This article provides an overview of existing intelligent systems for monitoring the integrity of extended objects based on distributed fiber-optic sensors. The results of the development of measuring systems using distributed fiber-optic sensors are presented. The analysis of existing solutions for constructing measuring systems is carried out. The main types of measuring schemes and principles of determining the integrity of extended objects are considered. The basic principles for the construction of fiber-optic sensors are defined. The analysis demonstrated achievements in improving measurement accuracy using various optical reflectometry methods. Shortcomings of measuring systems are revealed, and ways of elimination are established. Methods of filtration from overvoltages and temperature influences are considered. The difficulties of applying wellknown works are determined. A review of the current state of development of artificial intelligence in the field of its application in measurement systems is carried out. It is revealed that the systems built based on an optical time domain reflectometer (OTDR) using a convolutional neural network (CNN) show higher quality indicators. The possibilities of using different types of neural networks to recognize various mechanical influences using machine learning are considered. The disadvantages and advantages of using neural network systems in measuring systems based on distributed fiber-optic sensors are identified, and the most optimal type is selected. The direction for further research and development of a technical condition monitoring system based on distributed fiber-optic sensors has been determined.
Fiber-optic sensor, optical fiber, monitoring system, neural network
Currently, extended facilities are used to perform any technological tasks, a violation of the technical
condition which can lead to small-scale or global accidents in the worst case. For this reason, intelligent
continuous monitoring systems are needed to warn of integrity violations promptly. It is worth noting
that the leading causes of damage can be both natural, and man-made phenomena, and sabotage,
violations caused by the human factor. Special systems based on an optical time domain reflectometer
(OTDR) are used to monitor the technical condition and ensure the safety of extended facilities. The
systems consume a small amount of electricity, and electromagnetic interference is not induced in them.
Sensors with phase-sensitive reflectometry are the most widely used, due to their sensitivity to
mechanical vibrations. At the same time, increased sensitivity is the cause of false positives, which
leads to an excess of information. Standard solutions do not allow for obtaining reliable results;
therefore, it is necessary to improve the systems. Most often, such a solution is to amplify the optical
signal [
2022 Copyright for this paper by its authors
deformations or vibrations by circulating the optical signal of the laser and the signal from the fiber
under the test (FUT) [
Artificial intelligence systems are used to solve the problem of interference caused by humans and
technology, namely, based on machine learning and neural networks. The result obtained is
characterized by high accuracy in determining the source and nature of the alarming event [
Consequently, quite often, there are systems using distributed fiber-optic sensors built on
phasesensitive OTDR and CNN [
Bibliometric analysis of the Elsevier group database (Scopus) was used as the main method when writing the article. 5,035 articles were found on this topic. The selection criteria for the study were: a) articles published from 2012 to 2022 were considered, b) systems for monitoring the state of extended objects used optical fiber as a sensor, c) systems for monitoring the state of extended objects had an intellectual component. The figure shows the statistics of the Scopus database on publications on this topic, broken down by year. Further, the research area was narrowed, and 20 sources were selected for analysis.
When building the system, they are mainly used by the ϕ-OTDR, c-OTDR method. Moreover, each
system performs certain functions, but the main direction developed is monitoring the state of extended
objects. For ϕ-OTDR systems, interferometry systems are the most common [
A pair of unbalanced Mach-Zender interferometers (MCI) with a time delay, created by a fiber, having a length 500m, was used as the measuring instrument of the HSL to measure the phase of the signal determined by time bands. The signal is represented as intensity and is measured by a photodetector (Figure 2). To eliminate interference, acoustic-optical modulation (AOM), which increases the optical signal amplitude, and an erbium-doped optical amplifier (EDFA), which restores the optical signal level, were installed in the standard circuit.
A laser with a small bandwidth frequency range acted as a source of information in the ϕ-OTDR system to achieve high efficiency of the adaptive pulse period (API) method, and low-frequency vibrations were measured.
