CN114266271A - Distributed optical fiber vibration signal mode classification method and system based on neural network - Google Patents
Distributed optical fiber vibration signal mode classification method and system based on neural network Download PDFInfo
- Publication number
- CN114266271A CN114266271A CN202111521196.5A CN202111521196A CN114266271A CN 114266271 A CN114266271 A CN 114266271A CN 202111521196 A CN202111521196 A CN 202111521196A CN 114266271 A CN114266271 A CN 114266271A
- Authority
- CN
- China
- Prior art keywords
- layer
- neural network
- convolutional
- optical fiber
- vibration
- 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.)
- Pending
Links
Images
Abstract
The invention discloses a distributed optical fiber vibration signal mode classification method based on a neural network, which comprises the following steps: acquiring alarm real-time data sent by a sensor, and carrying out low-pass filtering on the data to eliminate high-frequency noise; positioning an alarm point according to the characteristics of the signal, and intercepting a signal segment with fixed length and meeting the requirement of the input length of the neural network according to the positioning result; normalizing the signal segments and inputting the normalized signal segments into a neural network; and training the network by adopting a batch random gradient descent algorithm to obtain a neural network model and parameters for final judgment. The invention can realize real-time judgment of the reason for triggering the alarm after the distributed optical fiber sensor generates the alarm signal, is convenient for managers to quickly position the place and the reason for the alarm, is convenient for the managers to quickly make corresponding reactions, has simple method and is suitable for general popularization. The invention also discloses a distributed optical fiber vibration signal mode classification system based on the neural network.
Description
Technical Field
The invention belongs to the technical field of signal processing, and particularly relates to a distributed optical fiber vibration signal mode classification method and system based on a neural network.
Background
The distributed optical fiber vibration monitoring system is widely applied to the fields of tunnel pipe gallery security and protection positioning, electric power system cable monitoring, information pipeline online monitoring, high-speed railway security and protection systems, long-distance pipeline security and protection, key facility perimeter security and protection and the like.
In the existing distributed optical fiber vibration monitoring scheme, field seedlings and the like adopt radial basis functions combined with empirical mode decomposition to realize the identification of four events of shearing, shaking, knocking and climbing, and the average identification rate reaches 85.75 percent; lixichen and the like utilize short-time Fourier transform and singular value decomposition to realize feature extraction, and a support vector machine is adopted to identify and classify vibration signals, wherein the identification accuracy is over 90 percent; xu et al propose a feature extraction scheme of multi-parameter fusion features, and the classification of four types of events such as trampling and knocking is realized by using a support vector machine, and the average recognition rate is more than 93%.
However, the above-mentioned identification schemes have some non-negligible drawbacks: firstly, they all require complex signal processing methods, and the system design is complex; secondly, the above schemes all use a feature engineering method, and the performance of the system depends on the effectiveness of the selected features, so that the adaptability to different vibration types is poor.
Disclosure of Invention
Aiming at the defects or improvement requirements of the prior art, the invention provides a neural network-based distributed optical fiber vibration signal mode classification method and system, and aims to solve the technical problems that the system design is complex due to the fact that a complex signal processing method is needed in the existing distributed optical fiber vibration monitoring scheme, and the system performance depends on the effectiveness of selected features due to the fact that feature engineering methods are used in all the schemes, and the adaptability to different vibration types is poor.
To achieve the above object, according to one aspect of the present invention, there is provided a neural network-based distributed optical fiber vibration signal mode classification method, including the steps of:
(1) sampling was performed using a distributed vibration sensing DAS apparatus to obtain num1 fibre vibration signals { x (1), x (2), …, x (num1) }, and x (i) }, { x1(i),x2(i),…,xnum2(i) Where i ∈ [1, num 1]]，xj(i) J ∈ [1, num2, j ∈ [1, num2 ] representing the jth vibration value of the ith fiber vibration signal]Num2 denotes the vector dimension in the ith fiber vibration signal;
(2) preprocessing a plurality of optical fiber vibration signals obtained by sampling in the step (1) to obtain a vibration sample set formed by the preprocessed plurality of optical fiber vibration signals;
(3) and (3) inputting the vibration sample set obtained in the step (2) into a trained one-dimensional convolutional neural network to obtain a corresponding classification result.
Preferably, the sampling frequency F of the DAS equipment in the step (1)samplingEqual to:
wherein R isspaceAnd VfiberThe spatial resolution of the DAS device and the propagation speed of the laser light in the fiber are separately indicated.
Preferably, the time resolution Rtime of the DAS device samples is equal to:
wherein NumaverageAnd FtriggerRespectively representing the accumulated times of accumulated noise reduction and the trigger frequency of sampling of the DAS equipment, and FtriggerEqual to:
wherein NumsamplingIs 8192.
