1. A track bed fault early warning method based on DRSN and person correlation coefficient is characterized by comprising the following steps:

acquiring a plurality of vibration response signals of the track bed;

calculating the effective value of each vibration response signal, wherein the effective value calculation formula is as follows:

in the formula, Val represents an effective value, n represents the number of time points, and m represents_iRepresenting the vibration response amplitude at time point i;

taking the effective value of one vibration response signal as a reference effective value, and carrying out amplitude scaling on the rest vibration response signals according to the ratio of the effective value to the reference effective value to obtain a training sample;

training a DRSN network model by using a training sample;

carrying out amplitude scaling on vibration response signals with normal track bed structure states and unknown structure states, then respectively inputting the signals into a DRSN network model, and extracting a characteristic vector;

and calculating the correlation of the two eigenvectors by using the person correlation coefficient, and comparing the correlation with a preset threshold value to judge whether the track bed has faults or not.

2. The track bed fault early warning method based on the DRSN and person correlation coefficients as claimed in claim 1, wherein the vibration response signals are both subjected to length cutting after amplitude scaling is performed twice.

3. The method of claim 2 wherein adjacent samples formed by cutting overlap data points.

4. The method of claim 3 wherein the overlap portion is half of the cut length.

5. The track bed fault pre-warning method based on DRSN and person correlation coefficient as claimed in any one of claims 2-4,

cutting the vibration response signals with unknown track bed structure state to form n samples, and recording the finally obtained characteristic vectors as

Cutting the vibration response signal with normal track bed structure state to form i samples, and recording the finally obtained characteristic vector as S ═ S₁，s₂，…，s_i}；

Calculating S^τThe correlation between each feature vector in S and each feature vector in S has n × i correlation values in total;

and counting the abnormal proportion of the correlation values, and alarming if the abnormal proportion exceeds an early warning threshold value.

6. The method of claim 1, wherein before the DRSN network model is trained, the training samples are normalized and the values are mapped between (-1, 1).

7. The track bed fault early warning method based on the DRSN and person correlation coefficients as claimed in claim 6, wherein the normalization processing formula is as follows:

where x represents the training sample, min (x) represents the minimum value in the input sample data, and max (x) represents the maximum value in the input sample data.

8. The track bed fault early warning method based on DRSN and person correlation coefficient as claimed in claim 1, wherein in order to suppress the interference of noise in signal to feature learning, the noise is suppressed by using soft threshold mode, the expression is as follows:

in the formula: x denotes a feature of the network learning, and t denotes a threshold value.

9. The track bed fault early warning method based on DRSN and person correlation coefficients as claimed in claim 1, wherein the DRSN network model training comprises the following specific steps:

1) setting the input size of an input layer, and sequentially passing through a convolutional layer, two residual error shrinkage layers, a global average pooling layer and a full-connection layer;

2) randomly assigning network parameters according to normal distribution;

3) training a DRSN network model by taking the training sample as input;

4) dividing the training samples into a plurality of batches of batch _ size;

5) setting learning rate learning _ rate as c and maximum training time epoch as k;

6) the connection mode from the input layer to the hidden layer is set asWhere relu is the activation function, w₁As a weight matrix of the output layer to the hidden layer, b₁For the bias value of the hidden layer, i is 1,2, …, batch _ size;

7) the connection mode of the hidden layer to the output layer is set as z_(i)＝relu(w₂·y_(i)+b₂) Where relu is the activation function, w₂Weight matrix from hidden layer to output layer, b₂As the bias value of the output layer, i is 1,2, …, batch _ size;

8) the training goal of the DRSN network model is to find an optimal set of network parametersLet the loss function L (w)₁，w₂，b₁，b₂) At a minimum, the loss function expression is as follows:

in the formula: the first term on the right side of the equation represents the sum of errors of the network input data and the network output data; the second term is a regularization constraint term used for preventing over-fitting training;and z_(i)An input vector and a reconstructed vector of an ith sample respectively;to representAnd z_(i)The mean square error between, expressed as:

9) setting appropriate learning _ rate and batch _ size, and training the DRSN model through a gradient descent algorithm to make the loss function converge.

