Shipborne Underwater Acoustic Communication System and Sea Trials with Submersible Shenhai Yongshi

The Shipborne acoustic communication system of the submersible Shenhai Yongshi works in vertical, horizontal and slant channels according to the relative positions. For ease of use, an array combined by a vertical-cone directional transducer and a horizontal-toroid one is installed on the mothership. Improved techniques are proposed to combat adverse channel conditions, such as frequency selectivity, non-stationary ship noise, and Doppler effects of the platform’s nonlinear movement. For coherent modulation, a turbo-coded single-carrier scheme is used. In the receiver, the sparse decision-directed Normalized Least-Mean-Square soft equalizer automatically adjusts the tap pattern and weights according to the multipath structure, the two receivers’ asymmetry, the signal’s frequency selectivity and the noise’s spectrum fluctuation. The use of turbo code in turbo equalization significantly suppresses the error floor and decreases the equalizer’s iteration times, which is verified by both the extrinsic information transfer charts and bit-error-rate performance. For noncoherent modulation, a concatenated error correction scheme of nonbinary convolutional code and Hadamard code is adopted to utilize full frequency diversity. Robust and low-complexity synchronization techniques in the time and Doppler domains are proposed. Sea trials with the submersible to a maximum depth of over 4500 m show that the shipborne communication system performs robustly during the adverse conditions. From the ten-thousand communication records in the 28 dives in 2017, the failure rate of the coherent frames and that of the noncoherent packets are both below 10%, where both synchronization errors and decoding errors are taken into account.


