An open discussion on machine learning and artificial intelligence applications in the shipping industry.
Welcome to my blog on Technology Space with Machine Learning (ML) and Artificial Intelligence (AI), a discussion platform on the present and future technologies that will often relate to ML and AI applications in a more scientific sense. Of course, this discussion is biased towards the shipping industry.
Lokukaluge Prasad Perera (Professor in Maritime Technology at UiT The Arctic University of Norway & Senior Research Scientist in Smart Data at SINTEF Digital)
About the Blogger…
L. P. Perera received BSc (1999) in Mechanical Engineering and MSc (2001) in Systems & Controls from Oklahoma State University, USA and PhD (2012) in Naval Architecture and Marine Engineering from Technical University of Lisbon, Portugal. Currently, he is a Professor in Maritime Technology at UiT The Arctic University of Norway, and also a Senior Research Scientist in Smart Data at SINTEF Digital, Norway. His research experience includes SINTEF Ocean – Norway (2014-2017), Centre for Marine Technology and Engineering – Portugal (2008-2012) and Advanced Technology Research Center – USA (1998-2001). His academic experience includes Naval & Maritime Academy – Sri Lanka (2005-2008) and Ocean University of Sri Lanka – Sri Lanka (2003-2005). Furthermore, Prof. Perera was a visiting lecture (2001-2005) for several academic institutes in Sri Lanka: University of Ruhuna, University of Moratuwa, Open University of Sri Lanka, Colombo International Nautical & Engineering College. His industrial experience includes Wartsila Finland – Finland (2012-2014). Dr. Perera has published more than 100 peer-reviewed papers in reputed international journals and conferences. He has also been categorized into a group of the WORLD’S TOP 2% SCIENTISTS (2021 – 2022) by a Stanford University study.
Collision avoidance has been an important topic in various transportation systems, due to its contribution to navigation safety. The interest in the same topic has been highlighted in recent years, due to the recent developments in autonomous navigation systems in various transportation sectors. The requirements of collision avoidance can vary from one transportation system or another and should eventually influence the respective decision-making process of collision avoidance. Possible close vehicle/vessel encounter situations for land and maritime transportation systems are considered in Fig. 1 and that may result in possible collision situations.
As presented in the figure, land transportation systems are facilitated by a structured environment, where the respective road infrastructure, traffic light systems, pedestrian crossings etc. are established to improve navigation safety. However, the same road infrastructure limits the space that a vehicle can navigate in close encounter situations, especially in road intersections (see possible crossing area in the left figure of Fig. 1). Hence, decision-making in a road intersection has been considered a complex problem to solve in autonomous navigation, especially under land transportation systems. The main reason is that it can introduce only one intersection point for both vehicle trajectories. Therefore, traffic light systems are introduced to manage vehicle navigation and avoid possible collisions in land transportation systems. Therefore, only one vehicle can pass this trajectory intersection point within a selected time interval due to such traffic light systems. If road intersections are not facilitated by traffic light systems, then collision avoidance can be a complex problem in land transportation systems, i.e. due to navigation limitations under the road infrastructure.
However, this limitation may not be applicable to maritime transportation, since ocean-going vessels are not limited to any road infrastructure, other than shipping lanes in some situations. Therefore, the navigation area for ocean-going vessels is somewhat wider (see possible crossing area in the right figure of Fig. 1). That can introduce infinite possible intersection points for both vessel trajectories. This condition makes ship collision avoidance a simpler problem since there are an infinite number of interaction points, i.e. or possible solutions, towards ship collision avoidance between these two vessels. These solutions can be categorized as possible sub-optimal ship navigation trajectories that can be used by the vessels to avoid close encounter situations. Therefore, possible ship collision situations can be avoided with the same solutions.
However, various advanced algorithms for supporting collision avoidance among manned vessels have been proposed in the recent literature (Zhang et al., 2021). On the other hand, these proposed methods should be investigated further to understand their capabilities in applying them to present and future ocean-going vessels, i.e. even with autonomous navigation capabilities. On the other hand, many algorithms are focused on finding optimal solutions for ship navigation environments. However, by considering Fig. 1, it can be concluded that ship collision avoidance can have many possible sub-optimal solutions, due to the reason that an infinite number of possible trajectory intersection points can be identified and that can be utilized towards eliminating possible close encounter situations, as mentioned before. These solutions represent situations where a slight change of course or speed conditions of one vessel can completely eliminate a possible close encounter situation of both vessels.
