CN110715929B - Distributed strain micro crack detection system and method based on stacking self-encoder - Google Patents

Abstract

The invention discloses a distributed strain crack detection system and method based on a stacked self-encoder, which utilizes the good feature representation ability of a deep neural network, regards crack detection as a two-class problem, and constructs a depth based on the stacked self-encoder. A neural network for crack and non-crack classification of structural strain subsequences. The method of the invention can accurately and without omission detect tiny cracks with an opening width of 32 μm on a laboratory steel beam, and provides a solution with good noise robustness for the detection of distributed strain cracks on the surface of the structure.

Description

A distributed strain micro-crack detection system and method based on stacked autoencoders

technical field

The invention belongs to the field of pattern recognition, and in particular relates to a distributed strain micro-crack detection system and method based on a stacked self-encoder.

Background technique

Crack detection has always been an important topic in the field of structural health monitoring. Crack detection methods include manual observation methods and non-destructive testing methods. The manual observation method requires specialized maintenance personnel to use professional tools to conduct regular inspections, which is inefficient and highly subjective. Non-destructive testing methods mainly detect structural cracks through the data obtained by ultrasonic, X-ray, ground penetrating radar and cameras. These sensors are all point-to-point sensors, which cannot measure the overall data of the structure and are prone to missing cracks.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a distributed strain micro-crack detection system and method based on stacked self-encoders, so as to overcome the above-mentioned defects in the prior art, and the present invention can effectively collect the overall strain data of the structure. The overall strain data is used for crack detection, which significantly improves the accuracy of crack detection, improves the detection effect of crack detection, and provides an efficient crack detection solution for structural health monitoring.

To achieve the above object, the present invention adopts the following technical solutions:

A distributed strain micro-crack detection system based on stacked self-encoders, comprising: a strain sequence acquisition module for acquiring distributed strain on the surface of a structure; a strain sequence preprocessing module: for collecting the acquired strain sequence Standardize z-score and truncate into strain subsequences; feature self-learning and characterization module based on stacked autoencoder: used to extract the features of the divided strain subsequences; Softmax classification and recognition module, used to extract the subsequences The features are classified into two categories to determine the probability that each subsequence belongs to a cracked subsequence and a non-cracked subsequence;

Further, the strain sequence acquisition module: lays the optical fiber sensor on the surface of the structure, and uses the distributed optical fiber sensing system based on BOTDA to collect the distributed strain on the surface of the structure; the strain sequence preprocessing module includes:

Further, the z-score normalization module and the sliding window interception module, the z-score normalization module normalizes the strain series to data with 0 mean and 1 standard deviation. The sliding window module intercepts a set of strain subsequences with a length of 21 from the normalized strain sequence through a sliding window with a length of 21 and a step size of 1. Feature Self-Learning and Representation Module Based on Stacked Auto-Encoder: It consists of 3 auto-encoder modules, and the encoded part of the 3 auto-encoder modules is used as feature representation.

之间的关系可以表示为f_θ(·)和g_θ'(·)两个函数，具体如下所示：Further, the feature self-learning and characterization module based on stacked autoencoders: composed of 3 autoencoder modules, used to extract the features of the divided strain subsequences, the autoencoder module for input data x, feature h and output

The relationship between can be expressed as two functions f _θ ( ) and g _θ' ( ), as follows:

h=f _θ (x)=s _f (Wx+b _h )

之间的连接矩阵和偏置向量，W^T为W的转置矩阵。s_f和s_g是激活函数。Among them, W and b _h are the connection matrix and bias vector between the input data x and feature h, respectively, W ^T and b _v are the feature h and output, respectively

Between the connection matrix and the bias vector, W ^T is the transpose matrix of W. s _f and s _g are activation functions.

The process of feature self-learning and representation module based on stacked autoencoder is as follows:

h _i =s _f (W _i x _i +b _h,i )

分别为第i(i＝1,2,…,n)个自动编码器模块的输入数据x，特征h与输出

W_i和W_i ^T为其连接权重矩阵，b_h,i和b_v,i为其偏置向量，x_i＝h_i-1，s_f和s_g为激活函数。Among them, x _i , h _i ,

are the input data x, feature h and output of the i-th (i=1,2,...,n) autoencoder module, respectively

Wi and Wi ^T are their connection weight matrices, b _{h,i and} b _{v,i are} their bias vectors, _xi =h _i-1 , s _f and s _g are activation functions.