Preferably, step (2) comprises the sub-steps of:
(2-1) setting the counter cnt1 to 1;
(2-2) judging whether cnt1 is larger than the total num1 of the optical fiber vibration signals obtained in the step (1), if so, entering the step (2-5), otherwise, entering the step (2-3);
(2-3) performing convolution operation on each vibration value in the cnt1 th optical fiber vibration signal x (cnt1) to obtain a low-pass signal y (cntl) corresponding to the cnt1 th optical fiber vibration signal;
(2-4) setting the counter cnt1 ═ cnt1+1, and returning to step (2-2);
(2-5) acquiring the maximum value max of num1 × num2 numbers included in a low-pass signal sequence { y (1), y (2), …, y (num1) } formed by low-pass signals corresponding to all optical fiber vibration signals;
(2-6) creating an empty signal list z { }, and setting a counter cnt2 { (1);
(2-7) judging whether cnt2 is greater than the total num1 (namely 2500) of the optical fiber vibration signals obtained in the step (1), if so, entering the step (2-10), otherwise, entering the step (2-8);
(2-8) calculating the maximum value max (cnt2) of num2 numbers of low-pass signals y (cnt2) corresponding to the cnt 2-th optical fiber vibration signal, judging whether max (cnt2) > alpha × max exists, if yes, adding the low-pass signal y (cnt2) corresponding to the cnt 2-th optical fiber vibration signal to the tail of the signal list z, and entering the step (2-9), otherwise, entering the step (2-9), wherein the value of the intermediate parameter alpha is 0.7;
(2-9) setting the counter cnt2 ═ cnt2+1, and returning to step (2-7);
(2-10) setting the counter cnt3 to 1;
(2-11) judging whether cnt3 is greater than the total number num3 of the low-pass signals in the list z, if so, entering the step (2-13), otherwise, entering the step (2-12);
(2-12) normalizing each value zk (cn3) of the cnt 3-th low-pass signal z (cnt3) in the list z to a [0, 1] interval to obtain a normalized vibration signal t (cnt 3);
(2-13) setting the counter cnt3 ═ cnt3+1, and returning to step (2-11);
(2-14) creating an empty sample list s { }, and setting a counter cnt4 { (1);
(2-15) judging whether cnt4 is greater than the total number num3 of the low-pass signals in the list z, if so, entering the step (2-22), otherwise, entering the step (2-16);
(2-16) obtaining the maximum value in the normalized vibration signal t (cnt4)
And the serial number index (cnt4) of the maximum value in the vibration signal t (cnt 4);
(2-17) judging whether the sequence number index (cnt4) acquired in the step (2-16) is less than 51, if so, setting the index (cnt4) to 51, then entering the step (2-18), otherwise, setting the index (cnt4) to num2-50, and then entering the step (2-18);
(2-18) mixing (t)index(cnt4)-50(cnt4)，...，tindex(cnt4)+50(cnt4)) to the end of the sample list s;
(2-19) setting the counter cnt4 ═ cnt4+1, and returning to step (2-15);
and (2-20) taking the sample list s as a finally obtained vibration sample set.
Preferably, the convolution operation on x (cnt1) in step (2-3) is performed by using the following formula:
wherein y isj(cnt1) shows the result of performing convolution operation on the jth vibration value in the cnt1 th fiber vibration signal x (cnt1),
and the value of the parameter sigma is 2;
in step (2-12), t (cnt3) ═ t1(cnt3)，t2(cnt3)，…，tnum2(cnt3) }, where t isj(cnt3) represents the jth value in the cnt3 low-pass signal z (cnt3) in list z, and has:
where maximize indicates taking the maximum value.
Preferably, the one-dimensional convolutional neural network comprises six layers, wherein the first, second and third layers are convolutional layers, and each layer consists of two convolutional layers and one maximum pooling layer and is used for extracting the characteristics of the vibration signal; the fourth layer is a Flatten layer which changes the characteristics obtained by the third convolution layer into a one-dimensional structure; the fifth layer is a fully-connected layer to classify the signal features, and the sixth layer is a Softmax layer which normalizes the results obtained by the fifth layer into probabilities.
Preferably, the first layer of the one-dimensional convolutional neural network is a first convolutional block, which consists of two convolutional layers and one max-pooling layer. Each convolutional layer uses a length-3 convolutional kernel, 16 filters per convolutional layer, the first convolutional layer does not use an excitation function, and the second convolutional layer uses a tanh excitation function. The window length of the final pooling layer is 3;
the second layer of the one-dimensional convolutional neural network is a second convolutional block, the input of which is the output of the first convolutional block, the second convolutional block consisting of two convolutional layers and one max-pooling layer. Each convolutional layer uses a length-3 convolutional kernel, 64 filters per convolutional layer, and a tanh excitation function per convolutional layer. The window length of the final pooling layer is 3;
the third layer of the one-dimensional convolutional neural network is a third convolutional block, the input of which is the output of the second convolutional block, and the third convolutional block consists of two convolutional layers and one max-pooling layer. Each convolutional layer uses a length-3 convolutional kernel, 64 filters per convolutional layer, and a tanh excitation function per convolutional layer. The window length of the final pooling layer is 3;
the fourth layer of the one-dimensional convolutional neural network is a Flatten layer, the input of the one-dimensional convolutional neural network is the output of a third convolutional block, and the feature obtained by the third convolutional block is elongated into a one-dimensional structure;
the fifth layer of the one-dimensional convolutional neural network is a full connection layer, and the input of the full connection layer is the output of a Flatten layer, so that the classification of signal characteristics is realized;
the sixth layer of the one-dimensional convolutional neural network is the Softmax layer, whose input is the output of the fully-connected layer, which normalizes the output to the probability of a class.