Background

Along with the growth of urban economy and the huge demand of population trip, more and more traffic ways emerge, wherein subways get more and more attention due to a series of advantages of strong bearing capacity, high speed, no occupation of ground space and the like. The service cycle of the subway line is long, and the subway track bed is used as an important underground infrastructure of subway traffic, and the structure of the subway line bed cannot avoid the degradation and the fault caused by the long-term repeated action of a train load place, the aging of materials and the coupling action of the materials; serious life and property losses and severe social influences can be caused if the subway track bed is not strictly inspected and maintained; but the cost of using manual inspection is prohibitively expensive and time consuming. In recent years, with the continuous development of communication network, signal processing, artificial intelligence and other technologies, the practicability and engineering application research of the subway monitoring system are accelerated. The subway structure health monitoring is an effective method for monitoring and evaluating the subway structure condition, and can provide decision basis and guidance for maintenance and management of a subway structure track bed. Long-term and continuous subway track bed monitoring data is collected and analyzed, and early warning and alarming can be carried out on potential diseases or faults of a service track bed structure.

Through distributed optical fiber sensing monitoring of the track beds along the line, the vibration response generated by the excited track beds in different measuring areas of the train in the running process can be collected. The combination of deep learning techniques with structural health monitoring is very useful for analyzing and processing large amounts of data collected by monitoring systems. Due to the fact that the subway track bed is different in surrounding environment and geological structure when being laid, when influence of train running speed difference on structural vibration response is ignored, vibration response characteristics generated in the same measuring area are consistent, and vibration response characteristics generated in different measuring areas are different. Under the premise, under the excitation of a train, the vibration response generated by the excitation of the same test area track bed is picked up, the correlation among vibration signals is measured, and guiding opinions can be provided for track bed structure fault early warning. Due to the complex geological environment of the subway structure, signal mutation and complex noise often exist in data collected by the health monitoring system, the correlation of original signals is directly measured, and the accuracy of track bed fault early warning is low. Therefore, the method inhibits the complex noise in the original signal and carries out the correlation measurement among the characteristics after extracting the characteristics, and has great significance for improving the accuracy of the fault early warning.

Disclosure of Invention

The invention aims to provide a track bed fault early warning method based on DRSN and person correlation coefficients, which can inhibit the noise of track bed vibration signals and improve the fault early warning reliability of a track bed structure.

The technical scheme adopted by the invention is as follows:

a track bed fault early warning method based on DRSN and person correlation coefficients comprises the following steps:

acquiring a plurality of vibration response signals of the track bed;

calculating the effective value of each vibration response signal, wherein the effective value calculation formula is as follows:

in the formula, Val represents an effective value, n represents the number of time points, and m represents_iRepresenting the vibration response amplitude at time point i;

taking the effective value of one vibration response signal as a reference effective value, and carrying out amplitude scaling on the rest vibration response signals according to the ratio of the effective value to the reference effective value to obtain a training sample;

training a DRSN network model by using a training sample;

carrying out amplitude scaling on vibration response signals with normal track bed structure states and unknown structure states, then respectively inputting the signals into a DRSN network model, and extracting a characteristic vector;

and calculating the correlation of the two eigenvectors by using the person correlation coefficient, and comparing the correlation with a preset threshold value to judge whether the track bed has faults or not.

Preferably, the vibration response signal is length cut after both amplitude scalings.

Preferably, adjacent samples formed by the cutting overlap part data points.

Preferably, the overlap is half the cut length.

Preferably, n samples are formed after cutting of the vibration response signals with unknown track bed structure states, and finally obtained characteristic vectors are recorded as

Cutting the vibration response signal with normal track bed structure state to form i samples, and recording the finally obtained characteristic vector as S ═ S₁，s₂，…，s_i}；

Calculating the correlation between each feature vector in the S tau and each feature vector in the S tau, wherein n x i correlation values exist in total;

and counting the abnormal proportion of the correlation values, and alarming if the abnormal proportion exceeds an early warning threshold value.

