, preprocessing the two noise-containing crack signals to obtain a normalized noise-containing signal z1 (i)、z2 (i)，
z1 (i)＝(x(i)+x(i) noise1)/max(abs(x(i)+x(i) noise1))，z2 (i)＝(x(i)+x(i) noise2)/max(abs(x(i)+x(i) noise2))；
Where max (-) represents the maximum value of the vector, and abs (-) represents the absolute value of the vector element. i is the sequence number of the signal sample in the training set. Crack signals were collected from tensile tests on steel rail test pieces. In the experimental process, a label corresponding to a crack signal is given and is mapped into an attribute label x in One-hot coding format(i) cE.g. elastic or plastic phase corresponding to a label of [0,1]]Or [1,0 ]]Will label x(i) cSpliced to a normalized noisy signal z1 (i)、z2 (i)Then, n is 3000 training samples (z)1，xc)n×l、(z2，xc)n×lThe same principle is that m is 200 test samples (z)t1，xct)m×l、(zt2，xct)m×lWherein, l is 2048 is the number of sampling points contained in the standard noise-containing signal, the sampling frequency is 5MHz, and the corresponding time length of each section of signal is 409.6μ s. The synthesized signal containing the wheel-track noise and the reference signal are shown in fig. 2, and the signal containing the white gaussian noise and the reference signal are shown in fig. 3.
And (5) executing the step two: an initialization process for a generative countermeasure network, comprising:
(1) the method for setting the network depth and the structure parameters of the generator network (G) and the discriminator network (D) comprises the following steps: network layer number, node number and each node activation function: in structural design, because a generator network is used for filtering, a structure of a noise reduction self-Encoder (DAE) is adopted, namely the noise reduction self-Encoder (DAE) comprises an Encoder (Enencoder) and a Decoder (Decode), the forward networks of G and D all adopt full connection layers, and the interlayer calculation formula is as follows:
Figure BDA0001789260810000051
wherein the content of the first and second substances,
Figure BDA0001789260810000052
is the output of the l-level j cell,
Figure BDA0001789260810000053
is the output of the l +1 level i cell,
Figure BDA0001789260810000054
the weight value of the connection l layer j unit and l +1 layer i unit in the G or D network,
Figure BDA0001789260810000061
is the bias of the l +1 layer i cells,
Figure BDA0001789260810000062
for the output of the i +1 layer i unit, f (·) represents an activation function, and a sigmoid function or a hyperbolic tangent function and the like can be selected, and the corresponding expression is as follows:
Figure BDA0001789260810000063
in the selection of the activation function, the activation function of the generator output layer should select a hyperbolic tangent function to simulate the fluctuation of an output signal in a positive and negative range; the discriminator network is used for judging whether the filtered signal still contains noise components, and a sigmoid function is needed to be adopted as an output layer function, so that the output value is easily compressed in the range of [0,1] and an approximate probability result of two classifications is given; to speed up the convergence process of the hidden layer, the hidden layer activation function selects the ReLu function. The neural network structure adopted in the present embodiment is as shown in fig. 4, and the input-output relationship between the generator network (G) and the discriminator network (D) is as follows:
Figure BDA0001789260810000064
Figure BDA0001789260810000065
wherein D (-) is the mapping operator of the discriminator neural network, thetadThe weight matrix parameter of the discriminator network (D) is used for discriminating the de-noised signal and the reference signal and giving out a discrimination result
Figure BDA0001789260810000066
For back propagation and parameter updating; g (-) is the mapping operator of the generator neural network, θgTo generate weight matrix parameters for the network (G) of generators, which act on noisy inputs z(i)Predicting and filtering to obtain corresponding de-noising signal
Figure BDA0001789260810000067
The objective function in the iterative optimization process of the discriminator network (D) and the generator network (G) is as follows:
Figure BDA0001789260810000068
Figure BDA0001789260810000069
