CN110738168A - distributed strain micro crack detection system and method based on stacked convolution self-encoder - Google Patents

distributed strain micro crack detection system and method based on stacked convolution self-encoder Download PDF

Info

Publication number
CN110738168A
CN110738168A CN201910974481.9A CN201910974481A CN110738168A CN 110738168 A CN110738168 A CN 110738168A CN 201910974481 A CN201910974481 A CN 201910974481A CN 110738168 A CN110738168 A CN 110738168A
Authority
CN
China
Prior art keywords
strain
encoder
self
stacked
convolution
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
CN201910974481.9A
Other languages
Chinese (zh)
Other versions
CN110738168B (en
Inventor
宋青松
王浩林
陈禹
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.)
Changan University
Original Assignee
Changan University
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 Changan University filed Critical Changan University
Priority to CN201910974481.9A priority Critical patent/CN110738168B/en
Publication of CN110738168A publication Critical patent/CN110738168A/en
Application granted granted Critical
Publication of CN110738168B publication Critical patent/CN110738168B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/80—Recognising image objects characterised by unique random patterns
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Abstract

The invention discloses distributed strain crack detection systems and methods based on stacked convolutional encoders, which are characterized in that firstly, good characteristic characterization capability of a deep neural network is utilized, crack detection is regarded as binary classification problems, deep neural networks based on stacked convolutional encoders are constructed, and the classification of cracks and non-cracks of a structural body strain subsequence is realized.