Preferably, the one-dimensional convolutional neural network is trained by adopting the following steps:
A. collecting vibration signals of a plurality of categories as a data set for training a neural network, and carrying out the vibration signals according to the following steps of 7: 3, dividing the ratio into a training set and a test set;
B. inputting the training set obtained by dividing in the step A into a one-dimensional convolution neural network;
C. updating and optimizing the weight parameters and the bias parameters of each layer in the one-dimensional convolutional neural network by using a BP algorithm to obtain an updated one-dimensional convolutional neural network;
D. and C, performing iterative training on the one-dimensional convolutional neural network updated in the step C until the loss function of the one-dimensional convolutional neural network reaches the minimum.
Preferably, in step C, the initial value of the weight parameter is a random value close to 0, the initial value of the bias parameter is set to 0, and the back propagation algorithm is to use an Adam optimizer and an adaptive learning rate.
In step D, cross entropy is used as a loss function, namely:
where batch _ size is the volume of the batch, the invention selects batchsize32; category _ num is the number of vibration types in the training set, licIs a sign function, if the true class of the ith vibration signal in the training set is equal to the vibration type c, then licGet 1, otherwise get 0, picIs the probability that the predicted class of the i-th vibration signal s (i) in the training set is equal to the vibration type c.
According to another aspect of the present invention, there is provided a neural network-based distributed optical fiber vibration signal mode classification system, including:
a first module for sampling using a distributed vibration sensing DAS device to obtain num1 fibre vibration signals { x (1), x (2),. ·, x (num1) }, and x (i) ═ x (x 1) }1(i)，x2(i)，...，xnum2(i) Where i ∈ [1, numl) ]]，xj(i) J ∈ [1, num2, j ∈ [1, num2 ] representing the jth vibration value of the ith fiber vibration signal]Num2 denotes the vector dimension in the ith fiber vibration signal;
the second module is used for preprocessing the plurality of optical fiber vibration signals sampled by the first module to obtain a vibration sample set formed by the plurality of preprocessed optical fiber vibration signals;
and the third module is used for inputting the vibration sample set obtained by the second module into the trained one-dimensional convolutional neural network so as to obtain a corresponding classification result.
In general, compared with the prior art, the above technical solution contemplated by the present invention can achieve the following beneficial effects:
(1) in addition, because the method does not directly use a characteristic engineering method, the method has better adaptability to different vibration types;
(2) the invention adopts the steps (2-10) to (2-13) to standardize the vibration value of the vibration signal to the closed interval [0, 1], thereby overcoming the problem of signal attenuation in the long-distance optical fiber;
(3) according to the method, the steps (2-14) to (2-19) are adopted, and a small section of signals before and after the maximum vibration value of the vibration signal are intercepted to be used as the input of the classifier, so that the classifier can focus on the most obvious part of the signal, and the classification performance of the model is improved;
(4) according to the invention, alarm real-time data sent by a sensor are acquired, and low-pass filtering is carried out on the data to eliminate high-frequency noise; positioning an alarm point according to the characteristics of the signal, and intercepting a signal segment with fixed length and meeting the requirement of the input length of the neural network according to the positioning result; normalizing the signal segments and inputting the normalized signal segments into a neural network; and training the network by adopting a batch random gradient descent algorithm to obtain a neural network model and parameters for final judgment. The alarm signal mode recognition system of the distributed optical fiber sensor based on the neural network can realize real-time judgment of the reason for causing the alarm after the distributed optical fiber sensor generates the alarm signal, is convenient for managers to quickly position the place where the alarm occurs and the reason for causing the alarm, is convenient for the managers to quickly make corresponding reactions, has a simple method, and is suitable for general popularization.
Drawings
FIG. 1 is a flow chart of a neural network-based distributed optical fiber vibration signal mode classification method of the present invention;
FIG. 2 is an architectural diagram of a one-dimensional convolutional neural network used in the method of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
The invention adopts a deep learning method to avoid the characteristic extraction step required by the prior method, uses a multi-layer one-dimensional convolution application network to automatically learn the classification characteristic, and then uses a full connection layer and a softmax layer to realize the classification of the vibration reason.