On the other hand, it would be difficult for vessels to follow optimal trajectories, i.e. in collision avoidance, due to several reasons. One should note that if a vessel is following an optimal trajectory there are three conditions that should be satisfied. i) the vessel should be in a specific navigation position in a specific time instant, ii) the orientation of the vessel, i.e. heading, should be arranged in a way that can create a proper course speed vector, and iii) the course speed vector should be tangent to the navigation trajectory. In general, ocean-going vessels may have difficulty satisfying all those conditions, simultaneously. The main reasons for such difficulties would relate to environmental conditions, i.e. due to various wind and wave conditions, ocean-going vessels may have difficulties in following given optimal trajectories. Since most vessels can be considered underactuated systems, harsh ocean environmental conditions can create additional difficulties in following optimal navigation trajectories in collision avoidance-type situations.
Hence, the lack of optimality in such navigation trajectories may lead to possible collision situations in some situations. Therefore, a slight change of course or speed conditions of one vessel to eliminate possible vessel encounter situations should be considered a practical solution in ship navigation. When the collision risk has been eliminated in the same situation, then both vessels should keep their course speed vectors as constant values until the trajectory intersection point is passed. That approach can eliminate additional possible ship close encounter situations between the same vessels, due to any changes in the course speed vectors. Such situations can be further complicated in future vessels, especially when such vessels are under intelligent algorithms for autonomous navigation. Therefore, required decision support systems to overcome such situations should be considered and that should be based on a proper collision risk estimation methodology.
The main concept of autonomous navigation started with driverless vehicles in the automotive industry and that can also be considered as the birthplace of autonomy in navigation. It is noted that the autonomy in navigation should satisfy several selected criteria and that can be summarized as (Wooldridge, and Jennings, 1995) : i) Autonomy, ii) Social ability, iii)Tech-Interaction, vi) Reactivity, & v) Pro-activeness. Autonomy relates to a situation, without any direct or indirect inference from humans, a system should operate with its own decisions and actions as internal states while following appropriate regulatory frameworks. Social ability represents a situation, where a system has an appropriate communication language, and that should be used the system to interact with other systems and relevant parties (i.e., including humans).
Tech-interactions represent a situation, where a system should interact with relevant technologies including decision support type features, and understand and process the information that is obtained from the same relevant technologies as required for decision marking. Reactivity relates to a situation should not only interact with its environment but also respond to the respective time-varying environmental changes and challenges. Pro-activeness is related to a situation, where a system is in the respective environment, and takes appropriate initiatives to satisfy goal-oriented behavior, i.e., to satisfy its navigation objectives. The main reason for this discussion is to define the main concept of autonomy in transportation systems. One should note that auto-pilot systems have been used by various transportation systems and those function under the supervision of human drivers or navigators. It is also noted that a majority of autonomous navigation systems experiments are somewhat based on well-functioning auto-pilot systems. Therefore, a clear separation between well-functioning auto-pilot systems and autonomous navigation should be established to support the technology development in future transportation systems. One should note that similar concepts can be adapted for autonomous navigation in the shipping industry.
Autonomy in Shipping
Autopilot systems have been used extensively in ocean-going vessels under human navigators. Therefore, advanced autopilot systems used by the shipping industry are extensively integrated with control systems on propeller, rudder, and thruster actuation. One should note that these systems are supported by limited feedback information of vessel position, speed, heading, course, and rudder-propeller-thruster conditions. Such systems consist of the main features of heading control, course control, speed control, path-following, way-point navigation, and dynamic positioning type features. These advanced features can often be claimed as autonomous ship navigation systems, even though they may not completely satisfy the selected criteria that were discussed before.
As mentioned in the previous step, the autonomy in shipping should be formulated beyond these advanced autopilot systems and the criteria that have been discussed in the previous section should be facilitated with the same systems. When it comes to autonomous shipping these vessels should be facilitated with additional decision support systems that consist of various features to support ship navigation. That includes the decision support systems for ship collision avoidance, especially under system-level ship navigation or under digital navigators (Perera, 2020) in future vessels. Therefore, appropriate decision support systems for ship collision avoidance should be developed and that can support both manned vessels as well as autonomous vessels (Kim et al., 2022).