Description

distributed strain micro crack detection system and method based on stacked convolution self-encoder
Technical Field
The invention belongs to the field of pattern recognition, and particularly relates to distributed strain micro crack detection systems and methods based on stacked convolution self-encoders.
Background
The crack detection is an important subject in the field of structural health monitoring, and the crack detection method comprises a manual observation method and a nondestructive detection method, wherein the manual observation method needs a maintenance person to regularly check by using a professional tool, the method has low efficiency and strong subjectivity, the nondestructive detection method mainly detects cracks of a structural body by using data obtained by ultrasonic waves, X rays, a ground penetrating radar, a camera and the like, and the sensors are point-to-point sensors, cannot measure the whole data of the structural body and are easy to miss cracks.
Disclosure of Invention
The invention aims to provide distributed strain micro crack detection systems and methods based on stacked convolution self-encoders, so as to overcome the defects in the prior art, the system and method provided by the invention can effectively acquire the overall strain data of the structural body, perform crack detection on the overall strain data of the structural body, obviously improve the accuracy of crack detection, improve the detection effect of crack detection, and provide efficient crack detection schemes for structural health monitoring.
In order to achieve the purpose, the invention adopts the following technical scheme:
A distributed strain crack detection system based on a stacked convolution self-encoder, comprising:
strain sequence acquisition module: the system is used for acquiring distributed strain of the surface of the structure;
a strain sequence preprocessing module: the strain acquisition device is used for performing z-score standardization on the acquired distributed strain and intercepting the distributed strain into a strain subsequence;
the self-learning and characterization module of the characteristics based on the stacked convolution self-encoder comprises: the system consists of 3 convolutional automatic encoder modules and is used for extracting the characteristics of the divided strain subsequences;
and the Softmax classification and identification module is used for carrying out secondary classification on the extracted subsequence characteristics and judging the probability that each subsequences belong to the fracture subsequence and the non-fracture subsequence.
And , a strain sequence acquisition module, which is used for laying the optical fiber sensor on the surface of the structure and acquiring the distributed strain of the surface of the structure by using a distributed optical fiber sensing system based on BOTDA.
, the strain sequence preprocessing module includes a z-score normalization module and a sliding window module;
the z-score normalization module normalizes the strain sequences to 0-mean 1 standard deviation data.
The sliding window module cuts the normalized strain sequence through a sliding window of length 24 and step size 1 into sets of strain subsequences of length 24.
, a feature self-learning and characterization module based on the stacked convolutional self-encoder, which is composed of 3 convolutional self-encoder modules for extracting the features of the divided strain subsequences, wherein the convolutional self-encoder module is used for inputting data x and features h2The relationship between them can be expressed as two formulas, which are shown in detail below:
Figure BDA0002233151600000021
h2＝pool(h1)
wherein h is1Features after convolution;
Figure BDA0002233151600000022
is a convolution; w is a convolution kernel; b is a bias vector; h is2Outputting the features for the encoder; pool indicates pooling operation; sfIs an activation function in the encoder. Characteristic h2And output
Figure BDA0002233151600000023
The relationship between them can also be expressed as three formulas, which are shown in detail below:
Figure BDA0002233151600000024
Figure BDA0002233151600000025
Figure BDA0002233151600000026
wherein the content of the first and second substances,
Figure BDA0002233151600000027
and
Figure BDA0002233151600000028
for the th convolution kernel and offset vector in the decoding process,
Figure BDA0002233151600000029
the characteristic of th convolution in the decoder, and upsamplale is the up-sampling process;
Figure BDA0002233151600000031
features after upsampling;and
Figure BDA0002233151600000033
convolution kernel and offset vector, s, for the second convolution in the decoding processgIs an activation function in the decoder.
The process of the feature self-learning and characterization module based on the stacked self-encoder is specifically as follows:
Figure BDA0002233151600000034
hi,2＝pool(hi,1)
wherein h isi,1For the convolved data in the ith encoder, hi,2For pooled data in the ith encoder, which is also characteristic of the output of the ith encoder, sfIs an activation function. WiAnd biThe convolution kernel and the offset vector in the ith encoder.
A distributed strain crack detection method based on a stacked convolution self-encoder, comprising the following steps:
step 1: strain sequence collection, namely laying an optical fiber sensor on the surface of a structure, and collecting distributed strain on the surface of the structure by using a distributed optical fiber sensing system based on BOTDA;