Preferably, before training the DRSN network model, the training samples are normalized and their values are mapped between (-1, 1).

Preferably, the normalization process formula is as follows:

where x represents the training sample, min (x) represents the minimum value in the input sample data, and max (x) represents the maximum value in the input sample data.

Preferably, in order to suppress the interference of noise in the signal to the feature learning, the noise is suppressed by using a soft threshold method, and the expression is as follows:

in the formula: x denotes a feature of the network learning, and t denotes a threshold value.

Preferably, the specific steps of the DRSN network model training are as follows:

1) setting the input size of an input layer, and sequentially passing through a convolutional layer, two residual error shrinkage layers, a global average pooling layer and a full-connection layer;

2) randomly assigning network parameters according to normal distribution;

3) training a DRSN network model by taking the training sample as input;

4) dividing the training samples into a plurality of batches of batch _ size;

5) setting learning rate learning _ rate as c and maximum training time epoch as k;

6) the connection mode from the input layer to the hidden layer is set asWhere relu is the activation function, w₁As a weight matrix of the output layer to the hidden layer, b₁For the bias value of the hidden layer, i is 1,2, …, batch _ size;

7) the connection mode of the hidden layer to the output layer is set as z_(i)＝relu(w₂·y_(i)+b₂) Where relu is the activation function, w₂Weight matrix from hidden layer to output layer, b₂As the bias value of the output layer, i is 1,2, …, batch _ size;

8) the training goal of the DRSN network model is to find an optimal set of network parametersLet the loss function L (w)₁，w₂，b₁，b₂) At a minimum, the loss function expression is as follows:

in the formula: the first term on the right side of the equation represents the sum of errors of the network input data and the network output data; the second term is a regularization constraint term used for preventing over-fitting training;and z_(i)An input vector and a reconstructed vector of an ith sample respectively;to representAnd z_(i)The mean square error between, expressed as:

9) setting appropriate learning _ rate and batch _ size, and training the DRSN model through a gradient descent algorithm to make the loss function converge.

The invention has the beneficial effects that: the invention provides a track fault early warning method aiming at the condition that monitoring data representing the structural state health of a track bed is large in quantity and can reflect fewer referenceable samples of track bed faults. Aiming at the influence of the train speed difference on the response amplitude, the effective value is proposed to weaken the difference of the amplitude. And (3) suppressing the noise of the signal and extracting the characteristics of the signal by utilizing a residual contracted convolutional neural network aiming at the conditions of serious environmental noise and sudden change of the signal.

Further, before training, the samples are subjected to normalization processing, and convergence of the network is accelerated; in order to improve the reliability of fault early warning, a statistical analysis strategy is introduced to analyze the proportion of characteristic abnormity.

Drawings

Fig. 1 is a flow chart of a track bed fault early warning method based on DRSN and person correlation coefficients according to the present invention.

FIG. 2 is a vibration signal and a cutting signal generated by exciting a measuring region of the present invention.

FIG. 3 is a schematic diagram of a residual shrinking network module of the present invention.

FIG. 4 is a schematic diagram of a network model for feature extraction in accordance with the present invention.

Detailed Description

The invention will be further described with reference to the accompanying drawings in which:

in order to better inhibit the noise of the original signal, the fault early warning reliability of the track bed structure is improved. Firstly, feature extraction is carried out on an original cutting signal based on a residual shrinkage convolution neural network. Secondly, carrying out correlation measurement on the extracted signal characteristics by using a person correlation coefficient, and realizing track bed fault early warning.

The first embodiment is as follows:

the technical scheme of the invention utilizes a residual shrinkage convolutional neural network (DRSN) to carry out feature extraction on a preprocessed original signal, utilizes a Pearson correlation coefficient to measure feature correlation, and comprises the following main steps as shown in figure 1:

(1) randomly selecting vibration response signals of a plurality of measurement areas (taking 3 as an example) for a plurality of days (taking two days as an example) from all measurement areas of the subway track bed, wherein each measurement area respectively comprises a original samples, and the sampling frequency of the samples is 1000 Hz.