(2) networkSetting iteration parameters and initial values of each layer: including training a minimum batch nbatch100, 500 iteration number N, 1 × 10 learning rate α-4Regularization weight λ1＝0.1、λ20.9, etc. The initial value of the weight is initialized randomly, and the gradient of the network weight and the bias gradient are both set to be 0.
And step three is executed: and inputting the denoised sample and the reference sample into a discriminator network together and updating the weight of the discriminator network according to the discrimination error. Sampling a batch of noisy signals
Figure BDA00017892608100000610
Inputting the filtered and denoised sample and a noise-free reference sample into a discriminator network (D), calculating Mean Square Error (MSE) between the filtered and denoised sample and the noise-free reference sample and random gradient of a counterloss function, and updating a weight matrix parameter theta of the discriminator network (D) according to the Mean square error and the random gradient descending direction of the loss functiond：
Figure BDA00017892608100000611
The mean square error is calculated from the vector two norm:
Figure BDA0001789260810000071
wherein x is(i)(k) Representing the ith sample vector x(i)The kth element of (1).
And step four is executed: and updating and generating a network weight according to the loss value of the discriminator and the denoised mean square error. After the arbiter optimizes k to 1 step, i.e. after inputting k training samples to the arbiter, according to the mean square error between the denoised sample and the reference sample and the loss value obtained by the arbiter identifying the noise signal in the function of resisting loss, the weighted combination of the two is propagated reversely between the layers of G, and the weight matrix parameter theta of the generator network (G) is updated in the direction of random gradient descentg：
Figure BDA0001789260810000072
Wherein λ is10.1, which is the regularization weight of the penalty function; lambda [ alpha ]20.9, is the normalized weight of mean square error, λ1+λ2＝1。
And executing the step five: and training the discrimination network and the generation network alternately, and finally realizing effective removal of noise by using a generator forward network. Repeating the third step and the fourth step, repeatedly inputting all batches of data in all training sets, and stopping after the maximum iteration times reaches 500 times through the alternate iteration and optimization of the data and the data; for the trained GAN network, the testing process directly adopts the forward network of the generator to process the test sample and then judges the Peak Signal to noise Ratio (PSNR) after denoising treatment:
Figure BDA0001789260810000073
wherein the content of the first and second substances,
Figure BDA0001789260810000074
for the k-th element, x, in the filtered i-th signal(i)(k) Is the ith kth element of the reference crack signal. The denoising result of the wheel rail noise by using the method is shown in FIG. 5, and the denoising result of the Gaussian white noise is shown in FIG. 6. And finally, randomly selecting 20 samples from the test sample set, and counting the mean value and standard deviation of the peak signal-to-noise ratio of the samples after denoising in different noise environments, wherein the statistical result is shown in a table I.
TABLE I mean and standard deviation of peak SNR in different noise environments
Kind of noise White gaussian noise Mechanical noise of wheel and rail
Peak signal-to-noise ratio mean (dB) 21.458 20.035
Peak signal-to-noise ratio standard deviation (dB) ±2.868 ±3.404
By integrating the analysis of the embodiment, for the denoising and detection of the rail crack acoustic emission signal in the rail flaw detection of the high-speed railway, the invention adopts the rail crack acoustic emission signal detection method based on the improved least square generation countermeasure network, and the method can be used for filtering various types of noise after being trained by different noise sample libraries by using the same filter structure, for example, noise modeling and denoising are performed for Gaussian white noise and wheel rail mechanical noise in the embodiment. The method is suitable for the actual railway field application with a complex noise environment, does not need any priori knowledge, only needs to ensure that a training process gives sufficient reference crack signals and different types of noise signals, and can still achieve the purposes of suppressing noise and detecting the crack signals under the condition that the steel rail crack signals are completely submerged in the noise signals.

