Resumen de Algorithms for propagation-aware underwater ranging and localization
Oceans occupy more than two-thirds of Earth, while remaining one of the most underexplored parts of our planet. Collecting more data about the ocean can help us strike a balance between human activities and the conservation of marine life, or to prevent the devastating effects of disasters such as earthquakes and floods. Moreover, several events can be relieved or solved through localization. This includes threats, such as intrusion in sensitive areas, or even attacks to key infrastructure (ports, ships, military navy areas, or energy/resource mining systems). However, it also includes distress calls, safety issues, dangerous situations that require to know where the event is actually taking place. Being able to localize a threat or a distressful event helps prevent, or at least limit the impact of the event.
Despite the importance of ocean exploration, the underwater environment is extremely difficult to study: the challenges related to human diving, the high complexity of special equipment for immersion, and, perhaps most importantly, the intricacy of underwater communications with which we could collect data from distributed sensors or navigate remote devices such as autonomous underwater vehicles (AUVs) still prevent humans from fully interacting with oceanic environments.
While cabled communications are still the standard for long-term infrastructure deployment (e.g., for underwater oil&gas extraction monitoring), laying cables at the bottom of the ocean is very costly and challenging. Therefore, when possible, wireless connections are a preferrable choice.
Most of the regular terrestrial radio wireless communication and satellite navigation systems are based on high-frequency radio waves. However, radio waves hardly propagate in salty ocean waters, which are electrically conductive and do not support well the propagation of electromagnetic waves. This characteristic strongly affects the way communication and navigation systems are built under the surface of the ocean. The creation of cheaper and faster deployment methods for various distributed underwater systems is limited by the need to localize underwater devices, as most algorithms employed in terrestrial radio networks cannot be directly applied in underwater networks.
Currently, most commercial off-the-shelf (COTS) underwater ranging and localization systems are based on acoustic signals. Depending on the frequency, these signals can propagate from a few meters (for high-frequency ultrasound signals) to tens or hundreds of kilometers (for very-low frequency, sub-100 Hz signals).
Nevertheless, the physical properties of acoustic signal propagation in an underwater environment have several implications on localization algorithms. First, the speed of propagation of sound in water is much lower than the speed of propagation of radio waves in air. Second, refraction in non-homogeneous water media and multiple reflections from different features of the environment create complex, rapidly changing multipath channels, often without line-of-sight (or specular) components. Third, underwater wireless sensor networks are characterized by the complexity of synchronization among network nodes, e.g., due to the influence of high pressure and temperature changes. Additional challenges also include a complex deployment process, as well as limited energy and space that prevent from embedding several sensors on the same node. For example, the latter implies that in some low-power deployments it may be convenient not to have a full transceiver, but only a receiver device for a given task.
In this thesis, we try to address most of these challenges and investigate different problems related to underwater ranging and localization. In particular, we focus on algorithms that consider underwater acoustic channel properties. This group of algorithms utilizes additional information about the environment and its impact on acoustic signal propagation in order to improve the accuracy of location estimates, or to achieve a reduced complexity, or a reduced amount of required resources (e.g., anchor nodes) compared to traditional algorithms.
For certain applications that consider localization of sporadic targets it is important to simultaneously observe broad underwater environments. While most modern localization methods assume the deployment of multiple sensors, the deployment of a large number of active sonars to cover wide areas is often challenging, and may be ineffective in terms of energy and expenses. Moreover, active acoustic equipment may introduce various changes in natural habitat of marine animals, and in any event it is not advisable to deploy active systems and leave them unattended for long periods of time. AUVs can be another key instrument for observation of marine life, but localizing them under water precisely and in real time requires significant equipment to date, posing further issues for their extensive use.
In this thesis, we first tackle the problem of passive range estimation using the differences in the times of arrival (TDoA) of multipath replicas of a transmitted acoustic signal. This is a cost- and energy- efficient algorithm that can be used for the localization of AUVs and utilizes information about signal propagation. First, we discuss the performance of this algorithm by considering that it is unfeasible, in typical scenarios, to have perfect knowledge about the physical characteristics of the environment. We analyze the sensitivity of our ranging method versus environmental parameter uncertainty by modeling the acoustic channel via the well-known ray tracing software Bellhop. This makes it possible to study the effects of erroneous input environmental data on ranging. We study the accuracy of our method in the simplified case of constant sound speed profile (SSP) and compare it to a more realistic case with various non-constant SSPs and in the presence of erroneous TDoA as well as bottom depth data. We also propose an auxiliary quantity called effective sound speed. This quantity, when modeling acoustic propagation via ray models, takes into account the difference between rectilinear and non-rectilinear sound ray paths. According to our evaluation, such capability offers improved range estimation results with respect to standard algorithms that consider the a local, single-value measurement of the speed of sound.