As shown in fig. 1, the present invention provides a distributed optical fiber vibration signal mode classification method based on a neural network, including the following steps:
(1) sampling is performed using Distributed fiber optic Sensing (DAS) equipment (in the present invention, the sampling period is 10 seconds) to obtain num1 fiber vibration signals { x (1), x (2), …, x (num1) }, and x (i) } x (i) is obtained1(i)，x2(i)，...，xnum2(i) Where i ∈ [1, num 1]]，xj(i) J ∈ [1, num2, j ∈ [1, num2 ] representing the jth vibration value of the ith fiber vibration signal]Num2 represents the vector dimension in the ith optical fiber vibration signal, and num2 is 8192;
specifically, the sampling frequency F of the DAS device in this stepsamplingThe formula of (c) is shown as follows:
wherein R isspaceAnd VfiberThe spatial resolution of the DAS device and the propagation speed of the laser light in the fiber are separately indicated. The spatial resolution typically used in experiments is 1m, so the sampling frequency of the DAS device is
The time resolution Rtime of the DAS device samples is determined by:
wherein NumaverageAnd FtriggerRespectively representThe DAS equipment accumulates the accumulated times of noise reduction and the trigger frequency of sampling, and in practical use, NumaverageNormally set to 8 times, and the sampling trigger frequency FtriggerDepends on the sampling frequency Fsampling and the number Num of sampling points of a single triggeraverage，FtriggerIs determined by:
in the experiment, Num is usually setsampling8192, the frequency F is triggeredtriggerMaximum is
So the final spatial resolution RtimeCan be minimized
That is, 1525 fiber vibration signals can be collected in 1 second at most. In practical use, the minimum time resolution is not generally used, and in the experiment, the trigger frequency is set to be Ftrigger2KHz, time resolution
That is, 250 optical fiber vibration signals are collected in 1 second, so that in a sampling period of 10 seconds, 2500 optical fiber vibration signals are collected in total, that is, num1 is 2500.
(2) Preprocessing a plurality of optical fiber vibration signals obtained by sampling in the step (1) to obtain a vibration sample set formed by the preprocessed plurality of optical fiber vibration signals;
the method comprises the following substeps:
(2-1) setting the counter cnt1 to 1;
(2-2) judging whether cnt1 is greater than the total num1 (namely 2500) of the optical fiber vibration signals obtained in the step (1), if so, entering the step (2-5), otherwise, entering the step (2-3);
(2-3) performing convolution operation on each vibration value in the cnt1 th optical fiber vibration signal x (cnt1) to obtain a low-pass signal y (cnt1) corresponding to the cnt1 th optical fiber vibration signal;
specifically, the convolution operation on x (cnt1) in this step uses the following formula:
wherein y isj(cnt1) shows the result of performing convolution operation on the jth vibration value in the cnt1 th fiber vibration signal x (cnt1),
and the value of the parameter sigma is 2;
(2-4) setting the counter cnt1 ═ cnt1+1, and returning to step (2-2);
the steps (2-1) to (2-4) have the advantages that the convolution operation of one Gaussian kernel is used for restraining high-frequency noise in the signal, so that the triggering alarm positioning in the subsequent steps is more accurate.
(2-5) acquiring the maximum value max of num1 × num2 numbers included in a low-pass signal sequence { y (1), y (2), …, y (num1) } formed by low-pass signals corresponding to all optical fiber vibration signals;
specifically, after the processing in the above steps (2-2) to (2-4), each of the low-pass signals in the obtained low-pass signal sequence includes num2 (i.e., 8192) values (i.e., convolution results, which correspond to num2 vibration values in each of the low-pass signals), and this step is to find the maximum value from the total num ml num2 values.
(2-6) creating an empty signal list z { }, and setting a counter cnt2 { (1);
(2-7) judging whether cnt2 is greater than the total number numl (namely 2500) of the optical fiber vibration signals obtained in the step (1), if so, entering the step (2-10), otherwise, entering the step (2-8);
(2-8) calculating the maximum value max (cnt2) of num2 numbers of low-pass signals y (cnt2) corresponding to the cnt 2-th optical fiber vibration signal, judging whether max (cnt2) > alpha × max exists, if yes, adding the low-pass signal y (cnt2) corresponding to the cnt 2-th optical fiber vibration signal to the tail of the signal list z, and entering the step (2-9), otherwise, entering the step (2-9), wherein the value of the intermediate parameter alpha is 0.7;
(2-9) setting the counter cnt2 ═ cnt2+1, and returning to step (2-7);
the steps (2-5) to (2-9) have the advantages that the signals most relevant to the vibration are selected from the collected signals through a threshold operation, the selected signals can better represent the characteristics of the vibration category to which the signals belong, and the classification accuracy of the classifier is improved.
(2-10) setting the counter cnt3 to 1;
(2-11) judging whether cnt3 is greater than the total number num3 of the low-pass signals in the list z, if so, entering the step (2-13), otherwise, entering the step (2-12);
(2-12) normalizing each value zk (cn3) of the cnt 3-th low-pass signal z (cnt3) in the list z to a [0, 1] interval to obtain a normalized vibration signal t (cnt 3);
in this step, t (cnt3) ═ t1(cnt3)，t2(cnt3)，…，tnum2(cnt3) }, where t isj(cnt3) represents the jth value in the cnt3 low-pass signal z (cnt3) in list z, and has:
where maximize indicates taking the maximum value.
(2-13) setting the counter cnt3 ═ cnt3+1, and returning to step (2-11);
the steps (2-10) to (2-13) have the advantages that the numerical value of the vibration signal is normalized to the closed interval [0, 1] through a data normalization operation, so that the influence of the attenuation difference of the optical signal caused by different distances between the vibration trigger event and the light source on the training of the classification model is avoided, and the model characteristic acquisition capability is improved; meanwhile, the data distribution is consistent with the data distribution assumed by the machine learning software library, so that the model training efficiency is improved.
