DOPPLER COMPENSATION FOR JANUS APPLIED TO DATA COLLECTED IN THE BALTIC SEA

Transcription

1 DOPPLER COMPENSATION FOR JANUS APPLIED TO DATA COLLECTED IN THE BALTIC SEA Giovanni Zappa a, Ivor Nissen b and John Potter a a NATO Undersea Research Centre, Viale San Bartolomeo 400, La Spezia, Italy. b Research Department for Underwater Acoustics and Marine Geophysics (FWG), WTD71, Kiel, Germany. Giovanni Zappa, NATO Undersea Research Centre, Viale San Bartolomeo 400, La Spezia, Italia, fax: , zappa@nurc.nato.int Abstract: The Research Department for Underwater Acoustics and Marine Geophysics (FWG), conducted transmission experiments in the Baltic Sea during March 2010 in which JANUS messages were transmitted from RV Planet in the frequency band khz over ranges up to 29 km while sailing at various speeds up to 4 m/s. JANUS packets (64 bits) are expected to be robust to Doppler up to ~2 m/s source-receiver motion along the line of sight. Slightly greater tolerance may be experienced in the case of strong multipath, where the mistiming of the Frequency Hopping - Binary Frequency Shift Keying (FH-BFSK) slots and the Doppler shift of chip frequencies may be partly compensated by multipath temporal and Doppler spread. Using a matched filter on the Hyperbolic Frequency Modulated (HFM) sweep to detect packet arrival, the three wake-up tones were captured and Fourier- Transformed to estimate the Doppler-induced frequency shift and hence determine the best Doppler-compensated sampling frequency to use for decoding the packet. This simple Doppler estimator performs well, compared to ground truth from the RV Planet GPS navigational system, increasing the number of successfully decoded packets from 82 to 207. As anticipated, 90% of the additional decoded signals had Doppler speeds >2.2 m/s. Applying other simple enhancements, including soft Viterbi decoding with increased traceback length significantly improved the decoding rate, increasing the number of correctly decoded transmissions to 395. A Greater Of - Constant False Alarm Rate (GO-CFAR) detector was also applied to reduce false detection triggers, without significant impact. Using the best combination of pre-processors, the majority of transmissions could still not be decoded, as the Signal to Noise Ratio (SNR) fell <9 db at most ranges, with significant multipath. Removing some of the multipath effects by means of a prototype channel equalisation allowed transmissions of slightly lower SNR to be correctly decoded. Widening the acceptance window in frequency and time also appears to help significantly, suggesting that future JANUS versions might usefully employ a wider set/unset bit frequency differential. Keywords: Underwater, Acoustic, Communication, JANUS, FWG, NURC, FSK, CFAR, PLL, Doppler, Baltic, Planet, Frequency Hopping, Binary Shift Keying, multipath, channel, equalization.

2 The Pre-RACUN experiment The purpose of the Pre-RACUN sea trial was 1. To collect data for acoustic channel characterisation for mobile underwater communication 2. To validate the stochastic ray-tracing acoustic propagation model MOCASSIN [1] and 3. To find limit bounds for different robust communication signals such as JANUS [2] in littoral waters, over a long range, using two acoustic bands [ khz and khz]. The JANUS protocol is an open-source robust scheme for standardising digital underwater communications [3]. JANUS has been developed at the NATO Undersea Research Centre (NURC) with the collaboration of academia, industry and government and is in the process of becoming a NATO standard (STANAG). The physical coding scheme is described in [2] and the spectrogram of an example signal is illustrated in Fig. 1. Fig. 1:Spectrogram of a JANUS packet in the 3-5 khz band The experiment was performed in the Baltic Sea, east of Bornholm Island (Fig. 2) in the wind shade of the island. JANUS transmissions were made from RV Planet (Fig. 2, nominal track shown in white) towing two portable omni-directional transducers (ITC4008 and ITC4009). RV Planet sailed at various speeds throughout the trial. JANUS signals were recorded by a moored vertical hydrophone chain consisting of seven hydrophones, though hydrophone #3 was not working (Fig. 2, shown as a block dot near the western end of RV Planet s track). The acoustic channel was characterised by a stable stratification resulting in an iso-velocity surface layer overlying a sound channel where the sound velocity was significantly lower. FWG s stochastic ray-tracing shallow-water acoustic propagation model, MOCASSIN [1], was used to predict Transmission Loss (TL) as a function of depth and range out to 60 km for the two frequency bands used for JANUS transmissions. Example TL results are shown in Fig. 3 for a source placed at 50m depth.

