1. Introduction
Fiber-optic sensors have become indispensable ingredients in many applications, too wide and diverse to enumerate [
1,
2,
3,
4,
5]. They can directly measure strain, temperature, electric and magnetic fields, as well as rotation, and indirectly many other measurands in both static and dynamic scenarios. The latter include, for example, the measurement of the temporally varying strain fields in structural health monitoring (SHM) of flying airplanes, traffic carrying bridges and other civil structures, under dynamic loading [
6]. Aiming at the detection of a damage at its embryonic stage, it is of prime importance to obtain an accurate and undistorted dynamic signature of the relevant strain field.
A leading sensor for these applications is the Fiber Bragg grating (FBG) [
7,
8], for which the wavelength (or frequency) of peak reflection,
, is directly related to the local strain and temperature (henceforth, the ‘measurand’). For sensing purposes, FBGs are written along the fiber, either at discrete locations, or continuously inscribed along the fiber during the manufacturing process [
9]. For interrogation, a few techniques have been developed to accurately extract the
of the tested FBG. A common technique, applicable to both the discrete and continuous cases (though of different complexities, [
7,
8,
10]), involves periodic wavelength scanning of a range of wavelengths, which encompasses the values that
may acquire under the anticipated loading conditions. During each scan and provided that the scan rate is fast enough (with respect to the temporal bandwidth of the measurand), the interrogator correctly measures the instantaneous reflection spectra of the various concatenated FBGs. These peaks which determine the sought-after
occur, however, at instants dictated by the actual values of the measurand under study, and are generally not time-synchronized with beginning/end of the scan. For example, let the periodic scan cover a wavelength range which includes the
of interest, starting at
and ending at
. Clearly, a low strain value will have its corresponding
occur near
, i.e., near the beginning of the scan, while a large strain value will have its
measured only towards the end of the scan. Thus, despite the periodic nature of the scan, the resulting
are actually obtained on a
non-uniform temporal grid. Nevertheless, commonly available frequency-scanning interrogators report their obtained measurand values (the
in the case of FBG interrogation) on a
uniform time grid, either at the beginning of the scan or at its end.
In this paper, we show for the first time to our knowledge, that reporting the obtained measurand values on a uniform time grid, while they were actually obtained on a non-uniform one, results in both harmonic distortion and time-domain errors. These errors grow in significance: (i) as the dynamic range of the measurand (in terms of its induced wavelength/frequency variations) fills the scanning span; and (ii) as the temporal bandwidth of the measurand approaches the Nyquist bound (i.e., half the scanning frequency). This type of harmonic distortion and time-domain errors may be of particular concern for dynamic frequency-scanning fiber sensing techniques, characterized by a tight trade-off between the scanning rate and scanning span. Following an exposition of the problem via simulations in
Section 2, a simple post-processing technique appears in
Section 3, where knowledge of the instants of sampling (when available) leads to a significant decrease in both the spurious harmonic levels and time-domain errors (since information on the exact time evolution of the scan is generally not made available, the instants of sampling can be estimated from the reported data).
Section 4 presents an experimental corroboration of the simulation’s main predictions, utilizing the frequency-scanning technique of Fast Brillouin Optical Time Domain Analysis (F-BOTDA, [
11]). Here, a fully controllable laboratory setup measures the strain of an optical fiber under longitudinal and quite pure sinusoidal vibrations (independently measured by a temporally uniform sampling interrogator). Based on a known digitally generated frequency-scanning waveform, the setup augments the reported sampled strain values with the instants at which they were obtained for different values of scan rate and span. Attributing the sampled values to a uniform temporal grid allows us to compare the measured harmonic distortion with the simulation results of
Section 2. Moreover, feeding the measured strain values together with their instants of acquisition to the post-processing technique of
Section 3, results in a significantly purer signal recovery.
Section 5 investigates the performance of a commercial frequency-scanning dynamic interrogator of ‘continuous’ draw-tower FBGs [
9,
10], which reports the measured strain values only on a temporally uniform grid. This interrogator trades-off scan rate and scan range: the faster the scan, the narrower the span. As expected, the experimental results show that for a fast scan rate and its associated constrained span, the harmonic distortion worsens with increasing signal frequency and/or strain amplitude. Note that when the instants of sampling are not reported, they cannot be deduced from the measured values without a full knowledge of the scan parameters, which are most often not available.
Section 6 discusses the findings and provides valuable conclusions.
2. Simulations: Strain Interrogation with Periodic Frequency Scanning
The reflection from an FBG [
7,
8], as well as the probe gain in Brillouin-based sensing [
11,
12,
13], are frequency dependent, having a characteristic spectral shape, commonly denoted here by
. In both cases, and for a given time
peaks at a frequency, which is a unique function of the value of the measurand at that instant. Let
, of sinusoidal shape, amplitude
, and frequency
,
Figure 1a (solid blue), represent that peak frequency, which is the local Brillouin frequency shift (BFS) in Brillouin-based sensing (see
Section 4), and
for an FBG (as in
Section 5).
The commonly used saw-tooth-type frequency-scanning signal,
in
Figure 1a (solid green curve), is assumed to be ideal, having a perfect linear ramp and abrupt return, a scan rate of
(
) and peak-to-peak scanning span of
(for optimum coverage, the signal mean value is placed at the center of the frequency scan,
). Both the normalized frequency,
, and filling factor,
, can range from near zero (for dense sampling and very small signal amplitude) to below 0.5 (the Nyquist limit, and a meaurand peak-to-peak dynamic range that fills the scan span). The simulation assumes that the signal mean value perfectly aligns with that of the scan,
Figure 1. This is optimal for best utilization of the scan span. Anyway, the simulation results are independent of the signal mean, as long as the full dynamic range of the signal lies within that scan span.
As the optical frequency of the interrogator’s source is periodically scanned, either continuously or in small frequency steps, a photodetector measures the frequency-dependent reflected power from the FBG [
7,
8], or the frequency-dependent power of the Brillouin-amplified probe [
11,
12,
13]. In the simulation,
is assumed to have a Lorentzian shape,
, characterized by maximum value at
, and a (scaled) full width half maximum of
. The time-dependent detector power,
. can then be expressed by:
The simulated detector power in the neighborhood of its local peak at each of the scan periods is represented by the red-dot curves in
Figure 1b (in arbitrary units). The vertical purple lines in
Figure 1b cut the time axis at those instants,
, at which the detector power vs. time records attain their maxima at each scan cycle (see the
, fifth period in
Figure 1b), where
is the total number of scan cycles. Ideally, in the absence of noise and other perturbations, these maxima occur at the intersections of the signal and the saw-tooth scanning curves (purple X’s in
Figure 1b). Mathematically,
are the solutions of:
Once the intersection points are found, their ordinates are the desired measurand values (see the 5th period):
Note that the actual shape of the detector record is not fixed but rather depends on the signal dynamics, and specifically on the slopes of the signal and scan waveform near the instant of intersection. The more parallel the signal and waveform are, the wider the recorded shape and vice versa (cf. the red-dot curves in the second and fifth periods).
In the simulated scenario of
Figure 1, sampling rate is more than four times the signal frequency, seemingly more than sufficient for accurate recovery of the signal from its samples.
However, since the signal varies in time from scan period to another, it is clear from the ramp-type nature of the scanning, as well as from the nonlinear characteristics of Equation (2) and graphically, from
Figure 1b, that the sampling instants have variable signal-dependent distances from the beginning of the scan cycles that encompass them. For example: while in the first period of
Figure 1 sampling occurs towards the end of the scan, it is just the opposite at the fourth period. Yet, as already mentioned in the introduction, it appears, that frequency-scanning interrogators report their acquired measurand values at uniformly spaced instants, associated with either the beginning,
or end of the relevant scan period,
. This is bound to result in erroneous reconstructions of the sampled signal. Indeed, high-granularity sinc-reconstruction based on the sampled measurand values,
(the ordinate values of the purple X’s)), when attributed to the beginnings of the scans (black full circles in
Figure 1b), produces the red curve of
Figure 2, which is not only (tolerably) time-shifted, but also a distorted reconstruction of the true signal (blue curve). Note that sinc-based reconstruction of the sine wave from its true signal values at the scans’ starting points,
, rather than from the quite different
, results in a curve indistinguishable from the blue curve of the figure.
Assigning the sampled measurand values to a uniform temporal grid makes it possible to spectrally analyze them using DFT/FFT (which implicitly attributes these values to a uniform time-grid).
Figure 3 shows the power spectrum of the signal of
Figure 1, based on its obtained samples values,
. Here, the sampling rate is more than twice the Nyquist rate,
, and the signal variations are well within the scan limits,
. Yet, instead of a single peak at the signal frequency,
, the spectrum shows many harmonics, where the one at
is a very strong second harmonic, only −13.73 dB below the signal. All other peaks are folded harmonics (e.g., the peak at
is the third harmonic at −23.96 dBc, while the one at
is the fifth harmonic at −40.69 dBc, etc.). The total harmonic distortion lies at −13.27 dBc. Clearly, the second harmonic is the dominant one. Power leakage to all these harmonics also lowers the peak at the signal frequency by 0.27 dB.
Our simulation assumed an ideal saw-tooth waveform, as well as highly precise determination of the measurand sampled values
and their instants of acquisition,
. In practice, however, these values are only estimated from the reflection/gain measurements (red-dot curves in
Figure 1) with accuracy affected by (i) the estimation algorithm; (ii) the scanning granularity (the frequency steps); (iii) scan calibration; and (iv) noise. Simulations that include these sources of inaccuracies still exhibit the same strong harmonics in the calculated spectra, quite similar to those of
Figure 3. The only exception is the presence of an elevated floor, which masks some of the weaker harmonics (Note that our analysis does not consider the effect of fiber length, since the round-trip time of light in the fiber is assumed to be negligibly small with respect to the scan period).
Figure 4 displays simulation results for the dependence of the second harmonic (scaled by the signal power) on
, and
for a sinusoidally varying measurand. Clearly, the larger the measurand amplitude the higher the required scanning frequency for a prescribed amount of harmonic distortion. Conversely, the closer the scanning frequency to the Nyquist rate, the smaller the allowed dynamic range to maintain a maximum permissible level of the second harmonic. Finally, harmonic distortion is a sign of nonlinear behavior of the interrogation process, indicating, for example, that the output of the sum of two inputs will not equal the sum of the individual outputs. Indeed, when the sum of two signals is sampled, the actual sampling instances of the sum are generally different from the actual sampling instances of each of these signals alone; therefore, the sampled values of the sum are generally different from the sum of sampled values of each of the signals. Attributing all samples to the same time grid therefore gives rise to an apparent non-linearity of the sampling mechanism, whereas if all samples are attributed to their correct time instances, linearity is implicitly maintained.
3. Distortion Mitigation Based on the Availability of the Sampling Instants
Let us assume that the interrogator reports the sampled values of the measurand,
, together their non-uniformly spaced instants of acquisition,
, Equations (2) and (3). While accurate signal recovery from non-uniformly time-spaced samples is not generally possible [
14,
15], yet, a sharp increase in the fidelity of the reconstructed signal can be achieved using interpolation.
Looking back at
Figure 1a, we note that since the mean value of our example signal coincides with the mean of the scan, the sampling instants,
, also gather around the (temporal) middle of the scan:
. More generally, let us define a shifted uniform temporal grid by:
Interpolation algorithms can now be applied to the measured data pairs,
, in order to estimate the signal values on the uniform grid of Equation (4):
. Thus, spectral analysis and time-domain recovery of the measured signals will now be based on
rather than on
. The results of processing the full data behind
Figure 1 in this manner, using Spline interpolation, are shown in
Figure 5 (time domain, where, again, sinc-based reconstruction was used to obtain the much denser displayed granularity), and in
Figure 6 in the frequency domain. The benefits of having access to the sampling instants are obvious: the folded third harmonic
is now the dominant one but at a level 37 dB below that of the signal. As for the time traces, the standard deviation of the difference between the true signal values at
of Equation (4) and their spline interpolated values,
, normalized by the standard deviation of the signal, is a mere 0.02.
6. Discussion and Conclusions
In dynamic measurand scenario, non-uniform sampling is unavoidable and inherent in periodic frequency-scanning interrogators. Static signals are uniformly sampled, since the scanning waveform always meets the signal at a constant time difference from the scan start. However, as the signal frequency and span increase (with respect to the sampling frequency and scan span, respectively), so does the non-uniformity in the sampling instants. For example, in the BOTDA experiment, the normalized spread in the sampling instants,
increases from 0.01 in
Figure 11 (left) to 0.03 in
Figure 10 (left), in line with the increase in the second harmonic level from −26 dB to −16.3 dB. Many if not all commercial frequency scanning fiber-optic interrogators report the sampled values of the measurand on a uniform temporal grid, whereas they were actually obtained on a non-uniform one. The sampling instants are not reported, nor is information about the exact time evolution of the scan, from which the instants of sampling could be possible estimated. We have shown by simulation and experiments that this common approach leads to distorted reconstruction in both the frequency and time domains. Exposing the erroneous nature of this approach is considered by us to be the main contribution of this paper. Note also that the reconstructed signal is also temporally
shifted from its true time dependence,
Figure 5, potentially leading to synchronization issues when the same physical effect is simultaneously measured by different types of sensors.
It should be noted, though, that the widely used frequency-scanning interrogators of discrete FBGs of non-overlapping reflection spectra, are normally much less affected by this type of errors. In spite of the fact that it also involves non-uniform sampling, the common scanning span is of the order of 40–100 nanometers (nm), while the dynamic range of the tested strain/temperature rarely exceeds 10 nm (~8000 micro strain/1000 °C at 1550 nm). Hence, the filling factor
is usually smaller than 0.1. Yet,
Figure 4 indicates that even for such a small value of
, the scan rate must be carefully chosen to meet a prescribed low level of harmonic distortion.
Nowadays, fiber-optic distributed sensing interrogators have become commercially available, based on either Rayleigh backscattering from standard single-mode optical fibers, or higher reflections from draw-tower ‘continuous’ FBGs. In some implementations, the faster the scan, the smaller the span. Thus, users quite often work with distortion-prone, very high filling factors,
.
Section 5 has examined a commercial interrogator of ‘continuous’ FBGs, exhibiting harmonic distortion that grows with
and
. In principle, the analysis of this paper also applies to Rayleigh-based, frequency-scanning interrogators [
21]. Practically, however, the magnitude of errors critically depends on the values of
and
in the relevant application.
Another worthy contribution of this work is a recommendation to the manufacturers of frequency-scanning interrogators to report the sampled values,
, together with their sampling instants,
.
Section 4 demonstrated that obtaining that type of data from the Brillouin setup allows the use of interpolation methods to obtain a much more accurate estimation of the signal values on a uniform temporal grid. Indeed, the right panes in
Figure 10 and
Figure 11 exhibit a substantially lower level of harmonics in comparison with their left counterparts.
In conclusion, the reporting of non-uniformly obtained measured sampled values on a uniform temporal grid results in erroneous harmonics in the frequency domain, and in distortion in the time domain. A proposed and demonstrated mitigation approach, based on the availability of the sampling instants in conjunction with post-processing algorithms, provides a sharp increase in the fidelity of the reconstructed signal.