(2-14) creating an empty sample list s { }, and setting a counter cnt4 { (1);
(2-15) judging whether cnt4 is greater than the total number num3 of the low-pass signals in the list z, if so, entering the step (2-22), otherwise, entering the step (2-16);
(2-16) obtaining the maximum value in the normalized vibration signal t (cnt4)
And the serial number index (cnt4) of the maximum value in the vibration signal t (cnt 4);
(2-17) judging whether the sequence number index (cnt4) acquired in the step (2-16) is less than 51, if so, setting the index (cnt4) to 51, then entering the step (2-18), otherwise, setting the index (cnt4) to num2-50, and then entering the step (2-18);
the purpose of this step is to intercept a segment of the sub-signal containing max (cnt 4). The normalized vibration signal t (cnt4) is a sequence with the length of num2 (8192 in the invention), and vibration usually only occurs in a small segment of the whole signal, and in order to improve the accuracy of classification, a small segment with the strongest signal (the small segment with the signal length of 101 in the invention) needs to be intercepted and input into the one-dimensional convolutional neural network. To ensure that the (2-18) step is effective in intercepting small segments of signal of length 101, it is necessary to deal with the situation where max (cnt4) is near the beginning or end of signal t (cnt 4). For example, if the index (cnt4) is 30, the index of the first signal value intercepted in step (2-18) will be index (cnt4) -50-20, which is obviously illegal, so that the index (cnt4) needs to be adjusted to 51, and thus the index of the first signal value intercepted in step (2-18) will be index (cnt4) -50-51-50-1.
(2-18) mixing (t)index(cnt4)-50(cnt4)，...，tindex(cnt4)+50(cnt4)) to the end of the sample list s;
(2-19) setting the counter cnt4 ═ cnt4+1, and returning to step (2-15);
the advantage of the above steps (2-14) to (2-19) is that the classifier is trained using only a small segment of the signal closest to the vibration trigger event, so that the classifier can focus on the most significant part of the signal, thereby improving the classification performance of the model.
And (2-20) taking the sample list s as a finally obtained vibration sample set.
(3) And (3) inputting the vibration sample set obtained in the step (2) into a trained one-dimensional convolutional neural network to obtain a corresponding classification result.
As shown in fig. 2, the one-dimensional convolutional neural network of the present invention includes six layers, wherein the first, second and third layers are convolutional layers, and each layer is composed of two convolutional layers plus a maximum pooling layer and is used for extracting the characteristics of the vibration signal; the fourth layer is a Flatten layer which "stretches" the features obtained from the third convolutional layer into a one-dimensional structure; the fifth layer is a fully connected layer that implements classification of signal features, and the sixth layer is a Softmax layer that normalizes the results of the fifth layer into probabilities.
Specifically, the first layer of the one-dimensional convolutional neural network of the present invention is the first convolutional block, which consists of two convolutional layers and one max-pooling layer. Each convolutional layer uses a length-3 convolutional kernel, 16 filters per convolutional layer, the first convolutional layer does not use an excitation function, and the second convolutional layer uses a tanh excitation function. The window length of the final pooling layer is 3;
the second layer of the one-dimensional convolutional neural network of the present invention is the second convolutional block, the input of which is the output of the first convolutional block, the second convolutional block is composed of two convolutional layers and one max-pooling layer. Each convolutional layer uses a length-3 convolutional kernel, 64 filters per convolutional layer, and a tanh excitation function per convolutional layer. The window length of the final pooling layer is 3;
the third layer of the one-dimensional convolutional neural network of the present invention is the third convolutional block, the input of which is the output of the second convolutional block, and the third convolutional block consists of two convolutional layers and one maximum pooling layer. Each convolutional layer uses a length-3 convolutional kernel, 64 filters per convolutional layer, and a tanh excitation function per convolutional layer. The window length of the final pooling layer is 3;
the fourth layer of the one-dimensional convolution neural network is a Flatten layer, the input of the one-dimensional convolution neural network is the output of the third convolution block, and the feature obtained by the third convolution block is elongated into a one-dimensional structure;
the fifth layer of the one-dimensional convolution neural network is a full connection layer, the input of the full connection layer is the output of a Flatten layer, and classification of signal characteristics is realized;
the sixth layer of the one-dimensional convolutional neural network of the present invention is the Softmax layer, the input of which is the output of the fully-connected layer, normalizing the output to the probability of a class.
The one-dimensional convolution neural network is obtained by adopting the following steps:
(1) collecting vibration signals of a plurality of categories as a data set for training a neural network, and carrying out the vibration signals according to the following steps of 7: 3, dividing the ratio into a training set and a test set;
specifically, the invention uses 5 types of vibration signals, namely normal, walking, jumping, trolley rolling and wind blowing; it should be understood that the present invention is not limited to the above category 5 samples, and any factor that causes DAS to generate an alarm can be included within the scope of the present invention;
(2) inputting the training set obtained by dividing in the step (1) into a one-dimensional convolution neural network;
(3) updating and optimizing the weight parameter and the bias parameter of each layer in the one-dimensional convolutional neural network by using a Back Propagation (BP) algorithm to obtain an updated one-dimensional convolutional neural network;
specifically, the initial value of the weight parameter is a random value close to 0, and the initial value of the bias parameter is set to 0;
the back propagation algorithm in this step is to use an Adam optimizer and use an adaptive learning rate.