Then we propose another approach for non-collaborative target localization inspired by traditional wireless sensor networks (WSN) fingerprint-based localization. Several fingerprinting methods work based on the correlation of some significant and distinguishable characteristics of the channel, such as the received signal strength or power-delay profile. In particular, the quantity of interest is measured when communicating with the desired target, and correlated with templates of the same information collected at different locations and stored into a fingerprint database. However, the underwater acoustic channel tends to be much more dynamic than terrestrial radio channels, with several multipath components from multiple reflection over the surface, bottom and volume scatterers.
In this approach, we utilize the dependence of the acoustic channel on the spatial diversity of the bathymetry profile. The proposed method is based on the modeling of expected acoustic channel impulse response (CIRs) for different possible locations of an acoustic source around a moored receiver. For out method to work, we thus assume that the receiver knows the bathymetry and SSP. Additionally we assume that the transmitted signal is known or can be estimated with the accuracy needed for CIR evaluation. For each step of the algorithm, we compare the channel properties of the received signal and those in the database of modeled CIRs by means of a normalized match filter. As surface reflections are typically very variable and may reduce the correlation with the modeled templates, we remove them from the model.
We collect several receptions from the mobile AUV and organize the locations showing the highest correlation into a trellis, over which we run a search algorithm.
Like the Viterbi algorithm, our trellis algorithm searches for the most likely source path. It accounts the probabilities of finding the target at each point covered by the modeled CIR database, and the transition probabilities from one position to another, for example, by limiting the maximum target speed. We limit the complexity of the algorithm by constraining the number of probable positions at each step to filter out locations that were unlikely visited. We also make the algorithm resistant to situations where no significant correlations between received CIRs and database-stored CIRs were found, and thus no probable positions could be estimated for a given transmission.
We evaluate the performance of our method through simulations (based on real bathymetry and sound speed information) in presence of imperfect knowledge about the environment. Then, we also validate the approach in a proof-of-concept sea trial. We compare the proposed algorithm with the classical Viterbi algorithm, as well as with common approaches such as a Kalman-filtered trajectory and a highest correlation method. The sea experiment demonstrates the applicability of our method to real sea conditions with a localization error as low as 5.8%, which is a remarkably good accuracy given the use of a single stationary receiver and the realistic, imperfect bathymetry and sound speed profile measurements. While these errors may seem large we argue that for many applications these values are acceptable.
In underwater scenarios, accurate localization typically requires acoustic arrays. These pieces of equipment encompass multiple hydrophones or acoustic transceivers. In search of a balance between localization accuracy and the cost and complexity of deployment, there could be situations when, for a given accuracy, it is necessary to use multiple pre-assembled arrays. Fusing information from multiple acoustic sensors can have a positive effect on the reliability of the results. Additionally, forming complex 3D designs can help discriminate arrivals that could not have been singled out using simpler 2D (e.g., linear or rectangular) arrays. Despite this, it is often the case that such “opportunistic” designs may incur suboptimal layouts. For example, due to design considerations such as cabling, batteries or other construction issues, joining multiple arrays into a single, larger array may not meet the spatial sampling constraint, which prescribes a distance of no more than half the minimum acoustic wavelength between subsequent elements. In this case, the ability to use information from all the sensors is limited, due to the spatial ambiguity caused by the improper spacing of array elements.
Additional challenges come from the fact that acoustic arrays often require wideband processing. Notably, most wideband array processing algorithms work with predefined array shapes, or are limited to 2D, to specific signals, or to a known number of targets. In some cases, the preferred solution is to directly employ particle velocity sensors.
Our work on the problem of designing “opportunistic” acoustic arrays was conceived in the context of the EU H2020 SYMBIOSIS project. The SYMBIOSIS platform is an optical-acoustic system for biodiversity monitoring. For this platform, we developed a system for the 3D acoustic localization of marine fauna based on acoustic arrays consisting of several sub-arrays combined together into an arbitrary layout.
We worked on a localization algorithm for pelagic fish species using a software-defined version of commercial ultra-short baseline (USBL) arrays of five elements.
The algorithm consists of several steps. First we apply bandpass and normalized matched filtering (NMF) to the raw acoustic data. Then we detect relevant peaks throughout the filter’s output, and cluster peaks from different channels using the DBSCAN algorithm. Each cluster is assumed to be a target detection. The crop of the NMF output that contains the likely target detection is then passed on to two separate processing chains. The first performs the actual wideband acoustic spatial processing by applying a wideband delay-sum direction of arrival (DoA) algorithm. The second exploits TDoA information related to the reception of the signal at different array elements to roughly localize the target via a TDoA-based multilateration algorithm. The latter provides a means to create a mask around the likely location of the target. This mask consists of a Gaussian kernel, and is applied to the spatial processing output to remove the ambiguity arising from suboptimal array layouts. In addition, should the algorithm detect a significant arrival from a surface or bottom reflection, TDoA information is used to filter out target location estimates that fall out of the watercolumn boundaries, assumed to be known.