A distributed strain crack detection method based on stacked autoencoders, comprising the following steps:

Step 1: Strain sequence acquisition;

Step 2: Use z-score normalization to standardize the collected strain sequence, use a sliding window with a length of 21 and a step size of 1 to intercept the strain sequence to obtain a strain subsequence, and mark the strain subsequence according to the interception position;

Step 3: Automatically learn features representing strain subsequences using a stacked autoencoder-based neural network;

Step 4: Use the Softmax classifier to implement binary classification of the features of the extracted strain subsequences to complete crack detection;

Further, the specific process of strain sequence acquisition in step 1 is as follows: the optical fiber sensor is adhered to the surface of the structure through epoxy resin, and the two ends of the optical fiber are connected to the BOTDA-based distributed optical fiber sensing system, and the BOTDA-based distributed optical fiber transmission system. The sensing system measures the Brillouin frequency shift of the fiber through two light sources—pump light and probe light, and obtains the distributed strain on the surface of the structure through the linear relationship between the Brillouin frequency shift and strain.

Further, the specific process of the strain sequence processing in step 2 is:

Step 2.1: Subtract the mean value of the acquired strain series and divide it by its standard deviation to obtain data with 0 mean and 1 standard deviation.

Step 2.2: Use a sliding window with a length of 21 and a step of 1 to slide along the acquired strain sequence, and cut the strain sequence into a set of strain subsequences with a length of 21.

Step 2.3: Label the obtained strain subsequence according to the interception position, mark the strain subsequence intercepted at the crack as the crack subsequence, and mark the left 3 strain subsequences and the right 4 strain subsequences as crack subsequences sequences, and the rest are marked as non-cracked subsequences.

Further, the specific process of using the stacked autoencoder-based neural network to automatically learn the characteristics of the strain subsequence in step 3 is as follows:

Step 3.1: Model initialization, determine the number of layers and neurons of the model. Randomly initialize the connection weight matrix and bias vector in the model. The number of neurons in the input layer is equal to 21, which is the length of the strain subsequence.

Step 3.2: Pretrain the stacked autoencoder, which consists of 3 autoencoders, and pretrain each autoencoder with the resulting strain subsequences. The loss function of the pretrained autoencoder is the mean squared error between the input and the output, as follows:

为自动编码器输出的重构数据，M为所有输入的应变子序列的数量，X^m、

分别是输入模型的第m条应变子序列和对应的输出重构的第m条子序列。where x is the input strain subsequence,

is the reconstructed data output by the autoencoder, M is the number of all input strain subsequences, X ^m ,

are the mth strain subsequence of the input model and the mth subsequence of the corresponding output reconstruction, respectively.

Further, in step 4, the Softmax classifier is used to realize the classification of the strain subsequences, and the specific method is:

Step 4.1: Build a Softmax classifier. For a given input z, use the hypothesis function h _δ (z) to estimate the probability value p(y=l|z) for each category l, l∈{0,1}, suppose The function h _δ (z) outputs a t-dimensional vector representing the t estimated probability values, t=2, assuming that the function h _δ (z) is as follows:

z⁽ⁱ⁾为输入，y⁽ⁱ⁾为输出，Softmax分类器将z分为类别l的概率为：Among them, δ ₁ , δ ₂ are all parameters of the Softmax classifier,

With z ⁽ⁱ⁾ as input and y ⁽ⁱ⁾ as output, the probability that the Softmax classifier classifies z into class l is:

where z ⁽ⁱ⁾ is the input and y ⁽ⁱ⁾ is the output;

Step 4.2: Pre-train Sofmax, input the strain subsequence into the pre-trained stacked autoencoder, get the output feature z ⁽ⁱ⁾ , pre-train Softmax with z ⁽ⁱ⁾ and its label category y ⁽ⁱ⁾ , the loss function is Cross entropy function, as follows:

为Softmax分类器的全部参数，

为输出的类别概率，λ₁为Softmax中连接权重矩阵和偏置向量正则项的权重系数，M为输入应变子序列的总个数，K为类别数，为2。in,

are all parameters of the Softmax classifier,

is the output category probability, λ ₁ is the weight coefficient of the regular term connecting the weight matrix and the bias vector in Softmax, M is the total number of input strain subsequences, and K is the number of categories, which is 2.