(2) In order to suppress the influence of the velocity difference on the amplitude of the vibration response, the relationship between the amplitude of the response and the velocity is expressed by using an effective value, which is calculated as follows:

effective value

Wherein n represents the number of time points of the sample, m_iFor the response amplitude at the current moment。

(3) And taking the effective value of one response signal as a comparison reference, and scaling each of the rest signals according to the ratio of the effective value to the reference effective value.

(4) Through the steps, the influence of the train running speed difference on the response amplitude can be restrained.

(5) After the preprocessing, the data are cut according to 1024 lengths, the purpose of cutting is mainly to unify the lengths of the model input signals, and 512 data points are overlapped on each cut sample. As shown in fig. 2, the upper part of fig. 2 is an uncut signal, and the lower part is two cut samples, and partial data of the two samples are overlapped.

(6) After the data cutting operation, the number of samples in each measurement area is m, and the total data volume of the samples is 3 × m.

(7) The total sample size was divided into training samples and test samples in a ratio of 80% to 20%.

(8) Training residual punctured convolutional neural network DRSN using training samples and testing the performance of the network with test samples.

(9) And extracting the sample characteristics of the test set by using the trained network model.

(10) And selecting a vibration signal generated by the fact that the subway train passes through one of the 3 measuring areas in other time periods and is excited, and copying, zooming and cutting the vibration signal. Thus, 1 signal can be divided into n signals, and then the characteristics of each signal section after being cut through network extraction are calculated and recorded as S_i＝{l₁，l₂，…，l_n}。

(11) When a certain measuring area structure of the subway track bed is in a safe and stable state, the characteristics of the vibration signals generated by the excited measuring area have consistency, namely the characteristics of the vibration signals can be linked with the state of the structure. When a certain measuring area structure has a fault, the correlation difference between the vibration signal characteristics generated by the excited measuring area and the vibration signal characteristics under the normal state of the structure becomes larger. Therefore, the correlation between the characteristics of the signal in the current time period and the characteristics of the vibration signal generated by the excited normal structure can be used as a basis for judging whether the structure has a fault or not.

(12) Calculating S using person correlation coefficient_iAnd setting a certain threshold value, and if the correlation between the current characteristic and the previous characteristic is lower than the threshold value, indicating that the area may be abnormal. In addition, the abnormal proportion of the characteristic correlation can be counted, and if the abnormal proportion exceeds a preset value, an early warning needs to be sent out to remind inspectors of timely checking whether the structure of the measuring area breaks down, so that the reliability is improved.

Example two:

the invention discloses a subway track bed early warning method based on a residual shrinkage convolutional neural network and a person correlation coefficient, and figure 1 is a general flow of the invention. The object of track bed fault early warning is vibration data of 3 measuring areas for two days when the sampling rate is 1 kHz.

In order to suppress the influence of the train speed on the response amplitude, the effective value Val1 is calculated with reference to a certain vibration signal. The effective value Val2 of the further signal is then calculated and the ratio of Val2 to Val1 is taken as the scale of the current signal amplitude scaling.

Each test area contains 21 complete driving data. Each complete row data is cut to length of 1024 and the samples are overlapped at 512 points. Thus, a complete piece of driving data can be cut into 12 samples. Finally, the training samples for each test area contained 252, and the total training samples contained 3 × 252. The complete driving data waveform and the waveform after cutting are schematically shown in fig. 2.

In order to accelerate the convergence of the network, the samples are normalized before training, and the values are mapped to (-1, 1), as shown in the following formula:

in the formula: x represents the pre-processed sample after cutting, min (x) represents the minimum value in the input sample data, and max (x) represents the maximum value in the input sample data.

The invention utilizes residual shrinkage volumeThe product neural network extracts signal features, and takes the layer 7 of the network as the features of the signal, and the model structure diagram is shown in fig. 4, wherein the residual shrinkage module diagram is shown in fig. 3. In fig. 3, the soft threshold is a threshold obtained automatically by training the neural network, and is denoted as T ═ T₁，t₂，…，t_c}. Wherein t is_cRepresenting the soft threshold of the c-channel.