3 Figure 2: Map of the experimental area. The waveguide ducting of acoustic energy in the 40-50m depth range is clearly visible, especially in the higher frequency band. The generally lower TL experienced by lower frequencies is also evident. Figure 3: Acoustic Transmission Loss (TL) modelled by MOCASSIN in the low (left hand panel) and high (right hand panel) frequency bands used for 1. DOPPLER EFFECTS ON JANUS SIGNALS We begin by noting that the received frequency of each single-path arrival of a signal is related to the transmitted frequency and line-of-site speeds by (1)

4 Where is the received frequency, is the transmitted frequency, is the ratio of the length of the transmitted signal to the length of that received, and is the speed of sound in water (~1500 m/s) with the subscripts r and s denoting receiver and source speeds along the line of sight. The effect is to apply a dilation factor,, to the transmit frequency, the same factor also affecting the temporal duration of the signal, so that the number of cycles remains constant. The scenario is complicated by multipath, when the received waveform may be modelled as a sum of time-delayed amplitude-modulated and Doppler-dilated replicas of the transmitted signal, where depth-dependent currents and surface waves confer different Doppler dilations for each path. The effect of Doppler on JANUS signals is therefore to both shift the frequencies of the various components, and to dilate the timing, affecting the synchronisation necessary to decode the signal. 2. A SIMPLE DOPPLER COMPENSATION PRE-PROCESSOR FOR JANUS Although the JANUS BH-FSK scheme is inherently robust to moderate Doppler dilations [4], in this experiment it was known that the RV Planet was travelling at speeds above the expected Doppler tolerance for JANUS packets (~+/-2 m/s) about 75% of the time. While the JANUS standard defines how a JANUS packet is to be encoded, there is no restriction on how to decode a received signal. We may therefore propose a Doppler compensation processor to improve performance. We choose to do so by pre-processing the received signal by re-sampling at a shifted (dilated) frequency. The advantage of this approach is that it is simple and does not affect the rest of the decoding process chain, based on the example implementations provided on the JANUS wiki site ( Also, re-sampling compensates for both the time dilation and frequency effects of Doppler. Our algorithm proceeds as follows: 1. Detect the arrival of a JANUS packet with a matched filter to locate the HFM sweep that is part of ver. 1.0 packets 2. Sum the waveforms of the three wake-up tones that precede the HFM and evaluate the Power Spectral Density (PSD) of the summed wake-up tone signals 3. Fit quadratic curves to the three main spectral peaks to estimate the received tone frequency peaks 4. Estimate the Doppler dilation from the shift of the fitted curves compared to the nominal transmit frequencies. Detection on the HFM is robust because, as is well known, the impact of a Doppler shift on an HFM signal is principally to appear as an arrival time delay, with some minor reduction in matched filter amplitude due to the end effects at the start and finish of the sweep. Doppler will also cause a temporal shift of the HFM matched filter peak, leading to an error in the estimated timing of the wake-up tones. Furthermore, the 400 ms between the wake-up tones and the HFM will also be modified. For a Doppler speed of +/- 5 m/s with a JANUS Centre frequency ~3.9 khz (Fig. 1) it is simple to show that the HFM location error is ~1 ms and the second source of error is ~2.2 ms, so that the impact on selecting the wakeup tones (each with duration 80 ms) is minor. Nevertheless, we chose to enlarge each of the three sampling windows to capture the wake-up tones even if shifted by the maximum Doppler considered, at the expense of capturing slightly increased noise energy. Once the wake-up tones are localised, the time series for the three tones are summed to provide a 1/ 3 gain in Signal to Noise Ratio (SNR) (assuming uncorrelated noise). The Fourier-transformed PSD provides separate best-fit estimates for a second-order polynomial

5 curve fitting to the primary lobes. The three fits are then jointly optimised to capture the maximum power in the signal to provide the Doppler dilation estimate. Once the dilation factor is estimated, the original received time series can be re-sampled at the dilated frequency to provide a Doppler-compensated input to the regular decoder. We could also have chosen to exploit the 400 ms between the wake-up tones and the HFM by estimating Doppler based on the dilation of this interval. This would require detecting the wake-up tones independently (e.g. with an energy detector), enabling an estimation of the dilated time lapse between these and the HFM. There are many other possible combinations. 3. EXPERIMENTAL RESULTS Doppler estimation The accuracy of the Doppler compensation can be measured by comparing Doppler speed estimates with navigational records from the RV Planet. The GPS-derived ship velocity was projected onto the chord between transmitter and receiver to give the Doppler speed. The RV Planet made several turns and during these the towed source might be assumed to follow a different path to the vessel, cutting the corner, also possibly changing depth. We can also expect the estimated Doppler to differ from the GPS-derived ship speed when the RV Planet is close to and passing the receivers, due to the physical offset between the GPS receiver and the towed source. Comparisons between the Doppler estimated speed (in red) and GPS-derived speed (in green) are plotted against detected packet number in Fig 7 where estimates in the low frequency band are shown in the left hand panel, and those for the high frequency band in the right. Fig.7: Doppler-estimated (red) and GPS-derived (green) speeds for hydrophone 4 plotted against signal decoding number, with distance plotted The Doppler-estimated speeds are generally in very good agreement with the GPS-derived speeds except for two short periods (at the beginning and again at about 20% of the way through the experiment), during which the two frequency bands are consistent with each other, but not with the GPS data. These segments correspond to RV Planet manoeuvring or passing the receivers, as expected. Note that detections cease on both frequency bands at an opening range of about 12 km, but restart soon after RV Planet turns round at the end of her track run at 25 km. This could be due to changes in the acoustic propagation environment or perhaps the asymmetry of propagating from the source behind RV Planet, through her long wake trail, versus forwards. Excluding those parts of the experiment where we have imprecise information regarding the source navigation or where the estimator completely fails because of impulsive or too

6 high noise, the statistics of the differences between the Doppler estimated speed and the GPS estimate were computed. The resulting distribution is approximately zero-mean, with long tails (perhaps due to a non-linear breakdown of the Doppler estimator) compared to a Gaussian distribution. Table I gives the mean, r.m.s. and the mean and s.d. of the best Gaussian distribution fitted to the observed histograms. Frequency Actual histogram Best-fit Gaussian distribution Hyd. # Band Mean speed [m/s] r.m.s. [m/s] Mean speed [m/s] s.d. [m/s] 1 High Low High Low High Low High Low High Low High Low Table 1: Speed estimation difference distributions While the actual histograms have somewhat irregular shapes, the best-fit Gaussian distributions are generally consistent, with an average s.d. of 0.67 with statistically insignificant bias in the mean. For a Gaussian distribution we can expect 95% of estimates to lie within 2σ of the mean, corresponding to +/ m/s. In the 3.9 khz band, the chip length is 20 ms. If we take a 95% confidence limit on a maximum tolerable timing error of half a chip time at the end of the message, this limits message lengths to 11 s, which allows some 500 bits to be sent (after coding) allowing for the wake-up tones, HFM and the waiting periods in between. After decoding, this gives us 160 bits of payload (20 bytes), in addition to the 64 bits of the JANUS header packet. Since most of the JANUS coding parameters scale with frequency, this 20 byte performance can be expected to hold over a wide range of centre frequencies. Decoding performance We chose to use a Greatest Of - Constant False Alarm Rate (GO-CFAR) detector, originally developed for use in detecting signals in clutter [5]. A simple HFM matched filter may incorrectly trigger on correlated nearby noise resulting from strong multipath or changes in amplitude of the signal. Normalising the output of the matched filter with the maximum of two Average Cells, helps reduce this problem. We also chose to use soft Viterbi decoding, whereas the baseline decoder implementation provided on the JANUS wiki ( is hard. The baseline decoding success rate, without Doppler compensation, was only 82 out of 2184 signals. The lowest SNR at which a successful decoding occurred was 0.1 db. The maximum range was 15.3 km. With Doppler compensation, this figure rose to 215, 90% of the additional correct decodings being at Doppler speeds >2.2 m/s. Clearly, Doppler compensation helps significantly at speeds over the predicted baseline JANUS tolerance of 2 m/s. Nevertheless, we still have <10% of all the signals decoded correctly. With the idea to explore other

7 possible simple enhancements to the decoder, several additional processor enhancements were considered. The first enhancement was to use a longer traceback length of 48, compared to the computationally-simpler baseline traceback length of only 8. The baseline decoder is deliberately sub-optimal to respect potential limitations on decoder computational power and memory. The shorter traceback length was chosen for the baseline decoder to reduce memory requirements in legacy hardware implementations. Applying GO-CFAR (which does no worse, on average, than the simple HFM matched filter in this case and has proven valuable with other datasets) and soft Viterbi with a longer traceback increases the number of successfully decoded packets to 395, the farthest being at a range of 24.2 km. If we permit a maximum of 4 bit errors in the decoded packet, this number rises to 435. Finally, multipath effects can be not only suppressed, but actively employed to advantage if we are able to perform an effective channel equalisation. Considering the noise to be white in band and extracting the peaks from the matched filter output to estimate the taps of the channel, we can compute the Wiener filter deconvolution. Applying the equaliser with and without Doppler correction, the number of correct decoded signals then increases to FUTURE WORK As discussed in Section I, the effect of Doppler on JANUS signals is to both dilate the time (resulting in a loss of synchronisation on the decoding of the chip sequence) and frequency content (resulting in a shift in the frequency of the chip sequence). In the case of strong multipath, where several paths contribute significant energy to the arrival, the strongest being not necessarily the direct path, the Doppler shifts may be different for each path. For surface-interacting paths, the Doppler may shift significantly on time scales much shorter than the packet length. Our current Doppler compensation scheme assigns only one Doppler estimate, presumed valid for the summed multipath arrival and for the duration of the packet. To improve on this scheme, our first idea is to use a Phase Locked Loop (PLL)-inspired tracker to continuously track timing variations through the packet. Such a technique would, however, require substantial re-programming of the decoder itself and would also need a much more powerful decoder than the baseline JANUS standard implementation assumes is available. In principle, the frequency dilation could also be tracked in a similar way, with a feedback decision loop that decodes each chip and simultaneously estimates the instantaneous frequency dilation by optimising a fit to the frequency shift. The output of this optimisation would probably be fed to a model-based tracker such as a Kalman filter to track Doppler frequency dilation. To address the issue of multipath Doppler spread, we conducted some preliminary tests to see if widening the frequency and temporal acceptance window in the decoder could improve performance, attempting to capture more of the chip energy. The results were very positive. These proposed enhancements could potentially provide dramatically improved performance in the receiver (at the cost of requiring a much more capable receiver hardware set) but have the advantage that they do not require any changes in the encoded signal. The problem with accommodating Doppler frequency spreading by widening the acceptance window is that the 0 and 1 frequency-shifted chip windows now overlap, which is obviously undesirable. To combat this, the encoded FSK could be arranged to place chips +/- n sub-carriers up or down, rather than +/- 1, where obviously n>1. It should be possible to maintain pseudo-orthogonality in the FH sequence (so long as bandwidth wrapping problems do not predominate) while allowing a wider decoding window to be used without overlap. Unfortunately, such a modification would require a change to the JANUS standard, since it

8 affects the transmitter and encoder, rather than just the receiver and decoder. Given the strong potential performance improvement this could bring, however, we propose that this idea should be pursued for possible future versions of the JANUS standard. 5. CONCLUSIONS JANUS was designed for simplicity and robustness to provide a first contact unsolicited beacon functionality and a low-level data exchange protocol. The baseline standard was kept deliberately very simple to reduce the requirements for legacy systems to reach compliance to a minimum. We have obtained a set of JANUS transmissions and receptions up to long range (25 km), at relatively high speed (4 m/s) in two frequency bands in the Baltic Sea, known to be very challenging for acoustic communications. These data have been used to guide the design and testing of a simple Doppler compensation scheme that works as a pre-processor to the baseline decoder. Our very low computational complexity algorithm performs well, fulfilling the theoretical needs of the decoder for up to 20 bytes payload. JANUS packets were successfully decoded at ranges up to 25 km and speeds up to 4 m/s, down to SNR levels of 0.1 db using this compensation. Nevertheless, the achieved decoding performance remains at the ~10% level due to the complexity of the multipath and low SNR. Several further enhancements are suggested (e.g. PLL tracking of the instantaneous time and frequency dilations and a Weiner-filter based channel equalisation) that promise significant improvements. Finally, we propose another possible approach to accommodate Doppler spread, consisting of widening the acceptance window to capture the energy of a single chip in a wider frequency slot. The acceptance window could also be widened in the temporal direction to combat multipath spread. This leads to the proposal to investigate modulation with a larger spacing of the frequency carriers of set and unset bits for future versions of JANUS. 6. ACKNOWLEDGEMENTS The authors sincerely thank and acknowledge the many people who have contributed to this work through the sharing of their ideas and in their support of experimentation aboard RV Planet and elsewhere. Special thanks to Michael Hamilton for suggesting the GO-CFAR detector. REFERENCES [1] H.G. Schneider, MOCASSIN sound propagation and sonar range prediction model for shallow water environments, Technical Report FWG, Kiel, Germany, [2] K. McCoy, JANUS: From Primitive Signal to Orthodox Networks, In Underwater Acoustic Measurements, Technologies and Results, Nafplion, Greece, J.S. Papadakis and L. Bjørnø, [3] [4] J.G. Proakis "Digital Communications", McGraw Hill, Fourth Edition, 2001 [5] H. M. Hansen Constant false alarm rate processing in search radar, In Proceedings of the IEE International Radar Conference, London, pp , October 1979