Step 4.3: Fine-tuning, the encoding part of the stacked autoencoder is followed by a Softmax classifier, which can make it have a classification function. Using the pretrained strain subsequences to fine-tune the connection weight matrix and bias vector between the encoding part of the stacked autoencoder and the overall structure of the Softmax classifier. The loss function during fine-tuning is a cross loss function, as follows:

where ω is the connection weight matrix and bias vector in the stacked _autoencoder , Θ is ω and δ, and λ2 is the weight coefficient of the regular term of the connection weight matrix and bias vector in the stacked autoencoder.

Step 4.4: The Softmax classifier receives the features output by the stacked autoencoder as its input, and outputs the category 0 or 1 of the strain subsequence, where 0 means non-slit and 1 means crack; for the feature z ⁽ⁱ⁾ output by the stacked autoencoder, Select the category l with the largest probability p(y ⁽ⁱ⁾ =l|z ⁽ⁱ⁾ ; δ) as the category corresponding to the feature. The distributed strain subsequence is restored to the truncated position and the adjacent fracture subsequences are merged into one fracture.

Compared with the prior art, the present invention has the following beneficial technical effects:

The invention realizes data collection by distributed optical fiber sensors, and changes the previous point-to-point sensing method. Standardization narrows the differences between different data. At the same time, the method based on stacked autoencoder overcomes the contradiction between high spatial resolution and low signal-to-noise ratio of distributed optical fiber sensor. Stacked autoencoders can extract highly robust and identifiable features for classification in low signal-to-noise ratio data. It has a significant effect in crack detection, can detect tiny cracks, and has improved the detection effect of tiny cracks.

Description of drawings

Fig. 1 is the schematic flow chart of the system of the present invention;

Fig. 2 is the schematic diagram of the autoencoder in the present invention;

Fig. 3 is the schematic diagram of the stacked autoencoder in the present invention;

Fig. 4 is the process schematic diagram of the inventive method;

FIG. 5 is a schematic diagram of specific pre-training and fine-tuning in the present invention.

Detailed ways

Below in conjunction with accompanying drawing, the present invention is described in further detail:

Referring to Figures 1 to 5, a distributed strain crack detection system based on stacked autoencoders includes a strain sequence acquisition module; a strain sequence preprocessing module; a feature self-learning and characterization module based on stacked autoencoders; Softmax classification and recognition module (the specific process is shown in Figure 1).

The strain sequence acquisition module is used to collect the distributed strain of the structure, and the collected distributed strain of the structure is a one-dimensional sequence;

The strain sequence preprocessing module includes: z-score normalization module and sliding window module. The z-score normalization module normalizes the strain sequence to data with 0 mean and 1 standard deviation. The sliding window module intercepts a set of strain subsequences with a length of 21 from the normalized strain sequence through a sliding window with a length of 21 and a step size of 1. Label the obtained strain subsequence according to the interception position, mark the strain subsequence intercepted at the crack as the crack subsequence, and mark the left 3 strain subsequences and the right four strain subsequences as the crack subsequence, and the rest marked as non-cracked subsequences.

之间的关系可以表示为f_θ(·)和g_θ'(·)两个函数，具体如下所示：The feature self-learning and characterization module based on stacked autoencoder includes: 3 autoencoder modules as shown in Figure 2, the autoencoder module for input data x, feature h and output

The relationship between can be expressed as two functions f _θ ( ) and g _θ' ( ), as follows:

h=f _θ (x)=s _f (Wx+b _h )

之间的连接矩阵和偏置向量，W^T为W的转置矩阵。s_f和s_g是激活函数。Among them, W and b _h are the connection matrix and bias vector between the input data x and feature h, respectively, W ^T and b _v are the feature h and output, respectively

Between the connection matrix and the bias vector, W ^T is the transpose matrix of W. s _f and s _g are activation functions.