(4) Performing iterative training on the one-dimensional convolutional neural network updated in the step (3) until the loss function of the one-dimensional convolutional neural network reaches the minimum;
specifically, this step is to use the cross entropy as a loss function, that is:
where batch _ size is the volume of the batch, the invention selects batchsize32; category _ num is the number of types of vibration in the training set, and is 5 in the invention; licIs a symbolic function, if the true class of sample s (i) (i.e. the ith vibration signal in the training set) is equal to vibration type c, then licGet 1, otherwise get 0, picIs the probability that the predicted class of sample s (i) is equal to vibration type c.
Specifically, 32 samples were randomly picked per batch, with 40 generations of training each time.
Because various samples in the total sample are unbalanced, different weights need to be assigned to each sample in the batch stochastic gradient descent algorithm, and the weight w of the ith sample in the training setiIs calculated by the following equation.
Wherein n isjIs the number of class j samples.
The samples are input into the one-dimensional convolutional neural network trained by the method, the network automatically identifies the intrusion type and gives an identification result, and the identification accuracy rate reaches 95.75%.
It will be understood by those skilled in the art that the foregoing is only 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 in the scope of the present invention.
Claims (10)
1. A distributed optical fiber vibration signal mode classification method based on a neural network is characterized by comprising the following steps:
(1) sampling was performed using a distributed vibration sensing DAS apparatus to obtain num1 fibre vibration signals { x (1), x (2), …, x (num1) }, and x (i) }, { x1(i),x2(i),…,xnum2(i) Where i ∈ [1, num 1]]，xj(i) J ∈ [1, num2, j ∈ [1, num2 ] representing the jth vibration value of the ith fiber vibration signal]Num2 denotes the vector dimension in the ith fiber vibration signal;
(2) preprocessing a plurality of optical fiber vibration signals obtained by sampling in the step (1) to obtain a vibration sample set formed by the preprocessed plurality of optical fiber vibration signals;
(3) and (3) inputting the vibration sample set obtained in the step (2) into a trained one-dimensional convolutional neural network to obtain a corresponding classification result.
2. The neural network-based distributed optical fiber vibration signal mode classification method according to claim 1, wherein the frequency F sampled by the DAS device in the step (1)samplingEqual to:
wherein R isspaceAnd VfiberThe spatial resolution of the DAS device and the propagation speed of the laser light in the fiber are separately indicated.
3. The neural network-based distributed optical fiber vibration signal mode classification method according to claim 1 or 2, characterized in that the time resolution Rtime of the DAS device samples is equal to:
wherein NumaverageAnd FtriggerRespectively representing the accumulated times of accumulated noise reduction and the trigger frequency of sampling of the DAS equipment, and FtriggerEqual to:
wherein NumsamplingIs 8192.
4. The neural network-based distributed optical fiber vibration signal mode classification method according to any one of claims 1 to 3, wherein the step (2) comprises the following sub-steps:
(2-1) setting the counter cnt1 to 1;
(2-2) judging whether cnt1 is larger than the total num1 of the optical fiber vibration signals obtained in the step (1), if so, entering the step (2-5), otherwise, entering the step (2-3);
(2-3) performing convolution operation on each vibration value in the cnt1 th optical fiber vibration signal x (cnt1) to obtain a low-pass signal y (cnt1) corresponding to the cnt1 th optical fiber vibration signal;
(2-4) setting the counter cnt1 ═ cnt1+1, and returning to step (2-2);
(2-5) acquiring the maximum value max of num1 × num2 numbers included in a low-pass signal sequence { y (1), y (2), …, y (num1) } formed by low-pass signals corresponding to all optical fiber vibration signals;
(2-6) creating an empty signal list z { }, and setting a counter cnt2 { (1);
(2-7) judging whether cnt2 is greater than the total num1 (namely 2500) of the optical fiber vibration signals obtained in the step (1), if so, entering the step (2-10), otherwise, entering the step (2-8);
(2-8) calculating the maximum value max (cnt2) of num2 numbers of low-pass signals y (cnt2) corresponding to the cnt 2-th optical fiber vibration signal, judging whether max (cnt2) > alpha × max exists, if yes, adding the low-pass signal y (cnt2) corresponding to the cnt 2-th optical fiber vibration signal to the tail of the signal list z, and entering the step (2-9), otherwise, entering the step (2-9), wherein the value of the intermediate parameter alpha is 0.7;
(2-9) setting the counter cnt2 ═ cnt2+1, and returning to step (2-7);
(2-10) setting the counter cnt3 to 1;
(2-11) judging whether cnt3 is greater than the total number num3 of the low-pass signals in the list z, if so, entering the step (2-13), otherwise, entering the step (2-12);
(2-12) Each value z in the cnt 3-th low-pass signal z (cnt3) in the list zk(cn3) normalized to [0, 1]]Interval to obtain normalized vibration signal t (cnt 3);
(2-13) setting the counter cnt3 ═ cnt3+1, and returning to step (2-11);
(2-14) creating an empty sample list s { }, and setting a counter cnt4 { (1);
(2-15) judging whether cnt4 is greater than the total number num3 of the low-pass signals in the list z, if so, entering the step (2-22), otherwise, entering the step (2-16);
(2-16) obtaining the maximum value in the normalized vibration signal t (cnt4)
And the serial number index (cnt4) of the maximum value in the vibration signal t (cnt 4);
(2-17) judging whether the sequence number index (cnt4) acquired in the step (2-16) is less than 51, if so, setting the index (cnt4) to 51, then entering the step (2-18), otherwise, setting the index (cnt4) to num2-50, and then entering the step (2-18);
(2-18) mixing (t)index(cnt4)-50(cnt4)，...，tindex(cnt4)+50(cnt4)) to the end of the sample list s;
(2-19) setting the counter cnt4 ═ cnt4+1, and returning to step (2-15);
and (2-20) taking the sample list s as a finally obtained vibration sample set.