step 2: standardizing the acquired strain sequence by using z-score, intercepting the strain sequence by using a sliding window with the length of 24 and the step length of 1 to obtain a strain subsequence, and marking the strain subsequence according to the intercepted position;
and step 3: automatically learning features characterizing the strain subsequence using a neural network based on a stacked convolution auto-encoder;
and 4, step 4: performing secondary classification on the extracted characteristics of the strain subsequence by adopting a Softmax classifier to finish crack detection;
, in step 1, the strain sequence acquisition process includes adhering an optical fiber sensor to the surface of the structure through epoxy resin, connecting two ends of the optical fiber to a BOTDA-based distributed optical fiber sensing system, measuring Brillouin frequency shift of the optical fiber through two light sources, namely pump light and detection light, and obtaining distributed strain of the surface of the structure through the linear relation between the Brillouin frequency shift and strain.
, the specific process of the strain sequence processing in step 2 is:
step 2.1: the mean of the collected strain sequences was subtracted by their standard deviations, and the data were obtained as 0 mean 1 standard deviation.
And 2.2, sliding along the acquired strain sequence by using a sliding window with the length of 24 and the step size of 1, and cutting the strain sequence into strain subsequences with the length of 24.
Step 2.3: marking the obtained strain subsequences as cut position mark labels, marking the strain subsequences which are cut by taking the crack as the center as crack subsequences, marking the strain subsequences at the left side and the right side of the strain subsequences as crack subsequences, and marking the rest strain subsequences as non-crack subsequences.
, the specific process of using the neural network based on the stacked convolution self-encoder to automatically learn and characterize the strain subsequence in step 3 is as follows:
step 3.1: and initializing the stacked convolutional self-encoder, and determining the number of layers and the number of neurons of the stacked convolutional self-encoder. Randomly initializing a connection weight matrix and a bias vector in the stacked convolutional auto-encoder. The number of neurons in the input layer is equal to 24, which is the length of the strained subsequence.
Step 3.2: the stacked convolutional autocoder is pre-trained, the stacked convolutional autocoder is composed of 3 convolutional autocoders, and each convolutional autocoder is pre-trained by using the obtained strain subsequence. Loss function of pre-trained convolutional autocoder is the mean square error L between input and output1The method comprises the following steps:
Figure BDA0002233151600000041
wherein x is an input strain subsequence,
Figure BDA0002233151600000042
for convolution of the reconstructed data from the encoder output, M is the number of all incoming strain subsequences, Xm、Respectively the mth strain subsequence of the input model and the corresponding reconstructed output mth subsequence.
, in step 4, a Softmax classifier is used to classify the sub-sequences of the variables, and the specific method is as follows:
step 4.1: constructing a Softmax classifier using a hypothesis function h for a given input zδ(z) for each classes l, a probability value p (y ═ l | z) is estimated, l ∈ {0,1}, assuming a function hδ(z) output t-dimensional vectors representing the probability values of the t estimates, t being 2, assuming the function hδ(z) is as follows:
Figure BDA0002233151600000051
wherein, delta1,δ2Are all parameters of the Softmax classifier,
Figure BDA0002233151600000052
z(i)to input, y(i)For output, the probability of the Softmax classifier classifying z into class l is:
Figure BDA0002233151600000053
wherein z is(i)To input, y(i)Is an output; t denotes the transpose of the matrix.
Step 4.2: pre-training Sofmax, inputting the strain subsequence into the pre-trained stacked convolution self-encoder to obtain the output characteristic z(i)In z is(i)And its label category y(i)Pre-training Softmax, wherein a loss function is a cross entropy function, and the method comprises the following specific steps:
Figure BDA0002233151600000054
wherein the content of the first and second substances,
Figure BDA0002233151600000055
for all the parameters of the Softmax classifier,
Figure BDA0002233151600000056
is the class probability of the output, λ1The weight coefficient is a weight coefficient connecting a weight matrix and a bias vector regular term in Softmax, M is the total number of input strain subsequences, and K is the category number and is 2.
Step 4.3: and fine adjustment, namely stacking the coding part of the convolutional self-encoder and then connecting a Softmax classifier, so that the convolutional self-encoder has a classification function. And utilizing the pre-trained strain subsequence to finely adjust a connection weight matrix and an offset vector of the coding part of the stacked convolutional self-encoder and the overall structure of the Softmax classifier. The loss function during trimming is a cross loss function, and specifically, the loss function is as follows:
where ω is the stack volumeThe connected weight matrix and the offset vector in the product-self encoder are provided, theta is omega and delta, lambda2The method is characterized in that weight coefficients for connecting a weight matrix and a bias vector regular term in a stacked convolution self-encoder are used.