Claims (6)

1. A rail crack acoustic emission signal detection method based on an improved least square generation type countermeasure network is characterized by comprising the following steps:
the method comprises the following steps: synthesizing noisy signal sample libraries under different noise backgrounds and normalizing to obtain training sample library { z(i)}；
Step two: initializing a generating countermeasure network, setting network depth and structure parameters of a generator network (G) and a discriminator network (D), including the number of network layers, the number of nodes and activation functions of each node, and initializing the two networks;
step three: inputting the denoised sample obtained by the forward network filtering of the generator (G) and the reference sample into a discriminator network (D) together and updating a discrimination network weight theta according to a discrimination errord；
Step four: updating generator network weight theta according to the loss value of the discriminator and the denoised Mean Square Error (MSE)g；
Step five: and alternately training the discrimination network and the generation network to finally use the generator forward network to realize effective noise removal, calculating the final peak signal-to-noise ratio and judging whether the requirements of the denoising precision are met.
2. The method for detecting the rail crack acoustic emission signal based on the improved least square generation type countermeasure network according to claim 1, wherein the first step is as follows:
collecting specific wheel-rail noise signal x generated during train running on railway site(i) noiseAnd maintaining the initial amplitude of the reference crack signal x for the same length of time(i)Superposing, synthesizing noise-containing crack signals, and preprocessing the noise-containing crack signals to obtain normalized noise-containing signals z(i)，z(i)＝(x(i)+x(i) noise)/max(abs(x(i)+x(i) noise))；
Wherein max (beta) represents the maximum value of the solved vector, abs (beta) represents the absolute value of the solved vector element, i is the serial number of the signal sample in the training set, gives the label corresponding to the crack signal and maps the label into the attribute label x of One-hot coding format(i) cE.g. elastic or plastic phase corresponding to a label of [0,1]]Or [1,0 ]]Will label x(i) cSpliced to a normalized noisy signal z(i)Then n training samples (z, x) are obtainedc)n×lIn the same way, m test samples (z) are obtainedt,xct)m×lAnd l is the number of sampling points contained in the standard noise-containing signal.
3. The steel rail crack acoustic emission signal detection method based on the improved least square generation type countermeasure network according to claim 1, characterized in that the second step is:
an initialization process for a generative countermeasure network, comprising: setting initial values and structural parameters of a generator network (G) and a discriminator network (D): including the number of network layers, the number of nodes and the initial value of weight, training the minimum batch nbatchThe forward networks of the invention G and D all adopt full connection layers, and the interlayer calculation formulas are as follows:
Figure FDA0001789260800000011
wherein the content of the first and second substances,
Figure FDA0001789260800000012
is the output of the l-level j cell,
Figure FDA0001789260800000013
is the output of the l +1 level i cell,
Figure FDA0001789260800000014
the weight value of the connection l layer j unit and l +1 layer i unit in the G or D network,
Figure FDA0001789260800000015
is the bias of the l +1 layer i cells,
Figure FDA0001789260800000016
for the output of the i +1 layer i unit, f (·) represents an activation function, and a sigmoid function or a hyperbolic tangent function and the like can be selected, and the corresponding expression is as follows:
Figure FDA0001789260800000021
specifically, in principle of selecting an activation function, because the generator network is used for filtering, a structure of a noise reduction auto encoder (DAE) is adopted and the activation function of an output layer should select a hyperbolic tangent function to simulate fluctuation of an output signal in a positive and negative range; and the discriminator network is used for judging whether the filtered signal still contains noise components, a sigmoid function is needed to be adopted as an output layer function, so that the output value is easier to be compressed in the range of [0,1], an approximate probability result of two classifications is given, and the input-output relation of the generator network (G) and the discriminator network (D) is as follows:
Figure FDA0001789260800000022
Figure FDA0001789260800000023
wherein D (-) is the mapping operator of the discriminator neural network, thetadThe weight matrix parameter of the discriminator network (D) is used for discriminating the de-noised signal and the reference signal and giving out a discrimination result
Figure FDA0001789260800000024
For back propagation and parameter updating; g (-) is the mapping operator of the generator neural network, θgTo generate weight matrix parameters for the network (G) of generators, which act on noisy inputs z(i)Predicting and filtering to obtain corresponding de-noising signal
Figure FDA0001789260800000025
The objective function in the iterative optimization process of the discriminator network (D) and the generator network (G) is as follows:
Figure FDA0001789260800000026
Figure FDA0001789260800000027
4. the steel rail crack acoustic emission signal detection method based on the improved least square generation type countermeasure network according to claim 1, characterized in that the third step is:
inputting the denoised sample and the reference sample into a discriminator network together and updating the weight of the discriminator network according to the discrimination error; sampling a batch of noisy signals
Figure FDA0001789260800000028
Inputting the filtered sample and a reference sample without noise into a forward network of the generator, substituting the filtered sample and the reference sample without noise into a discriminator network (D), calculating Mean Square Error (MSE) between the filtered sample and the reference sample and the random gradient of a countercheck loss function, and updating a weight matrix parameter theta of the discriminator (D) according to the descending direction of the random gradient of the loss functiond：
Figure FDA0001789260800000029
Wherein, D (-) is the mapping operator of the discriminator neural network, and G (-) is the mapping operator of the generator neural network.
5. The steel rail crack acoustic emission signal detection method based on the improved least square generation type countermeasure network according to claim 1, characterized in that the fourth step is:
after the discriminator optimizes the k steps, updating a generator (G) weight matrix parameter theta in the descending direction of the random gradient according to the loss value obtained by the discriminator identifying the noise signal in the mean square error and the antagonistic loss functiong：
Figure FDA0001789260800000031
Wherein λ is1To combat the regularizing weight of the loss function, λ2Is a regularized weight of mean square error, λ1+λ2The mean square error is calculated from the vector two norm as 1:
Figure FDA0001789260800000032
wherein
Figure FDA0001789260800000033
x(i)(k) Representing the ith sample vector x(i)The kth element of (1).
6. The rail crack acoustic emission signal detection method based on the improved least square generation type countermeasure network according to claim 1, characterized in that the fifth step is:
repeating the third step and the fourth step, inputting all training set data repeatedly, and stopping after the maximum iteration number N is reached through the alternating iteration and optimization of the training set data and the training set data; for the trained GAN structure, the testing process directly adopts a forward network of a generator to process a test sample, then judges the Peak Signal to Noise Ratio (PSNR) after the processing and checks whether the precision requirement is met:
Figure FDA0001789260800000034
wherein the content of the first and second substances,
Figure FDA0001789260800000035
for the k-th element, x, in the filtered i-th signal(i)(k) Is the ith kth element of the reference crack signal.
CN201811029104.XA 2018-09-05 2018-09-05 Steel rail crack acoustic emission signal detection method based on improved least square generation type countermeasure network Active CN110879254B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811029104.XA CN110879254B (en) 2018-09-05 2018-09-05 Steel rail crack acoustic emission signal detection method based on improved least square generation type countermeasure network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811029104.XA CN110879254B (en) 2018-09-05 2018-09-05 Steel rail crack acoustic emission signal detection method based on improved least square generation type countermeasure network