5. The neural network-based distributed optical fiber vibration signal mode classification method of claim 4,
the convolution operation for x (cnt1) in step (2-3) uses the following formula:
wherein y isj(cnt1) shows the result of performing convolution operation on the jth vibration value in the cnt1 th fiber vibration signal x (cnt1),
and the value of the parameter sigma is 2;
in step (2-12), t (cnt3) ═ t1(cnt3)，t2(cnt3)，...，tnum2(cnt3) }, where t isj(cnt3) represents the jth value in the cnt3 low-pass signal z (cnt3) in list z, and has:
where maximize indicates taking the maximum value.
6. The neural network-based distributed optical fiber vibration signal mode classification method according to claim 1, characterized in that the one-dimensional convolutional neural network comprises six layers, wherein the first, second and third layers are convolutional layers, each layer is composed of two convolutional layers plus one maximum pooling layer, and is used for extracting the characteristics of the vibration signal; the fourth layer is a Flatten layer which changes the characteristics obtained by the third convolution layer into a one-dimensional structure; the fifth layer is a fully-connected layer to classify the signal features, and the sixth layer is a Softmax layer which normalizes the results obtained by the fifth layer into probabilities.
7. The neural network-based distributed optical fiber vibration signal mode classification method of claim 6,
the first layer of the one-dimensional convolutional neural network is the first convolutional block, which consists of two convolutional layers and one max-pooling layer. Each convolutional layer uses a length-3 convolutional kernel, 16 filters per convolutional layer, the first convolutional layer does not use an excitation function, and the second convolutional layer uses a tanh excitation function. The window length of the final pooling layer is 3;
the second layer of the one-dimensional convolutional neural network is a second convolutional block, the input of which is the output of the first convolutional block, the second convolutional block consisting of two convolutional layers and one max-pooling layer. Each convolutional layer uses a length-3 convolutional kernel, 64 filters per convolutional layer, and a tanh excitation function per convolutional layer. The window length of the final pooling layer is 3;
the third layer of the one-dimensional convolutional neural network is a third convolutional block, the input of which is the output of the second convolutional block, and the third convolutional block consists of two convolutional layers and one max-pooling layer. Each convolutional layer uses a length-3 convolutional kernel, 64 filters per convolutional layer, and a tanh excitation function per convolutional layer. The window length of the final pooling layer is 3;
the fourth layer of the one-dimensional convolutional neural network is a Flatten layer, the input of the one-dimensional convolutional neural network is the output of a third convolutional block, and the feature obtained by the third convolutional block is elongated into a one-dimensional structure;
the fifth layer of the one-dimensional convolutional neural network is a full connection layer, and the input of the full connection layer is the output of a Flatten layer, so that the classification of signal characteristics is realized;
the sixth layer of the one-dimensional convolutional neural network is the Softmax layer, whose input is the output of the fully-connected layer, which normalizes the output to the probability of a class.
8. The method for classifying the vibration signal mode of the distributed optical fiber based on the neural network as claimed in any one of claims 1 to 7, wherein the one-dimensional convolution neural network is obtained by adopting the following training steps:
A. collecting vibration signals of a plurality of categories as a data set for training a neural network, and carrying out the vibration signals according to the following steps of 7: 3, dividing the ratio into a training set and a test set;
B. inputting the training set obtained by dividing in the step A into a one-dimensional convolution neural network;
C. updating and optimizing the weight parameters and the bias parameters of each layer in the one-dimensional convolutional neural network by using a BP algorithm to obtain an updated one-dimensional convolutional neural network;
D. and C, performing iterative training on the one-dimensional convolutional neural network updated in the step C until the loss function of the one-dimensional convolutional neural network reaches the minimum.
9. The neural network-based distributed optical fiber vibration signal mode classification method of claim 8,
in the step C, the initial value of the weight parameter is a random value close to 0, the initial value of the bias parameter is set to 0, and the back propagation algorithm adopts an Adam optimizer and uses the self-adaptive learning rate.