Step 4.4: the Softmax classifier receives the features output by the stacked convolution self-encoder as input, outputs the class 0 or 1 of the strain subsequence, 0 represents non-crack, and 1 represents crack; for feature z of stacked self-encoder output(i)Selecting the probability p (y)(i)＝l|z(i)(ii) a δ) the largest class i as the class to which the feature corresponds.
Compared with the prior art, the invention has the following beneficial technical effects:
the invention realizes data acquisition by the distributed optical fiber sensor and changes the traditional point-to-point sensing mode. The differences between different data are reduced by normalization. Meanwhile, the contradiction between high spatial resolution and low signal-to-noise ratio of the distributed optical fiber sensor is overcome by a method based on a stacked convolution self-encoder. Stacked convolutional autocoder can extract highly robust, distinguishable features for classification in data with low signal-to-noise ratio. The crack detection device has the advantages that the crack detection device is remarkable in crack detection, can detect the micro cracks, and is improved in the detection effect of the micro cracks.
Drawings
FIG. 1 is a schematic flow diagram of the system of the present invention;
FIG. 2 is a schematic diagram of a convolutional auto-encoder in the present invention
FIG. 3 is a schematic diagram of a stacked convolutional auto-encoder of the present invention;
FIG. 4 is a process schematic of the method of the present invention;
FIG. 5 is a diagram illustrating pre-training and fine-tuning in accordance with the present invention.
Detailed Description
The invention is described in further detail with reference to the drawings in which:
referring to fig. 1 to 5, distributed strain crack detection system based on a stacked convolution self-encoder comprises a strain sequence acquisition module, a strain sequence preprocessing module, a characteristic self-learning and characterization module based on the stacked convolution self-encoder, and a Softmax classification identification module (the specific flow is shown in fig. 1).
The strain sequence acquisition module is used for acquiring the distributed strain of the structural body, and the acquired distributed strain of the structural body is -dimensional sequences;
the strain sequence preprocessing module comprises a z-score standardization module and a sliding window module, wherein the z-score standardization module standardizes the strain sequence into data with 0 mean value and 1 standard deviation, the sliding window module cuts groups of strain subsequences with the length of 24 and the step length of 1 through a sliding window with the length of 24, the obtained strain subsequences are marked as cutting positions, the strain subsequences cut by taking the crack as the center are marked as crack subsequences, the left 3 strain subsequences and the right 4 strain subsequences are marked as crack subsequences, and the rest are marked as non-crack subsequences.
The self-learning and characterization module of the characteristics based on the stacked convolution self-encoder comprises: is composed of 3 convolutional autocoder modules for extracting the characteristics of the divided strain subsequence, and features h of input data x2The relationship between them can be expressed as three formulas, which are shown in detail below:
Figure BDA0002233151600000071
h2＝pool(h1)
wherein h is1Features after convolution;
Figure BDA0002233151600000072
is a convolution; w is a convolution kernel; b is a bias vector; h is2Outputting the features for the encoder; pool indicates pooling operation; sfIs an activation function in the encoder. Characteristic h2And outputThe relationship between them can also be expressed as two formulas, as follows:
Figure BDA0002233151600000075
wherein the content of the first and second substances,
Figure BDA0002233151600000077
and
Figure BDA0002233151600000078
for the th convolution kernel and offset vector in the decoding process,
Figure BDA0002233151600000079
the characteristic of th convolution in the decoder, and upsamplale is the up-sampling process;
Figure BDA0002233151600000081
features after upsampling;
Figure BDA0002233151600000082
and
Figure BDA0002233151600000083
convolution kernel and offset vector, s, for the second convolution in the decoding processgIs an activation function in the decoder.
The process of the feature self-learning and characterization module based on the stacked self-encoder is specifically as follows:
Figure BDA0002233151600000084
hi,2＝pool(hi,1)
wherein h isi,1For the convolved data in the ith encoder, hi,2For pooled data in the ith encoder, tooIs a characteristic of the ith encoder output, sfIs an activation function. WiAnd biThe convolution kernel and the offset vector in the ith encoder.
The Softmax classification module adopts a Softmax classifier to realize the classification of the corresponding subsequence, and the specific method is as follows:
constructing a Softmax classifier using a hypothesis function h for a given input zδ(z) for each classes l, a probability value p (y ═ l | z) is estimated, l ∈ {0,1}, assuming a function hδ(z) output t-dimensional vectors representing the probability values of the t estimates, t being 2, assuming the function hδ(z) is as follows:
Figure BDA0002233151600000085
wherein, delta1,δ2Are all parameters of the Softmax classifier,
Figure BDA0002233151600000086
z(i)to input, y(i)For output, the probability of the Softmax classifier classifying z into class l is:
Figure BDA0002233151600000087
wherein z is(i)To input, y(i)Is an output;