The process of feature self-learning and representation module based on stacked autoencoder is as follows:

h _i =s _f (W _i x _i +b _h,i )

分别为第i(i＝1,2,…,n)个自动编码器模块的输入数据x，特征h与输出

W_i和W_i ^T为其连接权重矩阵，b_h,i和b_v,i为其偏置向量，x_i＝h_i-1，s_f和s_g为激活函数。Among them, x _i , h _i ,

are the input data x, feature h and output of the i-th (i=1,2,...,n) autoencoder module, respectively

Wi and Wi ^T are their connection weight matrices, b _{h,i and} b _{v,i are} their bias vectors, _xi =h _i-1 , s _f and s _g are activation functions.

The classification of strain subsequences is realized by Softmax classifier, and the specific method is as follows:

Construct a Softmax classifier. For a given input z, use the hypothesis function h _δ (z) to estimate the probability value p(y=l|z) for each category l, l∈{0,1}, assuming the function h _δ (z) Output a t-dimensional vector representing the t estimated probability values, t=2, assuming that the function h _δ (z) is as follows:

z⁽ⁱ⁾为输入，y⁽ⁱ⁾为输出，Softmax分类器将z分为类别l的概率为：Among them, δ ₁ , δ ₂ are all parameters of the Softmax classifier,

With z ⁽ⁱ⁾ as input and y ⁽ⁱ⁾ as output, the probability that the Softmax classifier classifies z into class l is:

where z ⁽ⁱ⁾ is the input and y ⁽ⁱ⁾ is the output;

The Softmax classifier receives the features output by the stacked self-encoder as its input, and outputs the category 0 or 1 of the strain subsequence, where 0 means no crack and 1 means crack; for the feature z ⁽ⁱ⁾ output by the stacked self-encoder, the probability p is chosen (y ⁽ⁱ⁾ =l|z ⁽ⁱ⁾ ; δ) The largest category l is taken as the category corresponding to this feature. The distributed strain subsequence is restored to the truncated position and the adjacent fracture subsequences are merged into one fracture.

A distributed strain crack detection method based on stacked autoencoders, the specific steps are shown in Figure 5:

1), strain sequence collection;

2), use z-score standardization to standardize the acquired strain sequence, use a sliding window with a length of 21 and a step size of 1 to intercept the strain sequence, obtain a strain subsequence, and mark the strain subsequence according to the interception position;

2.1: Subtract the mean value of the acquired strain series and divide it by its variance to obtain data with 0 mean and 1 standard deviation.

2.2: Use a sliding window with a length of 21 and a step size of 1 to slide along the acquired strain sequence, and cut the strain sequence into a set of strain subsequences with a length of 21.

2.3: Label the obtained strain subsequence according to the interception position, mark the strain subsequence intercepted at the crack as the crack subsequence, and mark the 3 left and 4 right strain subsequences as the crack subsequence .

3), use the neural network based on the stacked autoencoder to automatically learn the characteristics of the strain subsequence;

3.1: Model initialization, determine the number of layers and neurons of the model. Randomly initialize the connection weight matrix and bias vector in the model. The number of neurons in the input layer is equal to 21, which is the length of the strain subsequence.

3.2: Pre-training stacked autoencoders. The stacked autoencoders consist of 3 autoencoders, and each autoencoder is pretrained with the resulting strain subsequences. The loss function of the pretrained autoencoder is the mean squared error between the input and the output, as follows:

为自动编码器输出的重构数据，M为所有输入的应变子序列的数量，X^m、

分别是输入模型的第m条应变子序列和对应的输出重构的第m条子序列。where x is the input strain subsequence,

is the reconstructed data output by the autoencoder, M is the number of all input strain subsequences, X ^m ,

are the mth strain subsequence of the input model and the mth subsequence of the corresponding output reconstruction, respectively.