, where is the laser frequency shift, is the pseudo–period of the signal intensity, and are the amount depending on the speed of light in space, and the refractive index.
A laser with a bandwidth of 100 Hz and a wavelength of 1550 nm was chosen as the source of
information. The laser light passes through an acoustic-optical modulator (AOM), and a fiber-optic
amplifier (EDFA), and the already amplified signal passes through the fiber operating under the test
(FUT). At the same time, through an optical connector (OC) with a ratio of 1 to 99, the optical signal
from the laser enters the data acquisition board. The piezoelectric transducer generates mechanical
vibrations using an electrical signal of a certain frequency. The reflected optical signal enters the optical
connector (OC) with a 50-to-50 ratio and enters the photodetectors, the output signals recorded by the
data acquisition board [
The design of the system proposed in [
The experimental model was tested using a special vibration exciter, a feed stream with a frequency of 150 Hz. The phase value also changed with increasing current and amplification of the amplitude of mechanical vibrations (Figure 4). At the same time, the polarization-phase fusion unfolding algorithm differs from standard methods in its ability to accurately measure dynamic deformations based on restoring the actual phase signal from the raw with an amplitude signal. Such changes made it possible to achieve an extensive range of amplitude measurements, determined by the polarization signal, and good sensitivity, determined by the phase signal. However, the use of several Bragg gratings complicates the design and reduces the reliability of the whole system. Although it becomes quite easy to detect the damage's specific location, the system can also not recognize interference caused by humans or machinery.
In [
In [
= 2 sin( ), where – amplitude of the phase modulation depth.
A fiber-optic sensor with single-point sensitivity is not enough to determine the location of the vibration source. The solution to this problem is to remove the delay line to reduce sensitivity. The sensor cable with flat sensitivity cannot detect the vibration position, so the delay line must be removed to reduce its sensitivity. At the same time, two sensors were used to determine the distance. The sensors were installed in such a way that the first sensor next to the terminal box had the least sensitivity to vibration, and the second sensor located next to the connector, on the contrary, had the maximum sensitivity to vibration. To determine the location is calculated by the ratio of the difference between the output signals to their sum.
The proposed system can process signals from interference, which is good for zone security systems
and unauthorized access. However, the system cannot distinguish between other interferences.
Vibrations from interference are the causes of a false alarm, due to which standard data processing
systems will not be able to work correctly [
[
The measurement of the system is not affected by the interference caused by mechanical overvoltages and temperature influences. At the same time, this solution allows you to determine the displacement of plates in the tunnel. The results obtained allow us to determine the technical condition of the object.
The solution described in [
By comparing the input and output signals, it is possible to determine the source of vibration and its location. The proposed system involves the use of deep learning, since in its absence the system cannot give an accurate result, namely, it cannot distinguish interference caused by a person or the movement of a machine from an alarm signal. The solution for the monitoring system requires a lot of space and complicates the structure of the system and its price. Having many devices to improve accuracy can also reduce the normal operation time of the device.
The system in [
As a result, the system can detect mechanical deformations of power cables under water. At the same time, the amplification of the optical signal makes it possible to eliminate interference caused by fluctuations in the information source. At the same time, such a system cannot be adapted to other extended objects since it does not consider temperature influences and mechanical overvoltages.
The possibility of using a Rayleigh reflectometer or C-OTDR is also considered [
For each pulse, the value of the backscattering intensity is selected at a certain point in time with a constant interval during the passage of the laser pulse in the optical fiber. At the same time, to determine the intensity of the optical signal at a certain time, it is calculated as the sum of incoherent Iinc and coherent Icoh components. At the same time, the amount of information for processing is reduced, but the information at a certain point in time is sufficient to determine the location of the vibration source.
The proposed method can work accurately to determine fluctuations in extended blocks. Unlike
conventional ϕ-OTDR systems, the information coming from the photodetector is less, allowing for a
short period to obtain accurate data. The main disadvantage is the complexity of the system and the
calculation of fluctuations. Thus, in the presence of many accidental external influences, the system's
operation may be incorrect. The study presented in [
In both cases, a narrow-band laser with distributed feedback (DFB-FL) with a power of 10 MW and a bandwidth of 5 kHz was used. The laser creates an optical signal, amplified using an acoustic-optical modulator (AOM), an optical amplifier doped with erbium (A). The amplified signal is filtered by a fiber-optic lattice filter (F) and enters the circulator (C), which is again converted by devices F and A. The installation for the ϕ-OTDR system uses a standard optical connector, where the signal is divided and fed to three photodetectors (PD1-PD3). In the case of OTDR interferometry, an additional connector is used, from which signals are sent to two rotating Faraday mirrors.
The obtained results showed that in the case of standard ϕ-OTDR systems, the polarization was insignificant, and for the OTDR interferometric system, the polarization was performed independently of the input and output optical signal and was reduced. The use of both methods has demonstrated that phase-sensitive reflectometry is less sensitive than interferometric. However, the use of rotating Faraday mirrors in the system increases the design and complicates the system, and may affect the reliable operation of the system
Systems for monitoring the technical condition of facilities with a length of up to 75 km were
presented in [
The results obtained during the study [
At the same time, well-known methods using interferometers are widely used since they are able to work accurately, but traditional methods do not imply a solution to eliminate interference. The proposed articles above offer solutions to eliminate interference caused by undesirable mechanical, and temperature influences. At the same time, these systems are difficult to apply in a more complex environment, where undesirable mechanical influences are not just natural, but also man-made. The main disadvantages of the use of interferometers are their complexity, size, and cost, as well as the small resolution of the receivers. In further studies, more optimal options are proposed for the implementation of a cable protection system relative to the measuring device.
The device designed by the authors does not use known solutions for the construction of fiber-optic
sensors: optical interferometry, reflectometry, fiber Bragg gratings, or long-period fiber gratings. A
quartz single-mode sensor of the G standard was used as a sensitive sensor.652. A light spot profile is
used to identify the impact on the sensor. The advantages of the developed system are the cost and the
possibility of practical implementation in mines, where the requirements for the safe operation of
systems and devices are increased.
b
Figure 10: Experimental scheme for ϕ-OTDR (a) and OTDR interferometry (b) [
In [
Machine learning in extended object monitoring systems with the fiber-optic sensor system is used to eliminate interference and noise and improve operation parameters.
In [
Further, energy distribution coefficients are used to determine the event that occurred in the system. These coefficients are obtained by decomposing the signals. When the disturbance increases, certain coefficients increase, and an event can be identified by determining the growth. To separate the cause of an alarming event, namely human intrusion, from environmental interference, a 3-layer neural BP ANN is constructed. The BP ANN architecture is shown in Figure 14.
The experiments showed the following results: Identification Rate (IR) - 89.19%, Probability of
Detection (PD) - 86.15%, and Nuisance Alarm Rate (NAR) - 1.75%. BP ANN accuracy rates are
inferior to solutions using other neural networks. The φ-OTDR zoonding systems use machine
learning based on a feature extractor that uses coincidence filtering (MF)to eliminate interference and
determine the nature of the impact in [
The process of obtaining information about an alarming event is done by decomposing the optical signal. To extract the features, the Level Crossing (LC), Short-Time Fast Fourier Transform (ST-FFT), and Discrete Wavelet Transform (DWT) tools were used. The random forest (RF) algorithm was used to classify perturbation regions. According to the results, the measurement error decreases with the number of trends (Figure 15). The disadvantage of this method is that noise is reduced only point-bypoint at the site of disturbances to reduce processing costs. In contrast, there are methods to reduce interference and noise for the entire sensing system.
The prediction is formed by the differential signal δ for each tth time data-vector:
( ) = ( ) − +1( ), ( ) = 2 ∑ −=11 ∑ = [sin( − ) − ( − − )],
Where and are the phasor angles of Rayleigh backscattering signal (RBS) time data vectors from the two regions of both location sides along with perturbation, , are the amplitudes of the according phases. The differential signal was acquired for the different numbers Y and Z of the samples for each region. The phase is the direct measurement of the particle displacement in the optic fiber and consists of both primary and unwanted phases:
= F( + ).
Here F is the mapping function of the ( ) representing the angle through combination of the and signals. Therefore, the correlation vector with consideration of the Spearman correlation coefficient is [ ] = 1 − 6 ∑ =1( ( )− ( +1))
2 ( 2−1) , (7) where varies from 1 to − 1.
In [
PD signals were extracted by decomposition and reconstruction. Next, one-dimensional data on PDC signals, which were collected by the sensing system, were transformed into two-dimensional maps of time-frequency characteristics. Next, images of MFCC objects are sent to the CNN classification model for recognition. The training time of the CNN model is reduced when using the characteristics of PDC signals in the time frequency domain. Experimental results showed accuracy of 96.3%, sensitivity - 96.4% and specificity - 98.7% achieved.
The CNN model used a dataset with 10x cross-validation. The training set is divided into 10 parts, 832 test samples were used. Figure 16 shows the architecture of the CNN model. The architecture of the CNN model consists of two convolutional layers. Both are followed by the top union layer. Convolutional layers apply various filters to the input image to extract characteristic features, and combining layers reduce the size of the output data of the convolutional layer. The efficiency of the CNN model was evaluated by the average value of the obtained accuracy, sensitivity, and specificity, the results obtained with the other six existing traditional methods. The CNN model has the best accuracy indicators.
In [
Due to continuous learning with the help of a training set, the neural network can match the input image with the vibration distribution. The process of practical application also coincides with the testing process. Pixel accuracy reaches 99.95%. DL-TSD also has adaptability and noise protection Figure 17 shows the process of testing and training DL-TSD phase variation generated by the vibration
The neural network was built on the proportion of the light intensity collected at time over the intensity of probe pulse 0, scattering coefficient , the phase of the -th scattering center and ∆ ∝ 02 ∑ = + 2 02 ∑ > ∑ = cos( − + ∆ ).
(8) (9)
is (10) (11) (12) ( , , ℎ) =
(∑ ℎ( − ∗ ℎ)( , )) = ( − ( , ))
Here is the output of the -th convolution operation, is the -th convolution kernel, the activation function, , , ℎ is the row, column, and channel of . The prediction on the two-dimensional temporal-spatial matrix converted into gray-scale was made . The mathematical equation for the
was acquired from convolution operations: ( , ) = ( ○ −1 ○ … ○ 1 ( ( , ))) = was obtained from the mean square error:
was used as labels. The loss function dimensional temporal-spatial matrix = [ 1, 2, … , ] . convolutional operation was applied:
The information the values collected through time for neural network was represented in the twoThe encoding of the through time and mapping to the high dimensional feature space the (1 + (− ( )( ( , )))
) arg
In [
In [
CNN models are a popular tool in constructing intelligent systems for monitoring the technical
condition of extended objects with and using fiber. So, in [
To work with the CNN +SVM model, 11,997 images were used for eight categories of events. Experimental results showed a neural network accuracy of 94.17%. The work contains the largest number of images used for training a neural network among all other works considered.
The article [
To improve classification accuracy in systems using fiber optic based on C -OTDR technology
under challenging environmental conditions and interference, [
According to the analyzed literature, the main problem in technical condition monitoring
systems is interference caused by mechanical overvoltages and temperature influences. So, in [
For very extended objects, as in [
Distributed fiber-optic sensors based on the φ-OTDR principle have become widely used in
monitoring systems for the technical condition of extended objects with a length of up to 1 km. A
different approach is used for objects with a distance of up to 75 km [
However, such devices have the following disadvantages: instability to interference, large size, and
high cost. At the same time, it is worth noting that distributed sensors will be sensitive to minor
vibrations as the length of the sensing element increases [