the Softmax classifier receives the features output by the stacked convolution self-encoder as input, outputs the class 0 or 1 of the strain subsequence, 0 represents non-crack, and 1 represents crack; for stacked convolution from the feature z of the encoder output(i)Selecting the probability p (y)(i)＝l|z(i)(ii) a δ) the largest class i as the class to which the feature corresponds.
distributed strain crack detection method based on stacked convolution self-encoder, the concrete steps are shown in fig. 5:
1) strain sequence acquisition;
2) standardizing the collected strain sequence by using z-score, intercepting the strain sequence by using a sliding window with the length of 24 and the step length of 1 to obtain a strain subsequence, and marking the strain subsequence according to the intercepted position;
2.1: the mean of the collected strain sequences was subtracted by their standard deviations, and the data were obtained as 0 mean 1 standard deviation.
2.2 slide along the acquired strain sequence using a sliding window of length 24 and step size 1, truncating the strain sequence into sets of strain subsequences of length 24.
2.3: marking the obtained strain subsequences as cut position mark labels, marking the strain subsequences which are cut by taking the crack as the center as crack subsequences, and marking the strain subsequences at the left side and the right side of the strain subsequences as the crack subsequences.
3) Automatically learning features characterizing the strain subsequence with a neural network based on a stacked self-encoder;
3.1: and initializing the stacked convolutional self-encoder, and determining the number of layers and the number of neurons of the stacked convolutional self-encoder. Randomly initializing a connection weight matrix and a bias vector in the stacked convolutional auto-encoder. The number of neurons in the input layer is equal to 24, which is the length of the strained subsequence.
3.2: pre-training a stacked convolutional auto-encoder, the stacked auto-encoder consisting of 3 convolutional auto-encoders, each convolutional auto-encoder being pre-trained with the obtained strain subsequence. The loss function of the pre-training convolutional self-encoder is the mean square error between the input and the output, and is specifically as follows:
wherein x is an input strain subsequence,
Figure BDA0002233151600000102
for convolution of the reconstructed data from the encoder output, M is the number of all incoming strain subsequences, Xm、
Figure BDA0002233151600000103
Are respectively transportedAnd (4) entering the m-th strain subsequence of the model and the corresponding m-th subsequence of the reconstructed output.
4) Secondly, performing secondary classification on the extracted characteristics of the strain subsequence by adopting a Softmax classifier to finish crack detection;
4.1: constructing a Softmax classifier using a hypothesis function h for a given input zδ(z) for each classes l, a probability value p (y ═ l | z) is estimated, l ∈ {0,1}, assuming a function hδ(z) output t-dimensional vectors representing the probability values of the t estimates, t being 2, assuming the function hδ(z) is as follows:
wherein, delta1,δ2Are all parameters of the Softmax classifier,
Figure BDA0002233151600000105
z(i)to input, y(i)For output, the probability of the Softmax classifier classifying z into class l is:
wherein z is(i)To input, y(i)Is an output;
4.2: pre-training Sofmax, inputting the strain subsequence into the pre-trained stacked convolution self-encoder to obtain the output characteristic z(i)In z is(i)And its label category y(i)Pre-training Softmax, wherein a loss function is a cross entropy function, and the method comprises the following specific steps:
Figure BDA0002233151600000107
wherein the content of the first and second substances,
Figure BDA0002233151600000108
for all the parameters of the Softmax classifier,is the class probability of the output, λ1The weight coefficient is a weight coefficient connecting a weight matrix and a bias vector regular term in Softmax, M is the total number of input strain subsequences, and K is the category number and is 2.
4.3: and fine adjustment, namely stacking the coding part of the convolutional self-encoder and then connecting a Softmax classifier, so that the convolutional self-encoder has a classification function. And utilizing the pre-trained strain subsequence to finely adjust a connection weight matrix and an offset vector of the coding part of the stacked convolutional self-encoder and the overall structure of the Softmax classifier. The loss function during trimming is a cross loss function, and specifically, the loss function is as follows:
where ω is the connected weight matrix and offset vector in the stacked convolutional auto-encoder, and Θ is ω and δ, λ2The method is characterized in that weight coefficients for connecting a weight matrix and a bias vector regular term in a stacked convolution self-encoder are used.
4.4: the Softmax classifier receives the features output by the stacked convolution self-encoder as input, outputs the class 0 or 1 of the strain subsequence, 0 represents non-crack, and 1 represents crack; for stacked convolution from the feature z of the encoder output(i)Selecting the probability p (y)(i)＝l|z(i)(ii) a δ) the largest class i as the class to which the feature corresponds.
Effects of the implementation
The method comprises the steps of pre-tensioning an optical fiber sensor, and adhering the optical fiber sensor to the surface of a steel structure in a laboratory through epoxy resin, wherein two ends of the optical fiber sensor are connected with a distributed optical fiber sensing system based on BOTDA, so that distributed strain data distributed on the surface of the steel structure along the radial direction of the optical fiber sensor are obtained.