4), adopt the Softmax classifier to realize the two-classification of the features of the extracted strain subsequence, and complete the crack detection;

4.1: Build a Softmax classifier, for a given input z, use the hypothesis function h _δ (z) to estimate the probability value p(y=l|z) for each category l, l∈{0,1}, the hypothesis function h _δ (z) outputs a t-dimensional vector representing the t estimated probability values, t=2, assuming that the function h _δ (z) is as follows:

z⁽ⁱ⁾为输入，y⁽ⁱ⁾为输出，Softmax分类器将z分为类别l的概率为：Among them, δ ₁ , δ ₂ are all parameters of the Softmax classifier,

With z ⁽ⁱ⁾ as input and y ⁽ⁱ⁾ as output, the probability that the Softmax classifier classifies z into class l is:

where z ⁽ⁱ⁾ is the input and y ⁽ⁱ⁾ is the output;

4.2: Pre-training Sofmax, input the strain subsequence into the pre-trained stacked autoencoder, get the output feature z ⁽ⁱ⁾ , pre-train Softmax with z ⁽ⁱ⁾ and its label category y ⁽ⁱ⁾ , and the loss function is crossover The entropy function, as follows:

为Softmax分类器的全部参数，

为输出的类别概率，λ₁为Softmax中连接权重矩阵和偏置向量正则项的权重系数，M为输入应变子序列的总个数，K为类别数，为2。in,

are all parameters of the Softmax classifier,

is the output category probability, λ ₁ is the weight coefficient of the regular term connecting the weight matrix and the bias vector in Softmax, M is the total number of input strain subsequences, and K is the number of categories, which is 2.

4.3: Fine-tuning, the encoding part of the stacked autoencoder is followed by a Softmax classifier, which can make it have a classification function. Using the pretrained strain subsequences to fine-tune the connection weight matrix and bias vector between the encoding part of the stacked autoencoder and the overall structure of the Softmax classifier. The loss function during fine-tuning is a cross loss function, as follows:

where ω is the connection weight matrix and bias vector in the stacked _autoencoder , Θ is ω and δ, and λ2 is the weight coefficient of the regular term of the connection weight matrix and bias vector in the stacked autoencoder.

4.4: The Softmax classifier receives the features output by the stacked self-encoder as its input, and outputs the category 0 or 1 of the strain subsequence, 0 means non-crack and 1 means crack; for the feature z ⁽ⁱ⁾ output by the stacked self-encoder, choose The category l with the largest probability p(y ⁽ⁱ⁾ =l|z ⁽ⁱ⁾ ; δ) is used as the category corresponding to the feature. The distributed strain subsequence is restored to the truncated position and the adjacent fracture subsequences are merged into one fracture.

Implementation Effect

The fiber optic sensor is first pre-tensioned and then adhered to the surface of the steel structure by epoxy. The two ends of the optical fiber sensor are connected to the distributed optical fiber sensing system based on BOTDA, and the distributed strain data distributed along the radial direction of the optical fiber sensor on the surface of the structure body are obtained. By adopting the micro-crack detection method based on the stacked self-encoder of the present invention, the micro-cracks with an opening width of 32 μm can be accurately detected based on the collected distributed strain data, which is a distributed detection method for micro-cracks on the surface of steel structures. effective method.

Claims (5)

1. A distributed strain microcrack detection system based on stacked self-encoders, comprising:

strain sequence acquisition module: the system is used for acquiring distributed strain of the surface of the structure; the strain sequence acquisition module specifically comprises: 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;

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 strain sequence pre-processing module comprises 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 intercepts a group of strain subsequences with the lengths of 21 from the normalized strain sequence through a sliding window with the length of 21 and the step length of 1;

the self-learning and characterization module of the characteristics based on the stacking self-encoder comprises the following modules: the system comprises 3 automatic encoder modules, a data processing module and a data processing module, wherein the automatic encoder modules are used for extracting the characteristics of the divided strain subsequences, inputting the characteristics into the strain subsequences and outputting the characteristics into the strain subsequences;

and the Softmax classification and identification module is used for judging the probability that each strain subsequence belongs to the crack subsequence and the non-crack subsequence so as to carry out secondary classification on the extracted characteristics of the strain subsequences.