SHIP COLLISION AVOIDANCE
Decision Support Systems
Decision support systems in ship collision avoidance applications can improve the safety and integrity of ship navigation in the present and future times, i.e. including autonomous shipping. Collision avoidance in the ocean is somewhat a simpler problem compared to land transportation. In land transport systems, the respective vehicles are somewhat limited to given road transport systems, therefore the navigation space is limited, as discussed previously. On the other hand, ocean navigation consists of a wider area that can be utilized by vessels to avoid possible close encounter situations, as discussed previously. Therefore, this advantage should be utilized by ocean-going vessels to avoid possible ship collision situations and that has been categorized as a simple problem in ship collision avoidance in this study.
Even though ship collision avoidance can be considered as a simpler problem, the complexity in such situations can come from the collision risk estimation methodology. Therefore, the respective decision support systems in ship collision avoidance should consider such complexities in ship navigation and that is the main scope of this study. Hence, the overall collision risk estimation methodology should consist of two levels. The collision risk estimation at local and global levels. In a global level, the associated risk in ship close encounter situations should be considered and that can be done by considering AIS data sets, where the respective vessels in long distances can be identified with respect to the possibility of having close encounter situations (Murray and Perera, 2021). Therefore, adequate collision avoidance decisions can be taken by human navigators even before close ship encounter situations. The applicability of COLREGs can even not be applicable in such long distances, therefore the respective vessels can take appropriate actions.
However, in some situations, these vessels can result in close encounter situations, where more advanced localized predictors to evaluate the respective close ship encounter situations should be considered. Therefore, the respective ship collision risk on a more localized scale should be considered in such situations (Perera, 2017). Appropriate techniques to facilitate a localized collision risk methodology in shipping by considering complex vessel encountering situations are the main scope of this study. Those techniques are discussed in the following sections.
These global and local collision risk prediction tools can eventually be integrated to develop a comprehensive collision risk estimation methodology in ocean navigation. One should note that there are many ship collision avoidance studies that have been done without such a proper collision risk estimation methodology. Therefore, the validity of ship collision avoidance without a comprehensive collision risk estimation methodology is somewhat questionable. That can result in situations, where collision avoidance decisions/actions have been taken by vessels even without any collision risk.
Collision Risk Estimation
The definition of a ship collision situation can play an important role in ship collision risk estimation and that concept is further discussed in this section. Hence, a two-vessel encounter situation is considered in this section and that can be extrapolated into a multi-vessel encounter situation, as required. A two-vessel encounter situation is used towards estimating the respective collision risk between the two vessels. One should note that this vessel encounter situation can be considered as a somewhat simplified situation and that can be utilized to explain the complexities in assessing possible collision risk among vessels in a later stage.
As the first step of developing a proper methodology to detect possible ship close encounter situations, the question of what is a ship collision situation should be addressed. That can be defined by considering a two-vessel encounter situation and presented in Fig. 2. Let’s assume that two vessels are navigating in an ocean environment, one vessel is denoted as the ‘own vessel’ and the other vessel is denoted as the ‘target vessel’. One should note that the heading and course speed vectors of these vessels are in the same direction, for the simplicity in this situation. On the other hand, having two different directions in the heading and course speed vectors can produce complex ship maneuvers that will be discussed in a later part of this study. That can introduce additional complexity to the ship collision risk detection methodology.
The figure represents that both vessels have navigation trajectories that are straight lines due to the same reasons, by assuming calm environmental conditions. The course speed vectors are tangent to the respective trajectories. Furthermore, two vessels are heading toward a trajectory intersection point (see Fig. 2). That situation can be categorized as a possible ship close encounter situation that may lead to a collision situation. However, a trajectory intersection point may not be a close encounter situation, in some situations. Because these vessels can navigate through the same intersection point at different time periods, without any possibility of having close encounter situations. In such a situation, both vessels should keep the initial course speed vectors to avoid any close counter situations in accordance with the COLREGs. On the other hand, a change of course speed vectors can lead to a possible close encounter or collision situation, resulting in ship collisions. Therefore, continuous monitoring of ship close encounter situations is required to eliminate such situations and that can be considered as a continuous collision risk monitoring approach. That should be based on several parameters that relate to a close ship encounter situation and that are discussed in the following sections.