Claims (9)

1, distributed strain microcrack detection system based on stacked convolution self-encoder, comprising:
strain sequence acquisition module: the system is used for acquiring distributed strain of the surface of the structure;
a strain sequence preprocessing module: the strain acquisition device is used for performing z-score standardization on the acquired distributed strain and intercepting the distributed strain into a strain subsequence;
the self-learning and characterization module of the characteristics based on the stacked convolution self-encoder comprises: the system consists of 3 convolutional automatic encoder modules and is used for extracting the characteristics of the divided strain subsequences;
and the Softmax classification and identification module is used for carrying out secondary classification on the extracted subsequence characteristics and judging the probability that each subsequences belong to the fracture subsequence and the non-fracture subsequence.
2. The distributed strain micro-crack detection system based on stacked convolution self-encoders, as claimed in claim 1, wherein the strain sequence collection module is specifically configured to lay a fiber sensor on the surface of the structure, and collect the distributed strain on the surface of the structure by using a BOTDA-based distributed fiber sensor system.
3. The distributed strain microcrack detection system based on stacked convolutional self-encoder of claim 1, wherein the strain sequence preprocessing module comprises a z-score normalization module and a sliding window module;
the z-score normalization module normalizes the distributed strain to data of 0 means 1 standard deviation;
the sliding window module cuts the normalized strain sequence through a sliding window of length 24 and step size 1 into sets of strain subsequences each of length 24.
4. The distributed strain microcrack detection system based on stacked convolutional self-encoders of claim 1, wherein the module for self-learning and characterization of the features based on stacked convolutional self-encoders is composed of 3A convolution self-encoder module for extracting the characteristics of the divided strain subsequences, wherein the convolution self-encoder module is used for inputting data x and characteristics h2The relationship between them is specifically as follows:
Figure FDA0002233151590000021
h2＝pool(h1)
wherein h is1Features after convolution;
Figure FDA0002233151590000022
is a convolution; w is a convolution kernel; b is a bias vector; h is2Outputting the features for the encoder; pool indicates pooling operation; sfIs an activation function in the encoder, feature h2And output
Figure FDA0002233151590000023
The relationship between them is specifically as follows:
Figure FDA0002233151590000024
Figure FDA0002233151590000025
Figure FDA0002233151590000026
wherein the content of the first and second substances,
Figure FDA0002233151590000027
and
Figure FDA0002233151590000028
for the th convolution kernel and offset vector in the decoding process,
Figure FDA0002233151590000029
the characteristic of th convolution in the decoder, and upsamplale is the up-sampling process;
Figure FDA00022331515900000210
features after upsampling;
Figure FDA00022331515900000211
and
Figure FDA00022331515900000212
convolution kernel and offset vector, s, for the second convolution in the decoding processgIs an activation function in the decoder;
the process of the feature self-learning and characterization module based on the stacked convolution self-encoder is specifically as follows:
Figure FDA00022331515900000213
hi,2＝pool(hi,1)
wherein h isi,1For the convolved data in the ith encoder, hi,2For pooled data in the ith encoder, which is also characteristic of the output of the ith encoder, sfTo activate a function, WiAnd biThe convolution kernel and the offset vector in the ith encoder.
5, A distributed strain crack detection method based on a stacked convolution self-encoder, which is characterized by comprising the following steps:
step 1: collecting distributed strain on the surface of the structure;
step 2: carrying out z-score standardization on the acquired distributed strain and intercepting the distributed strain into a strain subsequence;
and step 3: automatically learning features characterizing the strain subsequence using a neural network based on a stacked convolution auto-encoder;
and 4, step 4: and (4) performing secondary classification on the extracted characteristics of the strain subsequence by adopting a Softmax classifier, and completing crack detection.
6. The distributed strain crack detection method based on stacked convolution self-encoders, according to the claim 5, the distributed strain acquisition in step 1 is that the optical fiber sensor is adhered to the surface of the structure through epoxy resin, two ends of the optical fiber are connected to a BOTDA-based distributed optical fiber sensing system, the BOTDA-based distributed optical fiber sensing system measures the Brillouin frequency shift of the optical fiber through two light source pumping light and detection light, and the distributed strain of the surface of the structure is obtained through the linear relation between the Brillouin frequency shift and the strain.