2. The distributed strain microcrack detection system according to claim 1, wherein the self-learning and characterization module based on the self-encoder comprises 3 autoencoder modules for extracting the characteristics of the divided strain subsequences, and the autoencoder modules input data x, characteristics h and output data x, h and output data h

The relationship between is represented as f_θ(. and g)_θ'Two functions, specifically as follows:

h＝f_θ(x)＝s_f(Wx+b_h)

wherein W and b_hRespectively, a connection matrix and an offset vector, W, between the input data x and the feature h^TAnd b_vRespectively characteristic h and output

A connection matrix and an offset vector between, W^TIs a transposed matrix of W, s_fAnd s_gIs an activation function;

the process of the feature self-learning and characterization module based on the stacked self-encoder is specifically as follows:

h_i＝s_f(W_ix_i+b_h,i)

wherein x is_i，h_i，

Input data x, characteristic h and output of the ith automatic encoder module respectively

And i is 1,2, L, n, W_iAnd W_i ^TTo which a weight matrix is connected, b_h,iAnd b_v,iIs its offset vector, x_i＝h_i-1，s_fAnd s_gIs an activation function.

3. A distributed strain micro crack detection method based on a stacked self-encoder is characterized by comprising the following steps:

step 1: collecting distributed strain on the surface of the structure; the strain sequence acquisition specific process comprises the following steps: the method comprises the following steps that an optical fiber sensor is adhered to the surface of a structure body through epoxy resin, two ends of an optical fiber are connected to a BOTDA-based distributed optical fiber sensing system, the BOTDA-based distributed optical fiber sensing system measures Brillouin frequency shift of the optical fiber through two light source pumping light and detection light, and distributed strain of the surface of the structure body is obtained through the linear relation between the Brillouin frequency shift and strain;

step 2: carrying out z-score standardization on the acquired distributed strain and intercepting the distributed strain into a strain subsequence; the method specifically comprises the following steps:

step 2.1: subtracting the mean value of the acquired strain sequence, 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 a group of strain subsequences with the length of 21 by using a sliding window with the length of 21 and the step size 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 3 strain subsequences on the left side and 4 strain subsequences on the right side as crack subsequences, and marking the rest as non-crack subsequences;

and step 3: automatically learning features characterizing the strain subsequence using a stacked self-encoder based neural network;

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.

4. The distributed strain microcrack detection method based on the stacked self-encoder according to claim 3, wherein the specific process of using the neural network based on the stacked self-encoder to automatically learn and characterize the strain subsequence features in step 3 is as follows:

step 3.1: initializing a model, determining the number of layers and the number of neurons of the model, randomly initializing a connection weight matrix and a bias vector in the model, and inputting the number of the neurons of the layer to be equal to the length of a strain subsequence;

step 3.2: pre-training a stacked self-encoder, the stacked self-encoder consisting of 3 auto-encoders, pre-training each auto-encoder with the obtained strain sub-sequence, the loss function of the pre-trained auto-encoder being the mean square error L between input and output₁The method comprises the following steps:

wherein x is an input strain subsequence,

for the reconstructed data output from the autoencoder, M is the number of all the input strain subsequences, X^m、

Respectively the mth strain subsequence of the input model and the corresponding mth subsequence of the output reconstruction.

5. The distributed strain micro crack detection method based on the stacked self-encoder as claimed in claim 3, wherein a Softmax classifier is adopted in the step 4 to realize classification of 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 z_δ(z) for each class l, a probability value p (y ═ l | z) is estimated, l ∈ {0,1}, assuming a function h_δ(z) outputting a vector of dimensions t representing the probability values of the t estimates, t being 2, assuming the function h_δ(z) is as follows:

wherein,

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:

wherein z is⁽ⁱ⁾To input, y⁽ⁱ⁾Is an output;

step 4.2: pre-training Sofmax, inputting the strain subsequence into the pre-trained stacked 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:

wherein,

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 which is 2;

step 4.3: utilizing a pre-trained strain subsequence to fine-tune a connection weight matrix and a bias vector of an encoding part of the stacked self-encoder and an overall structure of the Softmax classifier, wherein a loss function during fine tuning is a cross loss function, and the method specifically comprises the following steps:

where ω is in a stacked self-encoderConnecting the weight matrix and the offset vector, theta is omega and delta, lambda₂The weight coefficient is a weight coefficient which is connected with the weight matrix and the bias vector regular term in the stacked self-encoder;

step 4.4: the Softmax classifier receives as its input the features stacked from the encoder output, outputs class 0 or 1 of the strain subsequence, 0 representing non-crack, 1 representing 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.

Images