Due to noise in an input sample, the network learning sample characteristics are influenced, and in order to inhibit interference of noise in a signal on characteristic learning, a soft threshold mode is utilized, and an expression is shown as follows:

in the formula: x denotes a feature of the network learning, and t denotes a threshold value.

The input layer in fig. 4 represents the cut pre-processed sample, with input dimensions of 1 × 1024 × 1. The output layer represents the corresponding measuring area of the input sample.

The specific steps for training the residual shrinkage convolutional neural network model (DRSN) are as follows:

1) firstly, the input size of an input layer is set to be 1 × 1024 × 1, and the input layer sequentially passes through a convolution layer, two residual shrinkage layers, a global average pooling layer and a full connection layer.

2) In order to enable the network to better learn data characteristics and fast converge, the parameters of the network are randomly assigned according to normal distribution.

3) The DRSCNN network is then trained using the training set data as input.

4) In order to achieve a better training effect for the network, the training set is divided into a plurality of batches of batch _ size.

5) The learning rate learning _ rate is set to c, and the maximum training time epoch is set to k times.

6) The connection mode from the input layer to the hidden layer is set asi is 1,2, …, batch _ size. Where relu is the activation function, relu (x) max (0, x), w₁As a weight matrix of the output layer to the hidden layer, b₁Is the bias value of the hidden layer.

7) The connection mode of the hidden layer to the output layer is set as z_(i)＝relu(w₂·y_(i)+b₂) I is 1,2, …, batch _ size. Wherein relu is an activation function; w is a₂Weight matrix from hidden layer to output layer, b₂Is the bias value of the output layer.

8) The training goal of the DRSN network model is to find an optimal set of network parametersLoss function L (w) of network₁，w₂，b₁，b₂) At minimum, the expression is shown as follows:

in the formula: the first term on the right side of the equation represents the sum of errors of the network input data and the network output data; the second term is a regularization constraint term used for preventing over-fitting training;and z_(i)An input vector and a reconstructed vector of an ith sample respectively;to representAnd z_(i)The mean square error between, expressed as:

9) setting proper learning _ rate and batch _ size, training the DRSCNN model through a gradient descent algorithm, and completing training when the loss function convergence and the verification set accuracy reach the maximum and are converged.

After the training of the feature extraction network model is completed, the track fault early warning process is as follows:

1) selecting multiple groups of vibration signals in one of the 3 measurement areas under the normal structural state of the measurement area, preprocessing the vibration signals, then using the reconstructed vector of the 7 th network as a feature vector through a DRSCNN network, and recording the reconstructed vector as S ═ S₁，s₂，…，s_i}。

In the formula, s_iRepresenting the features of sample i and S represents the feature set of all samples.

2) Extracting characteristics of a group of vibration signals generated by exciting the same measuring area with unknown structural state through a trained DRSCNN network to obtain a characteristic vector

3) And setting a threshold value t of a correlation coefficient, and when the correlation coefficient between the two characteristics is lower than t, representing that the correlation between the two characteristics is poor and an abnormality exists.

4) Calculating S^τThe correlation between each feature in S and each feature in S has a total of n × i correlation coefficient values. And counting the abnormal proportion of the correlation coefficient, and alarming if the abnormal proportion exceeds an early warning threshold value.

In conclusion, the invention discloses a subway track bed fault early warning method based on the combination of a residual shrinkage convolutional neural network and a person correlation coefficient. Firstly, preprocessing the acquired original distributed fiber bragg grating vibration signals representing the state of the track bed, and inhibiting the influence of speed difference on response amplitude. Secondly, extracting the features of the cut preprocessed signals by using a residual shrinkage convolutional neural network, and performing correlation measurement on the extracted features by using a person correlation coefficient. And then, counting the signal proportion of the correlation abnormality based on a statistical analysis strategy, and when the abnormal proportion of a certain area exceeds a preset threshold value, sending out a fault early warning to the track bed of the area.

It will be understood by those skilled in the art that the foregoing is merely a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included within the scope of the present invention.