7. The distributed strain crack detection method based on stacked convolution self-encoders as claimed in claim 5, wherein step 2 specifically includes:
step 2.1: subtracting the mean value of the acquired distributed strain, and dividing the mean value by the standard deviation to obtain data of 0 mean value 1 standard deviation;
step 2.2, intercepting the normalized strain sequence into strain subsequences with the length of 24 by using a sliding window with the length of 24 and the step length of 1;
step 2.3: marking the obtained strain subsequences as cut position mark labels, marking the strain subsequences which are cut by taking the crack as the center as crack subsequences, marking the strain subsequences at the left side and the right side of the strain subsequences as crack subsequences, and marking the rest strain subsequences as non-crack subsequences.
8. The distributed strain crack detection method based on stacked convolution self-encoder, according to claim 5, wherein the specific process of using the neural network based on stacked convolution self-encoder to automatically learn and characterize the strain subsequence in step 3 is as follows:
step 3.1: initializing a convolution stacking self-encoder, determining the number of layers and the number of neurons of the convolution stacking self-encoder, randomly initializing a connection weight matrix and a bias vector in the convolution stacking self-encoder, and enabling the number of the neurons of an input layer to be equal to the length of a strain subsequence;
step 3.2: pre-training a stacked convolutional auto-encoder, the stacked convolutional auto-encoder is composed of 3 convolutional auto-encoders, each convolutional auto-encoder is pre-trained by the obtained strain subsequence, and the loss function of the pre-trained convolutional auto-encoder is the mean square error L between the input and the output1The method comprises the following steps:
Figure FDA0002233151590000041
wherein x is an input strain subsequence,
Figure FDA0002233151590000042
for convolution of the reconstructed data from the encoder output, M is the number of all incoming strain subsequences, Xm、
Figure FDA0002233151590000043
Respectively the mth strain subsequence of the input model and the corresponding reconstructed output mth subsequence.
9. The distributed strain crack detection method based on stacked convolutional self-encoder according to claim 5, wherein a Softmax classifier is used in step 4 to classify the strain subsequences, and the specific method is as follows:
step 4.1: constructing a Softmax classifier using a hypothesis function h for a given input zδ(z) for each classes l, a probability value p (y ═ l | z) is estimated, l ∈ {0,1}, assuming a function hδ(z) output t-dimensional vectors representing the probability values of the t estimates, t being 2, assuming the function hδ(z) is as follows:
wherein the content of the first and second substances,
Figure FDA0002233151590000045
δ1,δ2is all the parameters of the Softmax classifier, z(i)To input, y(i)For output, the probability of the Softmax classifier classifying z into class l is:
Figure FDA0002233151590000046
wherein z is(i)To input, y(i)Is an output; t represents the transpose of the matrix;
step 4.2: pre-training Sofmax, inputting the strain subsequence into the pre-trained stacked convolution self-encoder to obtain the output characteristic z(i)In z is(i)And its label category y(i)Pre-training Softmax, wherein a loss function is a cross entropy function, and the method comprises the following specific steps:
Figure FDA0002233151590000051
wherein the content of the first and second substances,is the class probability of the output, λ1The weight coefficient is a weight coefficient connecting a weight matrix and a bias vector regular term in Softmax, M is the total number of input strain subsequences, K is a category number, and K is 2;
step 4.3: utilizing a pre-trained strain subsequence to finely adjust a connection weight matrix and a bias vector of an encoding part of the stacked convolution self-encoder and an overall structure of the Softmax classifier, wherein a loss function during fine adjustment is a cross loss function, and the method specifically comprises the following steps:
Figure FDA0002233151590000053
where ω is the connected weight matrix and offset vector in the stacked convolutional auto-encoder, and Θ is ω and δ, λ2A weight coefficient connecting a weight matrix and a bias vector regular term in the stacked convolution self-encoder;
step 4.4: the Softmax classifier receives the features output by the stacked convolution self-encoder as input, outputs the class 0 or 1 of the strain subsequence, 0 represents non-crack, and 1 represents crack; for feature z of stacked self-encoder output(i)Selecting the probability p (y)(i)＝l|z(i)(ii) a δ) the largest class i as the class to which the feature corresponds.
CN201910974481.9A 2019-10-14 2019-10-14 Distributed strain micro crack detection system and method based on stacked convolution self-encoder Active CN110738168B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910974481.9A CN110738168B (en) 2019-10-14 2019-10-14 Distributed strain micro crack detection system and method based on stacked convolution self-encoder