Publications (2)

Publication Number Publication Date
CN110879254A true CN110879254A (en) 2020-03-13
CN110879254B CN110879254B (en) 2021-09-17

Family

ID=69727053

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811029104.XA Active CN110879254B (en) 2018-09-05 2018-09-05 Steel rail crack acoustic emission signal detection method based on improved least square generation type countermeasure network

Country Status (1)

Country Link
CN (1) CN110879254B (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112014478A (en) * 2020-08-28 2020-12-01 武汉大学 Defect echo blind extraction self-adaptive method submerged in ultrasonic grass-shaped signal
CN112836570A (en) * 2020-12-16 2021-05-25 东南大学 Equipment anomaly detection method utilizing Gaussian noise
CN112926504A (en) * 2021-03-23 2021-06-08 重庆商务职业学院 Acoustic emission signal denoising method based on noise reduction self-encoder
CN113052267A (en) * 2021-04-28 2021-06-29 电子科技大学 Unsupervised transmitter phase noise parameter extraction method based on generation countermeasure network
CN113111720A (en) * 2021-03-17 2021-07-13 浙江工业大学 Electromagnetic modulation signal denoising method and system based on deep learning
CN113378644A (en) * 2021-05-14 2021-09-10 浙江工业大学 Signal modulation type recognition attack defense method based on generative countermeasure network
JPWO2022044150A1 (en) * 2020-08-26 2022-03-03
CN116250844A (en) * 2023-03-03 2023-06-13 山东大学 Electrocardiosignal noise reduction optimization method and system based on condition generation countermeasure network
CN117743768A (en) * 2024-02-21 2024-03-22 山东大学 Signal denoising method and system based on denoising generation countermeasure network and diffusion model

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104777230A (en) * 2015-04-16 2015-07-15 哈尔滨工业大学 Train-mounted movable detection system and train-mounted movable detection method for rail defects of high-speed railway on basis of acoustic emission technology
CN104977357A (en) * 2015-08-04 2015-10-14 哈尔滨工业大学 Steel rail crack sound emission signal extracting and denoising method based on kalman filtering
CN105092711A (en) * 2015-08-04 2015-11-25 哈尔滨工业大学 Steel rail crack acoustic emission signal detecting and denoising method based on Kalman filtering
CN106596735A (en) * 2016-12-09 2017-04-26 四川西南交大铁路发展股份有限公司 Denoising and feature extraction method and system for acoustic emission signals of rail cracks
CN106645424A (en) * 2016-12-09 2017-05-10 四川西南交大铁路发展股份有限公司 Method and system for filtering online monitored noise of steel rail cracks and judging cracks
CN108226892A (en) * 2018-03-27 2018-06-29 天津大学 A kind of radar signal restoration methods under complicated noise based on deep learning

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104777230A (en) * 2015-04-16 2015-07-15 哈尔滨工业大学 Train-mounted movable detection system and train-mounted movable detection method for rail defects of high-speed railway on basis of acoustic emission technology
CN104977357A (en) * 2015-08-04 2015-10-14 哈尔滨工业大学 Steel rail crack sound emission signal extracting and denoising method based on kalman filtering
CN105092711A (en) * 2015-08-04 2015-11-25 哈尔滨工业大学 Steel rail crack acoustic emission signal detecting and denoising method based on Kalman filtering
CN106596735A (en) * 2016-12-09 2017-04-26 四川西南交大铁路发展股份有限公司 Denoising and feature extraction method and system for acoustic emission signals of rail cracks
CN106645424A (en) * 2016-12-09 2017-05-10 四川西南交大铁路发展股份有限公司 Method and system for filtering online monitored noise of steel rail cracks and judging cracks
CN108226892A (en) * 2018-03-27 2018-06-29 天津大学 A kind of radar signal restoration methods under complicated noise based on deep learning