In step D, cross entropy is used as a loss function, namely:
where batch _ size is the volume of the batch, the invention selects batchsize32; cayegory _ num is the number of vibration types in the training set, licIs a sign function, if the true class of the ith vibration signal in the training set is equal to the vibration type c, then licGet 1, otherwise get 0, picIs the probability that the predicted class of the i-th vibration signal s (i) in the training set is equal to the vibration type c.
10. A distributed optical fiber vibration signal mode classification system based on a neural network is characterized by comprising the following components:
a first module for sampling using a distributed vibration sensing DAS device to obtain num1 fibre vibration signals { x (1), x (2), …, x (num1) }, and x (i) } x (x 1)1(i),x2(i),…,xnum2(i) Where i ∈ [1, num 1]]，xj(i) J ∈ [1, num2, j ∈ [1, num2 ] representing the jth vibration value of the ith fiber vibration signal]Num2 denotes the vector dimension in the ith fiber vibration signal;
the second module is used for preprocessing the plurality of optical fiber vibration signals sampled by the first module to obtain a vibration sample set formed by the plurality of preprocessed optical fiber vibration signals;
and the third module is used for inputting the vibration sample set obtained by the second module into the trained one-dimensional convolutional neural network so as to obtain a corresponding classification result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111521196.5A CN114266271A (en) | 2021-12-13 | 2021-12-13 | Distributed optical fiber vibration signal mode classification method and system based on neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111521196.5A CN114266271A (en) | 2021-12-13 | 2021-12-13 | Distributed optical fiber vibration signal mode classification method and system based on neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114266271A true CN114266271A (en) | 2022-04-01 |
Family
ID=80827656
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111521196.5A Pending CN114266271A (en) | 2021-12-13 | 2021-12-13 | Distributed optical fiber vibration signal mode classification method and system based on neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114266271A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115424399A (en) * | 2022-09-02 | 2022-12-02 | 复旦大学 | Distributed optical fiber well lid switch monitoring system and method |
| CN116026449A (en) * | 2023-03-30 | 2023-04-28 | 广东恒志信息技术有限公司 | Vibration positioning monitoring system based on single-core optical fiber sensing |
-
2021
- 2021-12-13 CN CN202111521196.5A patent/CN114266271A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115424399A (en) * | 2022-09-02 | 2022-12-02 | 复旦大学 | Distributed optical fiber well lid switch monitoring system and method |
| CN116026449A (en) * | 2023-03-30 | 2023-04-28 | 广东恒志信息技术有限公司 | Vibration positioning monitoring system based on single-core optical fiber sensing |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zou et al. | An adversarial denoising convolutional neural network for fault diagnosis of rotating machinery under noisy environment and limited sample size case | |
| CN110995339B (en) | Method for extracting and identifying time-space information of distributed optical fiber sensing signal | |
| CN114266271A (en) | Distributed optical fiber vibration signal mode classification method and system based on neural network | |
| CN109272017B (en) | Vibration signal mode identification method and system of distributed optical fiber sensor | |
| CN110929785B (en) | Data classification method, device, terminal equipment and readable storage medium | |
| Chen et al. | A real-time distributed deep learning approach for intelligent event recognition in long distance pipeline monitoring with DOFS | |
| CN102254394A (en) | Antitheft monitoring method for poles and towers in power transmission line based on video difference analysis | |
| CN111209853B (en) | Optical fiber sensing vibration signal mode identification method based on AdaBoost-ESN algorithm | |
| CN114926825A (en) | Vehicle driving behavior detection method based on space-time feature fusion | |
| CN112036425A (en) | Tunnel cavity state radar spectrum image recognition model construction method and tunnel cavity state radar spectrum image recognition method | |
| CN111476102A (en) | Safety protection method, central control equipment and computer storage medium | |
| CN112183582A (en) | Multi-feature fusion underwater target identification method | |
| CN115438765A (en) | Moving target identification method, device and equipment based on evolutionary neural network | |
| CN117197726B (en) | Important personnel accurate management and control system and method | |
| CN114510960A (en) | Method for recognizing distributed optical fiber sensor system mode | |
| CN108280950B (en) | Defense area type optical fiber perimeter protection algorithm based on high-frequency energy distribution | |
| CN111597994B (en) | Optical fiber perimeter security intrusion event identification model construction method and security system | |
| CN116838955A (en) | Two-stage oil and gas pipeline line interference identification method | |
| CN112257533A (en) | Perimeter intrusion detection and identification method | |
| Chiang et al. | Distributed acoustic sensor systems for vehicle detection and classification | |
| CN111951505B (en) | Fence vibration intrusion positioning and mode identification method based on distributed optical fiber system | |
| Bublin | Machine learning for distributed acoustic sensors, classic versus image and deep neural networks approach | |
| CN112434716B (en) | Underwater target data amplification method and system based on condition countermeasure neural network | |
| CN112232329A (en) | Multi-core SVM training and alarming method, device and system for intrusion signal recognition | |
| Uzundurukan et al. | Deep Learning Based Threat Classification for Fiber Optic Distributed Acoustic Sensing Using SNR Dependent Data Generation |
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 |