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910974481.9A CN110738168B (en) 2019-10-14 2019-10-14 Distributed strain micro crack detection system and method based on stacked convolution self-encoder

Publications (2)

Publication Number Publication Date
CN110738168A true CN110738168A (en) 2020-01-31
CN110738168B CN110738168B (en) 2023-02-14

Family

ID=69268869

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910974481.9A Active CN110738168B (en) 2019-10-14 2019-10-14 Distributed strain micro crack detection system and method based on stacked convolution self-encoder

Country Status (1)

Country Link
CN (1) CN110738168B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111510205A (en) * 2020-04-21 2020-08-07 北京邮电大学 Optical cable fault positioning method, device and equipment based on deep learning
CN111754445A (en) * 2020-06-02 2020-10-09 国网湖北省电力有限公司宜昌供电公司 Coding and decoding method and system for optical fiber label with hidden information

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2015180368A1 (en) * 2014-05-27 2015-12-03 江苏大学 Variable factor decomposition method for semi-supervised speech features
CN108876796A (en) * 2018-06-08 2018-11-23 长安大学 A kind of lane segmentation system and method based on full convolutional neural networks and condition random field
US20180357542A1 (en) * 2018-06-08 2018-12-13 University Of Electronic Science And Technology Of China 1D-CNN-Based Distributed Optical Fiber Sensing Signal Feature Learning and Classification Method
CN109389171A (en) * 2018-10-11 2019-02-26 云南大学 Medical image classification method based on more granularity convolution noise reduction autocoder technologies

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2015180368A1 (en) * 2014-05-27 2015-12-03 江苏大学 Variable factor decomposition method for semi-supervised speech features
CN108876796A (en) * 2018-06-08 2018-11-23 长安大学 A kind of lane segmentation system and method based on full convolutional neural networks and condition random field
US20180357542A1 (en) * 2018-06-08 2018-12-13 University Of Electronic Science And Technology Of China 1D-CNN-Based Distributed Optical Fiber Sensing Signal Feature Learning and Classification Method
CN109389171A (en) * 2018-10-11 2019-02-26 云南大学 Medical image classification method based on more granularity convolution noise reduction autocoder technologies

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
宋俊芳: "基于CNN的路面裂缝图像分类", 《信息与电脑(理论版)》 *
邓俊锋等: "基于自动编码器组合的深度学习优化方法", 《计算机应用》 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111510205A (en) * 2020-04-21 2020-08-07 北京邮电大学 Optical cable fault positioning method, device and equipment based on deep learning
CN111510205B (en) * 2020-04-21 2022-07-12 北京邮电大学 Optical cable fault positioning method, device and equipment based on deep learning
CN111754445A (en) * 2020-06-02 2020-10-09 国网湖北省电力有限公司宜昌供电公司 Coding and decoding method and system for optical fiber label with hidden information
CN111754445B (en) * 2020-06-02 2022-03-18 国网湖北省电力有限公司宜昌供电公司 Coding and decoding method and system for optical fiber label with hidden information

Also Published As

Publication number Publication date
CN110738168B (en) 2023-02-14

Similar Documents

Publication Publication Date Title
CN108074231B (en) Magnetic sheet surface defect detection method based on convolutional neural network
CN112884747B (en) Automatic bridge crack detection system integrating cyclic residual convolution and context extractor network
CN106548182A (en) Based on deep learning and the causal analytic pavement distress survey method and device of master
CN110738168B (en) Distributed strain micro crack detection system and method based on stacked convolution self-encoder
CN113436169A (en) Industrial equipment surface crack detection method and system based on semi-supervised semantic segmentation
CN108960342B (en) Image similarity calculation method based on improved Soft-Max loss function
CN110060251A (en) A kind of building surface crack detecting method based on U-Net
CN114581782B (en) Fine defect detection method based on coarse-to-fine detection strategy
CN102521561A (en) Face identification method on basis of multi-scale weber local features and hierarchical decision fusion
CN103745238A (en) Pantograph identification method based on AdaBoost and active shape model
CN111627007B (en) Spacecraft defect detection method based on self-optimization matching network image stitching
CN114494222A (en) Vision transducer-based rolling bearing fault intelligent identification method
CN110715929B (en) Distributed strain micro crack detection system and method based on stacking self-encoder
CN114714145A (en) Method for enhancing, comparing, learning and monitoring tool wear state by using Gelam angular field
CN117252459A (en) Fruit quality evaluation system based on deep learning
CN115130599B (en) Semi-supervision method for identifying open pit mine card state under time sequence GAN data enhancement
CN115937651A (en) Cylindrical roller surface detection method and system based on improved yolov5s network model
CN115082797A (en) Early wheat scab monitoring and early warning method and system in field environment
CN113222920A (en) Suction pipe defect detection method based on pruning Yolov3
Liu et al. Automatic terahertz recognition of hidden defects in layered polymer composites based on a deep residual network with transfer learning
CN113311081A (en) Pollution source identification method and device based on three-dimensional liquid chromatography fingerprint
CN115452957B (en) Small sample metal damage identification method based on attention prototype network
KR102499986B1 (en) Method for detecting micro plastic
CN115326722B (en) Ocean red tide early warning method based on hyperspectral remote sensing data
CN112434590B (en) SAR image wind stripe recognition method based on wavelet transformation

Legal Events

Date Code Title Description
PB01 Publication
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