The first parameter that relates to a close ship encounter situation can be defined as the closest point of approach (CPA). That represents the situation where the distance between two vessels is minimal at a specific navigation instant (see Fig. 2). The distance to CPA (DCPA) defined that second parameter can be a measure of the collision risk associated with the close encounter situation of two vessels. The time to the CPA (TCPA) defined as that third parameter can be a measure of the respective time associated with the close encounter situation of two vessels. One should note that these parameters can be used to evaluate the risk of possible collision situations and that can be done by considering predicted future trajectories of ocean-going vessels as presented in the same figure. However, the calculations of the same parameters for collision risk estimation from the actual navigation conditions can be a challenge, therefore the relative navigation conditions can be considered and that will be discussed in the following sections. On the other hand, ship navigators use a somewhat simplified methodology to predict a possible ship collision situation. That method consists of the relative bearing of one vessel with respect to another vessel and that represents the angle between the true north and the other vessel’s position. Continuous monitoring of the relative bearing can be used to predict a possible collision situation, i.e. those two vessels are moving toward a close encounter situation, where the respective vessel course speed vectors and distances are in a good combination for such a situation. The relative bearing concept can be considered as the navigator’s observations of the relative motions of one vessel with respect to another vessel. Therefore, this similar concept can be adopted for evaluating the respective collision risk between vessels, where the calculations of the parameters for collision risk estimation can be simplified.
Hence, the relative motions of the own vessel with respect to the target vessel can be used to detect possible ship-close encounter situations, that may lead to ship collision situations. The relative motions of the target vessel with respect to the own vessel can be calculated by applying an equal & opposite course speed vector of the own vessel to both vessels. Therefore, the course speed vectors of both vessels will change. One should note that the course speed vector of the own vessel will have a zero magnitude, therefore that vessel comes to a standstill under such relative motion conditions. However, the target vessel will have a relative course speed vector and that creates the relative navigation trajectory of the target vessel (see Fig. 3). That can be utilized to identify possible vessel close encounter situations. The relative CPA is a situation where the distance between the standstill vessel (i.e. own vessel) and the relative navigation trajectory of the target vessel is minimal at a specific navigation instant. Hence, the calculations of the parameters for collision risk estimation can be simplified by this situation.
Collision Situation Definition
The relative vessel motions discussed in the previous section can be used to make the definition of a ship collision situation. Hence, that can be utilized to evaluate the risk of ship collisions. The definition should consist of the main concept of vessel domain, where the vessel domain of the own vessel is presented in Fig. 3. The size and shape of the vessel domain may relate to the dimensionality of the respective vessel particulars and speed conditions. When the relative DCAP is inside (or violates) the vessel domain that can be categorized as a ship collision situation or near miss situation. Hence, the same concept can be utilized for estimating the respective collision risk among ocean-going vessels. However, this ship collision definition has been implemented for a somewhat simplified ship navigation situation in this figure. In general, ship navigation situations can be more complex than that and vessels may not move in straight lines.
COLLISION RISK COMPLEXITY
Complex Ship Manoeuvres
The ship maneuvers that are discussed in the previous section are somewhat simplified situations. However, ocean-going vessels may not move in straight-line trajectories and their heading and course speed vectors can have different directions due to various complex environmental conditions and vessel under actuation situations. That can result in complex relative navigation trajectories. Therefore, predicting such complex navigation trajectories can play an important role in detecting possible close vessel encounter situations, that may lead to near miss or collision situations. Since the collision avoidance rules and regulations, i.e. the COLREGs, are implemented with respect to the heading vector, the interactions between the heading and course speed vectors can play an important role, which that can even create some confusion due to navigation complexities.
A two-vessel complex maneuvering situation is presented in Fig. 4. As presented in the figure, the own and target vessels are moving in circular or parabolic type trajectories and that can be considered as somewhat realistic ship motion conditions under various environmental and vessel under actuation conditions, where the heading and course speed vectors can have two separate directions. However, the parameters derived in the previous situation, i.e. trajectory intersection point, CPA, DCPA and TCPA can also be applicable under this situation.