The proposed approach yields the following advantages: (i) it provides a framework to merge together smaller arrays into a larger “array of opportunity” to achieve better DoA estimation accuracy; (ii) it provides a method to rule out the ambiguity that may result from the suboptimal spacing of the array elements; (iii) it works with wideband signals and arbitrary array topologies; (iv) it yields good performance in emulated sea environments, in a proof-of-concept experiment and also in larger experiments involving marina fauna localization in the wild. We conducted these trials in different environments, including lakes and sea water bodies, with a variety of targets and in different acoustic conditions.
For emulation results, the key idea is to employ measurements of noise and acoustic clutter from a lake experiment, in order to achieve a more realistic representation of the signal received by the array elements and test the algorithm in a controlled manner on various array designs. In particular, emulation results show that, at a Signal-to-Noise Ratio (SNR) of −20 dB, a 15-element array of opportunity achieves lower average and median localization error (27 m and 12 m, respectively) than a 30-element array with proper λ/2 element spacing (33 m and 15 m, respectively).
In the proof-of-concept lake experiment, we additionally show that our algorithm achieves a 90th-percentile DoA estimation error of 4◦ and a 90th-percentile total location error of 5 m when applied to a real 10-element array of opportunity.
The tests in the wild were scheduled throughout 2020, and therefore the final design of the SYMBIOSIS platform hosting our algorithm was challenged by the difficulties related to the well-known pandemic-related events of 2020. In this regard, changes were made to the algorithm that implemented several additional functions, and made it possible for the algorithm to still wok, even if other key elements of the platform design should fail. In order to provide a full solution for the detection, localization and tracking of underwater fauna, we extended and updated the algorithm. The algorithm forecasts when fish targets approach the platform and triggers image acquisitions from the platform's cameras. Therefore, the algorithm has to consider historical information from different acoustic records taken over time. The extended version of the algorithm includes the detection of possible targets, discrimination of stationary acoustic arrivals (that typically result from reflections of transmitted signals off environment features and platform parts), and the tracking of detected targets. Here we apply several additional steps of DBSCAN clustering.
The SYMBIOSIS platform and our algorithm were deployed and tested in two different sites. One is a deep water area, where the deep THEMO observatory is moored, a few tens of km west of the Israeli coast, in the Mediterranean sea. The second one is a mooring in the Israeli red sea, in the proximity of Eilat. The deep THEMO deployment represented a perfect opportunity to verify and tune the capabilities of the algorithm to discriminate between static and slowly moving targets. In this experiment, the SYMBIOSIS team released a rehabilitated turtle. In the recorded acoustic signals we observe stationary arrivals detected that are probably reflections from parts of the deployed platform. It was successfully removed from further processing by our algorithm. The algorithm shows consistent and slowly changing bearing angle estimates as the turtle swims away.
The Eilat deployment, in contrast, has very different acoustic channel properties with very shallow reef bottom and rich multipath from environmental features. The dataset consist of one working day consecutive records. In line with previous results, we observe different clusters of detections in this datased, one associated to bearing angles of 100°, and a second one with bearing angles of about 330°. Considering only the paths tracked, most of the detected targets were moving slowly, at less than 0.1 m/s.
The outcomes show that the algorithm can localize not just highly reflective or active targets but also smaller and weaker targets such as the fish species of interest for the SYMBIOSIS project. The algorithm works as expected in different environments.
Summarizing the results of these tests, the outcomes show that the algorithm effectively leverages “arrays of opportunity” obtained by merging different sub-arrays into a larger array, in order to detect underwater targets both in controlled experiments and in the wild.
As additional contributions, in this thesis we present an algorithm designed for underwater Light Detection And Ranging (LiDAR) signal processing, and experimental work related to its validation. Moreover, we report the details of a freely shared data set for underwater network emulation named ASUNA. This dataset joins acoustic link quality measurements from several experiments carried out in different scenarios, and is meant as a test tool for developers of underwater communication and networking algorithms to test their solutions before actually testing them at sea with real hardware.
Summarizing the main contributions of this work, first we have investigated the algorithms that improve the accuracy of ranging by using additional information about the underwater acoustic propagation environment, and studied the effects of imprecise environmental knowledge on the accuracy of underwater localization and range measurements. Then we also presented a scheme to estimate the direction of arrival of acoustic signals reflected by underwater targets using wideband hydrophone arrays of opportunity. Our proposed scheme solves this issue by fusing direction-of-arrival information with coarse multilateration outputs. This makes it possible to eliminate most of the ambiguity, and yields accurate direction-of-arrival estimates. Our results show that our scheme achieves satisfactory direction of arrival estimates.
Overall, this work has shown the effectiveness of using additional information about acoustic signal propagation and its benefits in further application of this approach to similar problems.