Introduction
With the increasing trend of deep ocean activities, underwater acoustic vertical communication has been successfully tested and used in many scenarios. Representative results include communications with the human occupied vehicles (HOV) (Roberts et al., 2012;Zhu et al., 2013), data collection from the seafloor sensors to the unmanned surface vehicle (USV) (Kurowski et al., 2015), real-time seafloor image transmission from the Autonomous Underwater Vehicle (AUV) (Ahn et al., 2017), and communication between the gliders and the satellite surface moorings (Trask and Farrar, 2018). The acoustic communication sys-tem of the HOV Deepsea Challenger, which documented human communication across full ocean depth for the first time, was described by Roberts et al. (2012). When the HOV was diving to the depth of 10.9 km, in order to keep away from the noise of the mothership, the surface receiver was lowered from an inflatable boat to receive voice and short messages from the submersible. The acoustic communication system of the HOV Jiaolong (Zhu et al., 2013) faced the same problem of the high-level noise of the mothership. The mothership Xiangyanghong-09 was built in 1978 and was quite noisy (Liu et al., 2010). To suppress the noise, a transducer array of 4 elements was lowered to 200-300 m depth, and then four communication methods, including the coherent modulation, the noncoherent modulation, the spread spectrum and the single sideband voice modulation, were simultaneously verified at the distance of up to 7.7 km in the sea trials (Zhu et al., 2013). Kurowski et al. (2015) tested the slant communication and positioning between the USV and the underwater part. A reliable link up to a depth of 6000 m in heavy sea states was assured, and the packages success ratio was smaller than 90% over the distance of 4500 m. In addition, the investigation about the ship noise showed that a passing ship with 8 knots at a 100 m distance enhanced the noise level by larger than 20 dB. Ahn et al. (2017) used acoustic communication at a speed of up to 16 kbps to transmit compressed images from AUV at a maximum depth of 2000 m. Other experiments of low Signal-to-Noise Ratio (SNR) or high data rate communication towards potential vertical applications have also been carried out. Wu et al. (2015a) proposed a concatenated code of nonbinary Low-Density Parity-Check Code (LDPC) and constant weight code to work under low SNR, and carried out noncoherent communication experiments in the vertical channel of 5 km. By iteratively decoding with single receiving channel, the SNR threshold of 2 dB at a data rate of 357 bps was obtained. Suzuki et al. (2017) verified a 31.8 kbps transmission based on rate-7/8 convolutional code and 64 Quadrature Amplitude Modulation to a depth of 1000 m, and with one branch receiver, the SNR threshold was about 26 dB.
To improve the communication performances, channel characteristics should be first studied. The vertical communication channel is characterized with short multipath spread (Stojanovic and Beaujean, 2016) and has common characteristics with other kinds of underwater acoustic channels. Comparisons between the vertical and horizontal channels are listed as follows. (1) Multipath: In horizontal long-distance communication, the signal is transmitted by boundary reflections or through the deep ocean channel, leading to rich multipath and spatial diversity. In vertical long-distance communication, the direct path is strong, and boundary reflections are usually suppressed by directional transducers.
(2) Propagation loss: The propagation loss is the superposition of absorption loss and spreading loss. Absorption loss has identical effects on both channels and leads to the frequency selectivity of the wideband signal at a long-distance transmission. In vertical communication, the spreading plane is spherical and the spread loss is deterministic and proportional to the square value of distance. The spreading loss in horizontal communication is complicated, depending on the existence of the waveguide and the loss values of boundary reflections. (3) Doppler effect: The Doppler scale is the ratio of the relative transmitter-receiver velocity to the speed of sound. The channel's time-variant is more obvious with the surface platform than that with the subsurface mooring (Freitag and Singh, 2015). When com-munication with the surface platform, its undulating movement has a larger component vertically than that horizontally, and therefore in vertical communication, the non-constant Doppler scale is much more obvious.
To narrow the gap between the experimental results and the theoretical channel capacity (Stojanovic and Beaujean, 2016), the vertical communications with higher data rate, higher robustness, and lower SNR threshold are necessary and feasible. Depending on the conditions of application, main problems and potential solutions in vertical communication are listed as follows.
(1) Non-stationary and highlevel noise: The mothership in deep ocean experiments is in large tonnages and in most cases cannot be halted for safety reasons. Therefore, the ship noise often leads to a low SNR. Many effective engineering methods have been applied, such as decreasing the data rate, using the low-noise surface platforms (Roberts et al., 2012;Kurowski et al., 2015), or lowering the transducer array to a certain depth (Zhu et al., 2013). Advanced communication techniques without loss of operational convenience or data rate, such as capacity approaching code and modulation schemes (Stojanovic and Beaujean, 2016;Tao, 2016), are more attractive. (2) Consistent frequency selectivity among receivers: As the result of frequency dependence of acoustic absorption factor and the fluctuation of the transmitter's frequency response, frequency selectivity is serious especially for the wideband signal. The spatial diversity gain is not as effective as it is in the horizontal channel, because in the vertical channel, different receivers undergo the similar selectivity. To combat frequency selectivity, single-carrier turbo equalizer (Choi et al., 2011;Stojanovic and Beaujean, 2016) or frequencycoded multicarrier (Zhou and Wang, 2014) schemes would be studied. (3) Non-constant Doppler scale caused by the platform movement: Non-constant Doppler scale is more difficult to be estimated and compensated than the constant Doppler scale. In actual experiments under nonstationary conditions, synchronization errors are of large proportion, and robustness to the actual channel of low-complexity synchronization technique should be improved.
A shipborne system for the submersible Shenhai Yongshi is designed to work robustly under the unfavorable conditions as mentioned above. Although the maximum distance is achieved in the vertical channel of the deep ocean, from practical views, the communication system is also used in the horizontal shallow water and required to cover the varying direction when the submersible is diving or ascending. To provide high operational flexibility, we use a shipborne array formed by two different transducers: a vertical-cone directional transducer and a horizontal-toroid one. Therefore, the receive algorithm should be adaptive to different channel models by using only two transducers. The superiority of the proposed system includes: (1) Coherent modulation: A turbo-coded single-carrier scheme is used in the transmitter's signaling, and in the receiver, the sparse decision-directed Normalized Least-Mean-Square (NLMS) soft equalizer automatically adjusts the tap pattern and weights according to the channel path structure, the two receivers' asymmetry, the signal frequency selectivity and the noise spectrum fluctuation. The use of the turbo code in turbo equalization significantly suppresses the error floor and decreases iteration times with the equalizer. Its superiority to ordinary adaptive turbo equalization is verified by means of extrinsic information transfer Charts (EXIT) and bit error rate (BER) performances based on the actual data.
(2) Noncoherent modulation: A concatenated error correction scheme of the nonbinary convolutional code and Hadamard code is adopted to utilize full frequency diversity. BER performance simulations under Rayleigh fading channel show the nonbinary scheme's superiority to the binary scheme of the standard JANUS (Potter et al., 2014), where the binary convolutional code and Binary Frequency Shift Keying (BFSK) modulation are adopted. (3) Synchronization techniques: Low complexity but robust synchronization techniques in time and Doppler domains are proposed. The time detection is based on the Generalized Logarithm Ratio Test (GLRT), which is more robust to the non-stationary condition than to the ordinary matched filter (MF) method. A Doppler detection scheme is proposed to work robustly in the time-varying multipath channel based on three processing steps of the preamble and postamble Linear Frequency Modulation (LFM) waveforms: matchedfilter, absolute value, and correlation.
The sea trials with the submersible to a maximum depth of over 4500 m show that the shipborne communication system performs robustly under the adverse condition. From the tenthousand communication records of the 28 dives in 2017, the failure rate of the coherent frames and that of the noncoherent packets are both below 10%, where both synchronization errors and decoding errors are taken into account.

Single carrier coherent modulation and sparse turbo equalization
The high-data-rate coherent modulation can be realized through two approaches, single carrier (Stojanovic and Beaujean, 2016) and Orthogonal Frequency Division Multiplexing (OFDM) (Zhou and Wang, 2014). Compared with OFDM, the single carrier method has the benefits of low Peak-to-Average Power Ratio (PAPR) and fast-tracking ability (Tao, 2016), which are important in the deep ocean application. The well-known single carrier receiving scheme in underwater acoustic communication is the multichannel adaptive equalizer with phase tracking (Stojanovic and Beaujean, 2016;Zhu et al., 2013). However, equalizing with insufficient spatial diversity (Pajovic and Preisig, 2015) and exploring the inherent sparsity of the equalizer (Tao et al., 2017) are still open problems in both algorithm research and system realization. Turbo equalization (Choi et al., 2011;Yellepeddi and Preisig, 2015;Stojanovic and Beaujean, 2016;Duan et al., 2018) is a powerful tool to approach the non-ISI decoding performance. To suppress the error floor phenomenon (Yellepeddi and Preisig, 2015) of turbo equalization with single-component channel code, we use multiple-component channel code in turbo equalization to introduce an additional iterative component. In our scheme of shipborne communication system, as shown in Fig. 1, the sequence of one-frame bits is firstly turbo encoded at half rate, which is realized by two parallel Recursive System Codes (RSC), and then mapped by Quadrature Phase Shift Keying (QPSK). The QPSK symbols are pulse shaped into passband waveform. After inserting the training symbols and synchronization signals, the waveform is transformed into the acoustic signal by the transducer.
The main differences of our receiving scheme from the well-known phase-tracking adaptive equalizer technique are the sparse placement of the equalizer taps, and the exchange of soft information iteratively among the turbo decoder's two components and the soft equalizer. In the receiver, after the steps of down carrier conversion, frame synchronization and Doppler compensation, the waveform sampled at two times of symbol rate is iteratively processed by a sparse linear turbo equalizer (Wu et al., 2015b). Soft information is exchanged iteratively among the soft-in-soft- x n out (SISO) equalizer and the turbo decoder's two components. Bahl-Cocke-Jelinek-Raviv (BCJR) algorithm is used to decode the RSC components separately, and simplified by the Max-Log-MAP method independent of SNR estimation. The SISO equalizer is an adaptive decision-directed equalizer with sparse taps. Given the received symbol vector y n and the a priori mean vector , the soft estimation of the nth symbol x n is written as: , and w n is the weight vector of sparse taps. The sparseness of the equalizer is described by the support sequence s, whose elements are of binary-value. The weight w n is adaptively updated by NLMS, as: where is the adaptive step size, is the estimation error, and is the hard decision based on . The sequence s is updated for each frame in a multiple-threshold method as follows. At the beginning of the equalization, all elements of s are set to 1. If a tap's amplitude at the end of the training symbols is smaller than a given fraction of the max-imum of all taps, the corresponding element of s is set to 0. The sparse equalization is repeated several times with the same training sequence and the newly generated s, and the fraction thresholds are set to 0.1, 0.2, and 0.4, respectively. The span of the feed-forward filter is set to 200 samples (0.02 s) of each channel at a low convergent speed firstly. After sparse processing, the number of the useful taps is reduced to nearly its one-tenth.
The EXIT charts (Pajovic and Preisig, 2015;Duan et al., 2018) are used to track the iteration trajectories, and comparisons between the proposed scheme and the scheme of RSC code and full-tap equalizer are shown in Fig. 2. The two RSC components of the turbo decoder, which are combined as one component in the EXIT analysis, are internally iterated one time after each equalization iteration. It is shown that the sparse equalizer has a higher outcome at its output axis leading to a lower SNR threshold, and the turbo code converges more quickly to full mutual-information output leading to a lower error floor. The BER performances of two schemes in Fig. 3 show the proposed scheme's superiority on the error floor and iterative times, which verifies the EXIT analysis.

Nonbinary coded multicarrier noncoherent modulation
For noncoherent modulation, the concatenated scheme (Wu and Zhu, 2014) of 32-ary convolutional code (dual-5 code) and a Hadamard code (20, 5) are used, as shown in Fig. 4. It is modulated by the on-off keying modulation (OOK) and transmitted through 120 orthogonal subcarriers, which changes the frequency-selective fading channel into non-selective subchannels. The cyclic prefix is further used to remove inter-symbol interference, similarly as it works in OFDM system. A simple scheme of random phasing, clipping, and filtering is effective to reduce the PAPR to 6.7 dB without obvious SNR loss. The nonbinary coding is more powerful than the scheme of binary convolutional code and BFSK in the standard JANUS. Fig. 5 shows the simulation results under Rayleigh fading channel. To achieve a BER of 10 -4 , the requirement of SNR per bit of the nonbinary scheme is 11.3 dB and is 3 dB better than the binary scheme of JANUS.

Robust synchronization in the adverse condition
For low computation complexity, the Doppler-insensitive waveform LFM (Zhou and Wang, 2014) is used as the synchronization signal. Synchronization techniques in the time and Doppler domains need to be enhanced to fit the channel multipath, the non-stationary noise, and the nonlinear platform movement.

Preamble detection based on GLRT
In the shipborne application, the level of the manufactured noise is rapidly time-varying at a large dynamic range, and on the submersible side, the dynamic range of the filtered noise is over 40 dB. Therefore, the matched-filter detection method with the fixed threshold or the adaptive threshold is not applicable. A practicable method supposes that the variances of noise and signal are both supposed to be unknown, and therefore the signal is detected based on GLRT. The binary hypothesis is written as where x[n] is the receiving waveform in baseband, w[n] is the additive white Gaussian noise with unknown variance; c[n] is the original signal to be detected. The symbol A stands for the received signal's complex amplitude, which is also unknown. The detection statistic is constructed as: where and stand for the ML estimations of the noise variance under H 0 and H 1 respectively, and The simplified and equivalent statistic is then written as: The GLRT detection has the same expression as the normalized matched filter (Diamant, 2016). The advantage of GLRT analysis is that it can be easily extended to more complex problems, such as time detection with unknown Doppler scale, Channel Impulse Response (CIR) and other nuisance parameters. In our application, we choose N=256 and c[n] is the LFM. The detection threshold is set as T(x)>0.12, which is a constant false alarm detection. When the noise is stationary in the detection window, the false alarm rate is 3×10 -6 . The advantage compared with that of the traditional detection method is that the false alarm rate is limited when the noise is suddenly enhanced during the detection window. When SNR=-3 dB, the missed detection rate is smaller than 10 -4 . Sliding window detection is realized in transversal filters.

Doppler scale estimation under time-varying multipath
Doppler scale is estimated based on the preamble and postamble LFM signals for each frame. A simple method is  based on the positions of the maximum outputs of the matched filter when fed with two received LFMs (MF-ABS-MAX) (Zhou and Wang, 2014). The MF-ABS-MAX method is not robust to multipath. For example, the CIR has two equal-amplitude components, and when taking different components as the maximums of the preamble and postamble CIRs, the MF-ABS-MAX has a wrong estimation. A remedy method is based on the maximum position of the correlation between the two outputs (MF-CORR-MAX). However, when the multipath is time-variant, MF-CORR-MAX may also fail. For example, the CIR is formed by two equal-amplitude paths, and the sliding time of one path equals nearly half the carrier cycle, then its CIR coefficient will be reversed, leading to the cancellation of the correlation.
The proposed Doppler estimation method is based on three steps: matched-filter, absolute value, and correlation (MF-ABS-CORR-MAX), as shown in Fig. 6. The difference from the MF-CORR-MAX methods is the absolutevalue operation before correlation, by which the correlation's cancellation is avoided. The LFM duration and the frame duration are 25.6 ms and 500 ms, respectively, and the estimation accuracy of Doppler scale is 5×10 -5 .
When the ship is heaving, the CIR's main path position is varying at a period of about 7 s, and the Doppler scale is rapidly fluctuating, as shown in Fig. 7. In our system, the Doppler scale was estimated for each 0.5 s, and compensated by the Farrow resampling filter in the baseband and the following frequency-offset correction. The compensation accuracy meets the need of the noncoherent demodulation. In the coherent communication, the residual Doppler offset is tracked by phase-locked-loop.

Sea trials with the submersible Shenhai Yongshi
A 4500 m-depth manned submersible, named Shenhai Yongshi (Deep-sea Warrior), was designed and developed in China to obtain better reliability and maintainability than the former submersible Jiaolong. The new communication system inherited all the function of multiple transmission services from the Jiaolong system (Zhu et al., 2013): exchanging various sensor data periodically, non-scheduled communication in voice or text, and upward transmission of in-situ images. A shipborne transducer array, functionally replacing the lowered transducer array, was installed on the mothership's bottom. Compared with the lowered array, the shipborne array does not restrict the manipulation of the ship during the launch and recovery of the submersible. The mothership Tansuo-1 was reformed from an old cable-laying ship and had nearly the same noise level with the submersible Jiaolong's mothership Xiangyanghong-09. The noise level measured by the shipborne longitudinal transducer on the ship Tansuo-1 was 26 dB higher than that of the lowered array for the ship Xiangyanghong-09. Table 1 shows the digital transmission performances of the different HOVs. All systems faced the problem of the high-level noise of the motherships. The proposed algorithms ensured that the Shenhai Yongshi acoustic system can work in the adverse shipborne condition.
The combined transducer array was installed on the mothership Tansuo-1 by the lift mechanism for ease of use, as shown in Fig. 8. The array was formed by two different transducers as shown in Fig. 9, and both of them were horizontally omnidirectional. The mosaic transducer, whose directional pattern is shown in Fig. 10a, is preferred when the submersible is near the surface. When the submersible is  WU Yan-bo et al. China Ocean Eng., 2018, Vol. 32, No. 6, P. 746-754 751 diving in depth, the longitudinal transducer has a higher SNR, as its pattern shown in Fig. 10b. The transmitter uses only one of the transducers, switched according to the submersible's depth. The waveforms received from both transducers are combined in coherent or noncoherent method to archive a higher demodulation SNR.
During an over 50-day expedition in 2017, the manned submersible Shenhai Yongshi reached a maximum depth of 4534 m to test its functions and performances. The shipborne acoustic communication system performed very well and achieved robust transmission of sensor data, text, images, and voice. The SNR of each noncoherent packet was measured and recorded, and its relationship with distance is shown in Fig. 11. For the reason of the lower PAPR, the SNR of the single-carrier coherent waveform was nearly 3 dB better than that of the multicarrier noncoherent waveform. The SNR curves predicted by the spherical spreading and the absorption loss are also shown. When the submersible was near the surface, the transducers' direction and the acoustic channel caused serious SNR loss. When the submersible was diving or ascending, the link direction was relatively favorable, and the SNR varied according to the distance. When the submersible cruised near the seafloor at the depths of 1500, 3500, and 4500 m, the link was sometimes in the slant direction and the ship increased its speed, causing the reduction of SNR. It is shown that at the distance of 4500 m, the SNR of the vertical directional transducer was below 0 dB, and the noncoherent communication was still working.
Information sources and the data size of the digital communication are shown in Table 2. If no image was transmitted, the sensor data and text were encoded in one noncoherent packet, automatically updated every minute.
When images were transmitted upward, the coherent communication was used. The original true color image was resized up to 260000 pixels, compressed in JPEG-2000 format, and transmitted together with other digital sources  (Roberts et al., 2012) Inflatable boat Unknown 10.9 Non-avail. Non-avail. SMS messages Jiaolong (Zhu et al., 2013) Lowered to 300 m depth 50 7.7 300 5000 16 Shenhai Yongshi Shipborne 76 4.5 446 5000 40 Fig. 8. Lifting mechanism for the shipborne array. in one coherent packet. The coherent packet was divided into frames, with a typical setting of 72 data frames appended by 18 Reed-Solomon-encoded redundant frames. The duration of the encoded packet including synchronization was 45 s. In-situ images received by the shipborne acoustic communication system are shown in Fig. 12.
The shipborne system's performances during 28 dives of the submersible are shown in Table 3. The failure rates were the ratios of the numbers of packets successfully decoded by the ship to the numbers of packets transmitted by the submersible, and failures in both synchronization and decoding were included. The total failure rates of noncoherent packets and coherent frames were both below 10%. The failure rate of the coherent packets was remarkably decreased from 50.0% to 18.5% by inter-frame RS coding with little redundancy in each packet, which avoided the retransmission's disadvantages including high delay and low channel utilization.

Conclusion
The proposed techniques in the coherent modulation, noncoherent modulation, and synchronization for the shipborne system provided the robust communication with the manned deep submersible and avoided the use of the lowered array on a noisy ship. The total failure rates of noncoherent packets and coherent frames were both below 10% from the records of the 28 dives in 2017. Although designed for the submersible Shenhai Yongshi, the shipborne system can also be applied to communicate with autonom-   ous underwater vehicles, and submerged buoys and other platforms. The proposed algorithms will be enhanced to develop a high-data-rate shipborne communication system for full ocean depth, where a planar multi-element array is taken into consideration to improve the SNR by narrow beamforming.