Both vessels have navigation trajectories, and the course speed vectors are tangent to the respective predicted navigation trajectories. While two vessels are heading towards a trajectory intersection point, that situation can be categorized as a possible ship close encounter situation that may lead to a collision situation, similarly. The CPA, where the distance between two vessels is minimal at a specific navigation instant, and the DCPA can be a measure of the collision risk associated with two vessels. However, calculating the DCPA may not be an easy task in such navigation situations due to the complex trajectories. The TCPA can be a measure of the respective time associated with the close encounter situation of two vessels. Continuous monitoring of the relative bearing vector can also be used to predict a close encounter situation, where the respective vessel course speed vectors and distances are in a good combination for such a situation.
The relative bearing vector among those two vessels can be an important tool under complex ship maneuvering conditions and that is presented in Fig. 5. Since these vessels are not moving in straight lines, this may be the only simple methodology that can be utilized for estimating a possible collision situation between these vessels. One should note that even though these vessels are changing their heading and course speed vectors, the respective collision risk can also be observed by this somewhat simpler methodology in ship navigation. This relative bearing concept can be considered as the navigator’s observations of relative motions of the own vessel with respect to the target vessel in this situation also. The course speed vector of the own vessel will have a zero magnitude, therefore that vessel comes to a standstill. The relative course speed vector of the target vessel will change the magnitude and direction. The relative course speed vector of the target vessel creates the target vessel’s relative navigation trajectory. That can be utilized to identify possible vessel encounter situations and such a situation is presented in Fig. 6.
The relative CPA is the situation where the distance between the standstill vessel (i.e. own vessel) and the relative navigation trajectory of the target vessel is minimal at a specific navigation instant. When the relative DCAP is inside (or violates) the vessel domain that can be categorized as a near miss situation or ship collision situation. The navigation trajectories may consist of complex ship maneuvers and varying weather conditions can be the reasons, therefore adequate collision risk estimation methodology, as well as continuous monitoring of the same, should be considered.
Advanced ship predictor-like technologies that are discussed previously can be developed on these tools and can play an important role under complex ship maneuvers. Such technologies can also be a part of future decision support systems that can be used for not only manned vessels but also autonomous vessels (Perera and Batalden, 2019).
The main concepts that relate to ship collision avoidance are discussed in this study with the main objective of guiding the research community toward developing more realistic applications in maritime transportation. Unfortunately, a majority of research studies on ship collision avoidance are extensively based on searching optimal algorithms without a proper collision risk estimation methodology. This study classifies ship collision as a simpler problem due to the reason that a slight change of course or speed action of one vessel can completely eliminate a possible close encounter situation of the respective vessels. Therefore, extensive utilization of computational resources to calculate optimal algorithms in ship collision avoidance can be eliminated due to practical reasons.
It is discussed that the navigation area for ocean-going vessels is somewhat wider (see possible crossing area in the right figure of Fig. 1). That can introduce infinite intersection points for both vessel trajectories. This condition supports ship collision avoidance as a simpler problem since there are an infinite number of interaction points, i.e. or possible solutions, towards ship collision avoidance. Therefore, many slight changes of course or speed change conditions can be taken by either vessel.
Even though ship collision avoidance can be considered a simpler problem, the complexity in such situations can come from collision risk estimation. Therefore, advanced ship collision risk detection methodology should be developed and that should be a part of decision support systems of ship collision avoidance. These situations are extensively discussed in this study and presented in many graphical formats to illustrate the fundamental conditions in realistic ship navigation. These decision support systems can be a part of future vessels, where autonomous navigation-type frameworks will be utilized. Hence, the autonomy in ship navigation that should satisfy several selected criteria to make realistic ocean navigation has also been discussed in this study.
The development of such tools both in global and local scales is further discussed in this study, with the main emphasis on the local scale ship collision risk estimation. This is based on the relative navigation trajectory of one vessel with respect to the vessel domain of another vessel. It is expected that such navigation tools should support future manned and autonomous vessels to improve the safety and integrity of future ship navigation. Therefore, the associated risk in quantifying a ship collision should be a part of collision avoidance situations in ship navigation.