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP7392166B2 (en) 2020-08-26 2023-12-05 三菱重工業株式会社 Image generation device, image generation method and program
JPWO2022044150A1 (en) * 2020-08-26 2022-03-03
WO2022044150A1 (en) * 2020-08-26 2022-03-03 三菱重工業株式会社 Image generation device, image generation method, and program
CN112014478A (en) * 2020-08-28 2020-12-01 武汉大学 Defect echo blind extraction self-adaptive method submerged in ultrasonic grass-shaped signal
CN112836570A (en) * 2020-12-16 2021-05-25 东南大学 Equipment anomaly detection method utilizing Gaussian noise
CN112836570B (en) * 2020-12-16 2024-02-06 东南大学 Equipment abnormality detection method utilizing Gaussian noise
CN113111720A (en) * 2021-03-17 2021-07-13 浙江工业大学 Electromagnetic modulation signal denoising method and system based on deep learning
CN112926504A (en) * 2021-03-23 2021-06-08 重庆商务职业学院 Acoustic emission signal denoising method based on noise reduction self-encoder
CN113052267B (en) * 2021-04-28 2022-06-14 电子科技大学 Unsupervised transmitter phase noise parameter extraction method based on generation countermeasure network
CN113052267A (en) * 2021-04-28 2021-06-29 电子科技大学 Unsupervised transmitter phase noise parameter extraction method based on generation countermeasure network
CN113378644A (en) * 2021-05-14 2021-09-10 浙江工业大学 Signal modulation type recognition attack defense method based on generative countermeasure network
CN113378644B (en) * 2021-05-14 2024-03-22 浙江工业大学 Method for defending signal modulation type recognition attack based on generation type countermeasure network
CN116250844A (en) * 2023-03-03 2023-06-13 山东大学 Electrocardiosignal noise reduction optimization method and system based on condition generation countermeasure network
CN116250844B (en) * 2023-03-03 2024-04-26 山东大学 Electrocardiosignal noise reduction optimization method and system based on condition generation countermeasure network
CN117743768A (en) * 2024-02-21 2024-03-22 山东大学 Signal denoising method and system based on denoising generation countermeasure network and diffusion model
CN117743768B (en) * 2024-02-21 2024-05-17 山东大学 Signal denoising method and system based on denoising generation countermeasure network and diffusion model

Also Published As

Publication number Publication date
CN110879254B (en) 2021-09-17

Similar Documents

Publication Publication Date Title
CN110879254B (en) Steel rail crack acoustic emission signal detection method based on improved least square generation type countermeasure network
CN110728654B (en) Automatic pipeline detection and classification method based on deep residual error neural network
CN104792530B (en) Deep-learning rolling bearing fault diagnosis method based on SDA (stacked denoising autoencoder) and Softmax regression
CN103398769B (en) Transformer on-line fault detecting method based on sampling integrated SVM (support vector machine) under wavelet GGD (general Gaussian distribution) feather and unbalanced K-mean value
CN107657250B (en) Bearing fault detection and positioning method and detection and positioning model implementation system and method
CN110879253B (en) Steel rail crack acoustic emission signal detection method based on improved long-time and short-time memory network
CN107702922B (en) Rolling bearing fault diagnosis method based on LCD and stacking automatic encoder
Ai et al. Source localization on large-scale canisters for used nuclear fuel storage using optimal number of acoustic emission sensors
CN109147817B (en) Denoising frequency characteristic extraction method based on variation-limited Boltzmann machine
CN105973595A (en) Diagnosis method of rolling bearing fault
Boussetta et al. Study of the filament wound glass/polyester composite damage behavior by acoustic emission data unsupervised learning
CN112231975A (en) Data modeling method and system based on reliability analysis of railway power supply equipment
CN111353640B (en) Method for constructing wind speed prediction model by combination method
CN115587290A (en) Aero-engine fault diagnosis method based on variational self-coding generation countermeasure network
CN114660180A (en) Sound emission and 1D CNNs-based light-weight health monitoring method and system for medium and small bridges
CN115014789A (en) CNN-GCN-based dual-sensor aeroengine case fault source acoustic emission positioning method
CN114330096A (en) Deep learning-based rail transit motor bearing fault degree classification model construction method
Dai et al. Acceleration-guided acoustic signal denoising framework based on learnable wavelet transform applied to slab track condition monitoring
CN115356599B (en) Multi-mode urban power grid fault diagnosis method and system
CN112183260A (en) One-way valve fault diagnosis method based on total variation noise reduction and RQA
CN115343573A (en) Voiceprint model evolution equipment fault fusion judgment method facing power grid monitoring
CN113409213A (en) Plunger pump fault signal time-frequency graph noise reduction enhancement method and system
CN113592028A (en) Method and system for identifying logging fluid by using multi-expert classification committee machine
CN110555457A (en) Engine lubricating oil wear signal characteristic processing system and method
CN112926504A (en) Acoustic emission signal denoising method based on noise reduction self-encoder

Legal Events

Date Code Title Description
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant