1. Introduction

The distribution of the back-reflected light along the fiber-optic transmission link is an important parameter for identifying the problems in the outside plant. An optical time-domain reflectometer (OTDR) has been widely used to measure this distribution. A modern OTDR can provide sufficient spatial resolution and dynamic range required for the characterization of the transmission link. Thus, there have been many efforts to utilize the OTDR for the in-service monitoring of various types of fiber-optic transmission systems [1–4]. In these techniques, a supervisory channel is used for the OTDR pulses so as not to disturb other wavelength-division multiplexed (WDM) channels in service. However, it is not straightforward to apply these techniques in WDM networks due to the WDM multiplexers and de-multiplexers placed along the transmission link [3, 4]. For example, in a WDM passive optical networks (PON) implemented by using an arrayed-waveguide grating (AWG) at the remote node (RN), these techniques cannot monitor the failures in the drop fibers (which connects the RN and each subscriber) since the OTDR pulse is blocked at the RN. Several techniques have been proposed to solve this problem by implementing additional couplers to bypass the AWG at the RN [5], using a tunable OTDR [6], or generating the OTDR pulse for a specific drop fiber by using the corresponding WDM transmitter [7]. However, it should be noted that all these techniques require the termination of the service of the corresponding WDM channel during the process of monitoring the status of drop fibers.

In this paper, we propose and demonstrate a novel optical reflectometer based on the correlation detection. Unlike the conventional OTDR, the proposed technique does not need any short-pulse sources, but uses the data-modulated transmitter itself. The distribution of the back-reflected light is obtained by calculating the cross-correlation between the transmitted and back-reflected signals. Thus, this technique can utilize the existing optical transmitters without any modifications. In addition, the in-service monitoring is realized not at the supervisory wavelength but at the signal wavelength. In this paper, we describe the operating principle of the proposed technique and discuss its potential limitation on the dynamic range. The operating principle is similar to that of the random-modulation continuous-wave (CW) lidar [8,9]. However, unlike the lidar, in which an ideal pseudo-random code with perfect autocorrelation characteristics is used, we use the truly random data used in service. As a result, the proposed technique can suffer from the severe limitation on the dynamic range due to the imperfect autocorrelation characteristics. To overcome this limitation, we have developed a discrete-component elimination algorithm. For a demonstration, we implement a WDM PON and used the proposed technique for the in-service monitoring of the drop fibers.

2. Operating principle

The proposed technique is based on the correlation detection, which has been extensively used in spread-spectrum communication systems and also applied to random CW lidars [8]. Hence, the operating principle is more or less similar to those systems. In this section, we first explain the basic principle under the ideal condition with infinite integration time, and then discuss the limitation on the dynamic range, which is the inherent problem of the proposed technique due to the imperfect autocorrelation characteristics.

2.1 Basic principle

Figure 1 shows the schematic diagram of the proposed technique. We use a conventional optical transmitter which is driven with a binary signal from the data source at a bit-rate B and launch the signal light into the transmission fiber. Let s(t) be the transmitted binary signal having a value of +1 or -1 and T the bit duration of the signal (=1/B). The input optical power is expressed by P_in(t)=P_a(1+s(t)), where P_a is the average input power. If we denote the distribution of the reflectivity including the round-trip loss as R(z), the power of the light back-reflected to the transmitter can be expressed as P_ret(t) = P_in(t)⊗R(ν_ct/2) , where ⊗ stands for the convolution operation, and ν_c is the group velocity of light in the fiber. Then, we detect the back-reflected light by using a photo detector. The signal voltage after the detection ν_det(t) can be expressed by the sum of dc- and ac-components as follows:

where η is the conversion efficiency of the detector, ν ₀≡ηP_a and ν_det′≡ηP_a s(t)⊗R(ν_ct/2) are the dc- and ac-components of ν_det(t), respectively. Hereafter, we use the prime (′) to express the ac-component.

 

Fig. 1. The schematic diagram of the proposed technique.

Download Full Size | PDF

For correlation detection, we use the transmitted data signal as a reference signal. The voltage of the data signal ν_ref(t) (i.e., reference signal) can be expressed by using s(t) as follows:

where ν _ref0 and ν_ref′(t) are the dc- and ac-components of ν_ref(t), respectively. The cross-correlation function q(τ) between ν_det′(t) and ν_ref′(t) is given by

where 〈〉 stands for ensemble average and ϕ_s(τ) is the autocorrelation function of s(t). (ϕ_s(τ) = 1-|τ|/T for |τ|<T and ϕ_s(τ) = 0 for |τ|≥T). Since T is sufficiently short, ϕ_s(τ) can be well approximated by using a delta function as Tδ(τ) . Then, Eq. (3) becomes q(τ) = η P_a ν _ref0 TR(ν _c τ/2). Thus, the reflectivity distribution can be derived by using the cross-correlation function q(τ) as R(z)∝q(2z/ν_c).

Since the transmission speed is extremely fast (>1 Gb/s) in most optical networks, it would not be practical to implement the proposed technique described above due to the high sampling rate required for measuring ν_det′(t) and ν_ref′(t). In order to cope with this difficulty, we filter out the low-frequency components of ν_ref′(t) and ν_det′(t) by using a first-order low-pass filter with a cut-off frequency of f_c (≪B). Here, let ν _ref,f′(t) and ν _det,f′(t) denote the ac-components of the reference and returned signals after filtering, respectively. Then, the cross-correlation function between ν _ref,f′(t) and ν _det,f′(t) is given by

where ϕ _s,f(τ) is the autocorrelation function of s_f(t), which is the signal derived by low-pass filtering of s(t). Since f_c ≪ B, ϕ _s,f(τ) can be expressed by πf_cTexp(-2πf_c |τ|). Thus, the time width of the autocorrelation function is broadened due to this filtering. The dashed curve in Fig. 2(a) shows an example of ϕ _s,f(τ) calculated for f_c=3 MHz. It has a single peak with a finite time width and quickly converges to zero as |τ| increases. Since the cross-correlation obtained in Eq. (4) represents the convolution of this broadened autocorrelation function ϕ _s,f(τ) and the distribution of the reflectivity, the spatial resolution is limited by the time width of ϕ_s,f(τ). The spatial resolution is given as 0.22 ν_c/f_c when we define the spatial resolution by using the full-width at half maximum (FWHM) of ϕ _s,f(τ). For example, when f_c=3 MHz, the spatial resolution becomes 15.0 m.

 

Fig. 2. (a). The autocorrelation function of the reference signal. The dashed curve shows the autocorrelation function calculated by using the ensemble average with infinite averaging time. The solid curve shows the autocorrelation function with a finite length of the sampled data. (N=4096). The inset shows the same trace plotted in log scale. (b) The background noise suppression ratio (BNSR) calculated as a function of the number of sampling points. Dots and open circles show the BNSR simulated for cases when f_c/f_s=0.1 and 0.3, respectively.

Download Full Size | PDF

2.2 Limitation on the dynamic range and its improvement by the discrete component elimination algorithm

For the practical purpose, we assume that ν _ref,f′(t) and ν _detf′(t) are digitally processed after sampling and the data length is finite. Therefore, the ensemble average in Eq. (4) can be replaced with the average by using the sampled data with a finite length. In order to take this into account, we denote the sampling interval and the total number of sampling points by Δt and N, respectively, and rewrite Eq. (4) as follows,

where k is an integer, and the convolution is carried out in the discrete time domain. In Eq. (5), ϕ _{s, f} (kΔt) is the autocorrelation function of s_f (iΔt) given by

The autocorrelation function ϕ _{s, f} (kΔt) has a sharp peak at kΔt = 0, but it accompanies with the noise background due to the finite sampling length. As an example, the solid curve in Fig. 2(a) shows ϕ _{s, f} (kΔt) simulated for N=4096, Δt =100 nsec (sampling rate f_s=10 MHz), and f_c=3 MHz. In order to evaluate this background noise, we define the background noise suppression ratio (BNSR) by the ratio of the peak amplitude to the standard deviation (i.e., root-mean square) of the background noise of ϕ _s,f. By assuming that s(t) is a stationary random variable, we can derive the BNSR analytically after some algebraic calculations as follows:

The dotted and solid lines in Fig. 2(b) show the BNSR as a function of N for f_c/f_s=0.1 and 0.3, respectively. When f_C is chosen to be larger than 0.3f_s, the BNSR can be well approximated by N ^1/2. For example, when N=4096, the BNSR becomes 18 dB. Dots and open circles in Fig. 2(b) show the BNSR values obtained by the numerical simulation. They agree well with the theoretically calculated lines using Eq. (7).

 

Fig. 3. Limitation on the dynamic range due to the BNSR and its improvement by the discrete component elimination algorithm. (a) Simulation model. (b) The cross-correlation trace. (c) The cross-correlation trace after applying the discrete component elimination algorithm to (b).

Download Full Size | PDF

Due to this background noise, the dynamic range can be severely deteriorated. In order to estimate the influence of the background noise, we simulated the proposed technique by using the model shown in Fig. 3(a), in which two discrete reflection points are assumed at z=5 and 10 km with reflectivities of -20 dB and 0 dB, respectively. In this simulation, we ignored other noise sources such as the receiver noise and used a numerically generated random bit sequence. Figure 3(b) shows the result obtained by simply calculating the cross-correlation function using Eq. (5). This figure shows that the low reflection point at 5 km is hidden by the noise background caused by the higher reflection point hides. As a result, the dynamic range is severely limited.

This problem is due to the fact that we have used the true random signal (i.e. the data to be transmitted) as the reference signal. To avoid this problem, some special bit-sequences such as M-sequence and Golay code are used in the random-modulation CW lidar and the complementary correlation OTDR [8, 9]. These special bit-sequences have a delta-function-like autocorrelation function without background noise, and they are ideal as a reference signal for cross-correlation detection. On the contrary, since the proposed technique utilizes the data to be transmitted, there is no choice of the bit-sequence for the reference signal.

To improve the BNSR, we have developed the discrete component elimination algorithm. Since the reference signal ν_detf′(t) is known, (ϕ _s,f can be easily calculated. That is, we can calculate the background noises generated from the discrete reflection points precisely. Thus, we can cancel out the background noise by the following steps: (a) calculate the cross-correlation using Eq. (5) directly, (b) find the highest discrete reflection point in the cross-correlation trace, (c) calculate ϕ _s,f from the measured reference signal ν _ref,f′(t) and estimate the background noise components, and (d) subtract it from the original cross-correlation function. As an example, we applied this discrete component elimination algorithm to the cross-correlation trace shown in Fig. 3(b). The solid curve in Fig. 3(c) shows the result when we subtracted the background noise generated by the peak at 10 km by the proposed algorithm. The results show that the background noise level is drastically reduced and the small peak component located at 5 km can be clearly observed. The recursive use of this algorithm can substantially improve the BNSR even when there are many reflection points in the transmission line.

To implement the proposed technique, it should be considered that most transmitters and receivers used in the telecommunication network have ac-coupled input/output ports. Thus, they have the high-pass filter (HPF) response with a cut-off frequency in the order of a few hundreds kilohertz. Intuitively, one might think that the effect of the ac-coupling is not serious because the proposed technique only utilizes the ac-component. However, the loss of the low-frequency component can affect the performance of the proposed technique. If we assume that one of the transmitter or the detector is ac-coupled and take account of its HPF response, Eq. (6) is modified as follows:

where ξ (t) is the impulse response of the HPF. Usually, ξ (t) is an S-shape function having a negative tail and it makes the dead-zone after the discrete reflection point depending on the cut-off frequency. Also it becomes impossible to measure the distributed reflection like Rayleigh scattering.

 

Fig. 4. Experimental setup. The inset shows the example of waveforms of the reference and the detected back-reflection signals.

Download Full Size | PDF

3. Experiments

Figure 4 shows the experimental setup used to demonstrate of the proposed technique. We used a distributed-feedback (DFB) laser diode operating at 1550.6 nm and directly modulated it with 1.25-Gb/s non-return-to-zero (NRZ) data in order to simulate the system in service. The length of the pseudo-random bit sequence (PRBS) was 2³¹-1. The signal light was launched into the fiber through an optical circulator. The average input power launched into the fiber was 0.8 dBm. To detect the back-reflected light, we used a conventional ac-coupled avalanche photo diode (APD), which was designed for a 155 Mb/s synchronous digital hierarchy (SDH) receiver. The data and the detected electric signals were filtered out by 3-MHz LPF’s, and digitized by a 9-bit A/D converter at a sampling rate of 10 Ms/s. Then, we calculated their cross-correlation by using a personal computer. In this experiment, the spatial resolution was determined by the bandwidth of the LPF to be 15 m. The maximum number of sampling points of the A/D converter was 4096, and the measurement range was 40.9 km. The inset in Fig. 4 shows an example of the sampled waveform. Since the cut-off frequency of the LPF was much lower than the bit-rate, the waveforms of the reference and detected back-reflected signals looked like noises having continuous probability distribution functions. The period of the 2³¹-1 PRBS was 1.72 sec and it was much longer than the sampling duration (= NΔt) of 4 msec. Therefore, there was no need to take into account the periodicity of the PRBS during the sampling of N-point data, and we can assume the transmission data was random.

 

Fig. 5. (a). Cross-correlation trace when a 2.2-km fiber with open end was measured. (b) Cross-correlation trace when the discrete component elimination algorithm was applied to (a). The red dotted lines show the root-mean square of the background noise, and the BNSR was defined by the ratio of the peak to this background noise level. (c) The BNSR measured as a function of the number of sampling points. Open circles and dots show the BNSR measured with and without the discrete component elimination algorithm. The dashed line shows the theoretical BNSR calculated by using Eq. (7).

Download Full Size | PDF

We first investigated the influence of the background noise originating from the background of the autocorrelation of the reference signal. In order to neglect the influence of the other impairment factors such as the quantization noise and the receiver noise, we used a relatively short length fiber of 2.2 km with an open end as a system under test. Figure 5(a) shows the cross-correlation trace measured when N=4096. The reflection at the fiber end was clearly observed even without averaging, in spite of the background noise. We evaluated the BNSR by calculating the ratio of the peak of the cross-correlation to the root-mean square of the background noise. The closed circles in Fig. 5(c) represent the BNSR measured as a function of the number of sampling points. The solid line indicates the theoretically calculated value of BNSR by using Eq. (7). The square-root dependence on the number of sampling points N is clearly seen. The small discrepancy from the theoretically calculated line was attributed to the noises generated by the other impairment factors.

We then applied the discrete component elimination algorithm and evaluated the effectiveness in the improvement of the BNSR. Figure 5(b) shows the results obtained by applying the discrete component elimination algorithm to the same trace shown in Fig. 5(a). This figure shows that the proposed algorithm could reduce the background noise level substantially. The open circles in Fig. 5(c) show the improved BNSR by using the proposed algorithm. In this demonstration, the BNSR was improved by about 8 dB. The residual background noise was generated mainly by the background noise generated by other impairment factors such as the receiver noise. When the signal-to-noise (S/N) ratio of the reflected light signal ν _det,f′(t) is not kept high (e.g., that the back-scattering level is too low to be measured), the BNSR improvement by the proposed algorithm becomes small. In this case, the BNSR can be simply determined by the S/N ratio of ν _det,f′(t).

 

Fig. 6. In-service monitoring of the WDM PON system. (a) Without averaging. (b) With averaging of 400 traces. The dashed line shows the root mean square of the background noise. (c) Cross-correlation trace when a fiber break with a small reflection of -39.2 dB was intentionally made just before the ONU.

Download Full Size | PDF

The experimental setup shown in Fig. 4(b) was used to demonstrate the feasibility of using the proposed technique for the in-service monitoring in WDM PON. The system under test consisted of two AWG’s, 10.0-km long feeder fiber and 6.2-km long drop fiber. The 1.25-Gb/s receiver was attached at the end of the drop fiber to simulate the optical network unit (ONU). The insertion loss of each AWG was about 4.1 dB and the fibers had a propagation loss of 0.21 dB/km. Figure 6(a) shows the cross-correlation trace measured without averaging. The discrete reflections at the central office (AWG1), the remote node (AWG2), and the ONU could be observed clearly even without averaging. In order to improve the S/N ratio further, we measured the cross-correlation trace 400 times and averaged them. Figure 6(b) shows the result. The background noise level (shown by the dashed line) was below the Fresnel reflection at the ONU end by 27 dB, which corresponded to the return loss of 41 dB (64.8 dB including the round-trip loss). In the current standards on PON, the tolerable reflectivity (or optical return loss) is defined to be around -32 ~ -26 dB [10–12]. These values are much higher than the measurement limit of this setup. Thus, the proposed technique can be used for the detection of the abnormal reflections (i.e., caused by irregular connections or fiber breaks) along the transmission line. We noted that the noise level was still too high to measure the distributed Rayleigh scattering (~-72 dB/m) [9,13]. To demonstrate the possibility of using the proposed technique for localizing the fiber failure, we intentionally made a break point having a reflectivity of -39.8 dB just before the ONU, and measured the cross-correlation. The result in Fig. 6(c) confirmed that the proposed method could indeed localize this break point. Thus, we believe that the proposed technique could be used for the in-service monitoring of WDM PON.

4. Summary

We proposed a novel optical reflectometer and demonstrated the in-service monitoring of the distribution of the back-reflected light in WDM-PON. In the proposed technique, we used the data-modulated transmitter instead of the optical pulse source, and detected the reflection points by calculating the cross-correlation between the transmitted and the back-reflected signals. Since we used the optical transmitter (used in service) itself, the probe light was not blocked even if optical filters/AWG’s were placed along the transmission link. We theoretically showed that the dynamic range of the proposed technique could be limited due to the background noise of the imperfect autocorrelation characteristics of the data signal to be transmitted and that the background noise from the higher reflection point could hide other lower reflection points. To deal with this problem quantitatively, we defined the BNSR and derived the analytical expression as a function of the number of sampling points. Then, we proposed a novel discrete component elimination algorithm, and showed its effectiveness by computer simulation. For a demonstration, we implemented a WDM-PON system whose downstream bit-rate was 1.25 Gb/s and used the proposed technique for the in-service monitoring. We examined the feasibility of the discrete component elimination algorithm, and achieved the improvement of the BNSR by about 8 dB. In addition, we successfully identified the reflections from all the components used in the WDM PON system. The minimum measurable return loss observed from the transmitter end was 64.8 dB. This sensitivity was high enough to detect irregular reflections. Thus, we believe that the proposed technique could be used for the in-service monitoring of WDM PON.