ADAPTIVE COLLISION AVOIDANCE ADVISORY SYSTEM FIELD OF THE INVENTION
The present invention is directed towards a system to assist in the collision avoidance of vessels. In particular, the present invention is directed towards implementing an adaptive knowledge-based inference engine to provide advisory information to vessels for collision avoidance. BACKGROUND OF THE INVENTION
Habour management authorities are very conscious of the need to monitor and detect collision situations. This is particularly relevant with the ever- increasing transport by sea of hazardous chemicals and nuclear related products. Any maritime conflict involving these sensitive vessels can cause tremendous environmental and economical consequences. The adverse effects of such maritime conflicts, in the form of pollution and/or personal injury may be particularly imminent when these collisions involve a large passenger cruise liner or tanker carrying oil or chemicals.
To reduce the potential for conflict, knowledge-based systems have been developed based on the International Regulations for Preventing Collision at Sea (COLREGs) as supported by the International Maritime Organisation (IMO). These currently available systems apply rules in the knowledge base from the perspective of an own ship in relation to another target ship. As such, they are not adequate in resolving conflict involving multiple vessels. For example, the MANTIS system as presented at the 12th Ship Control System Symposium, incorporates an Automatic Collision Avoidance Advisory Service (ACAAS) based on this approach, but again fails to address how complex multi-ship encountering situations can be dealt with. This limited situation assessment based on own ship perspective is not able to cater for more complex real world environments with dynamic happenings.
There is therefore a need for an improved system that provides sound collision avoidance advisory for conflict involving more than two vessels, and is able to adapt to complex situations.
OBJECT OF THE INVENTION
It is an object of the present invention to provide a system that seeks to avoid potential conflict between vessels, and is capable of handling more than two vessels by adapting to the situation. SUMMARY OF THE INVENTION
With the above object in mind the present invention provides an adaptive collision avoidance advisory system for vessels in complex situations unlike current adaptive collision avoidance systems which cater only for adaptive modeling of vessels
The input to the preferred system is a global sea situation picture with all the potential conflicts. The system will detect and classify each conflict into a number of possible scenarios, namely head-on, crossing, or overtaking. Based on each classification, the system will derive the action to be taken by the necessary vessel(s). This constitutes one possible resolution. The system can then assess the different combinations of all possible resolutions and present the recommended combinations with rankings to provide an indication of how good the recommendations are. The system is able to adapt to influence from external environment, including human experts, to adjust the rankings given to the recommendations.
In one aspect the present invention provides a collision avoidance system for three or more vessels including: a path predicting engine to predict the future position of each said vessel; a conflict detection engine to determine whether any said vessels are in conflict based on the predicted path determined by said path predicting engine; and a conflict resolution engine to determine a course of action for any said vessels determined to be in conflict by said conflict detection engine.
The path predicting engine may include a ship maneuvering module, and may also predict the position of vessels based on submitted planned routes which may be stored in a planned route database. The path predicting engine may also take advantage of a sea situation picture.
The conflict detection engine may categorise the potential conflict into levels of seriousness, and the conflict resolution engine may use a scheduling engine for low level conflicts. BRIEF DESCRIPTION OF THE DRAWINGS
Figure 1 is a block diagram for the adaptive collision avoidance advisory system.
Figure 2 shows the pair-wise collision detection algorithm.
Figure 3 shows an example of the sweeping algorithm.
Figure 4 is a block diagram showing the different component of the adaptive knowledge-based inference engine.
Figure 5 is a typical manoeuvres chart used to classify conflicts into the various scenarios.
Figure 6 depicts the set of rules used to classify conflicts.
Figure 7 shows some examples for classifying conflicts.
Figure 8 shows an example of the route generation module.
Figure 9 shows the selection operation of the Genetic Algorithm (GA).
Figure 10 shows the crossover operation of the GA.
Figure 11 shows the mutation operation of the GA. DESCRIPTION OF PREFERRED EMBODIMENT
The present invention provides a system and method for implementing adaptive collision avoidance advisory system for vessels in complex situations using a hybrid of artificially intelligent techniques. A complex situation may involve three or more vessels. Referring to figure 1 the five main components in the preferred arrangement are the path predicting engine 3, the conflict detection engine 5, the conflict resolution engine 7, the route generation module 9, and the scheduling engine 8. The planned route database 1 , the ship manoeuvring model 2, and the sea situation picture (SSP) 6 form the preferred inputs to the system, while the display module 4 presents the output from the system. The adaptive collision avoidance advisory system of the present invention ideally provides two levels of alerts and advices, for multiple vessels in potential conflicting situation, one is for early warning with a further look ahead time (level 1), and the other is for close quarters within the nearer future time (level 2). It is noted that existing
collision avoidance systems are only designed for a single pair of vessels, and are not able to handle three or more vessels.
The path predicting engine 3 can utilize a ship manoeuvring model to predict the future positions of vessels, based on submitted planned routes (if available) as well as the current position, course, and speed of vessels in the SSP. The SSP can be formulated from input sensors, such as radars and an automatic identification system (AIS), which would provide position, course, and speed of all vessels detected. Alternatively these values may be inputted from another system that calculates position, course and speed based on sensor detection. The ship manoeuvring model is recommended to provide high fidelity prediction based on vessel hydrodynamics and environmental data, such as weather, sea state, sea lanes, boarding ground, etc. This can be used for level 2 alerts. The ship paths may be predicted based on the vessels' current speed and their submitted planned routes if available, otherwise the predicted path may be assumed to be straight ahead based on the vessels' current course. Typically, a vessel's predicted route consists of a start way point, an end way point and a number of other intermediate way points. These way points can then be joined by straight lines to form a representation of a vessel's route. Level 1 alerts will be given to vessels that have reported their intended sail route to the system. These are kept in the planned route database 1. Likewise, these planned routes are represented as way-points.
Based on the predicted paths from the path predicting engine 3, the conflict detection engine 5 will determine conflicting situations, and ideally prioritise these situations according to some predefined criteria, such as time to potential conflict. A pair-wise collision detection algorithm (see figure 2), based on the closest point of approach (CPA) and danger circle 10 computation, should form the core of this engine. The pair-wise collision detection algorithm is used between two way-points for a pair of vessels. If the CPA between two vessels is less than a certain well-established or predefined safety distance (i.e. within the danger circle), the algorithm should flag such a situation to be a potential collision. The collision algorithm works by grouping pairs of vessels for computation. If there are 3 vessels, for eg. Vessels A, B and C, then there will be 3 possible ways of grouping pairs of vessels given by AB, BC and AC. For a
multiple global conflict scenario, this may be primarily handled by the Genetic Algorithm and the Artificial Neural Network which is described later
In figure 2, the dotted lines with larger dots represent portions of the predicted path of two vessels. To compute the CPA, the time of closest point of approach, timeCPA, is first computed. The positions and velocities of the two vessels at the way-points just before the collision detection are given by (x1 , y1 ), (vx1 , vy1 ), (x2, y2), (vx2, vy2) respectively. The closest point of approach on predicted path for both vessels, CPA1 and CPA2, can then be computed respectively.
For example, given the velocity and displacement of a pair of vessels , their relative displacement can be computed based on kinematic equations. The closest point of approach can then be computed by minimizing the magnitude of the relative displacement. The minimization can be done using a variety of methods, for example in the preferred arrangement of the present invention differentiation techniques are employed.
The time for both vessels to reach their CPAs could be computed using the following:
- (dx • vx + dy ■ vy) tιmeCPA = — — — where dx = JC, - , vx + v dy = yλ -y2
vx = vx, - vx2
vy = vyl -vy2
CPA-[ and CPA2 can then be computed by extrapolating the respective current position to this time:
The mid-point between CPA1 and CPA2, CPA, will be the CPA for the two vessels. The danger circle 10, of diameter centered about the CPA, is used to
determine if the two vessels are in potential conflict with each other. The diameter of the danger circle can be computed by: speed
seP™ -
l speed ≤ pd^ danger circle ■
seP
• !
■ speed ≥ spd
mm ( speed -spd
τ sep
min + (sep
τnm -sep
min) otherwise
In Figure 2, I is the higher value of the diameters for the minimum turning circle of both vessels; sepmin and sepmax denotes the safety separations between the vessels necessary given the threshold speeds of spdmin and spdmax. The typical values for sepmin, sepmax, spdmin, and spdmax are 2, 4, 4, and 12 respectively. These are rule-of-thumb given by experienced mariners.
The time to CPA may be computed by: dx = xl -x2, dy = yλ -y2 vx = vx, -vx2, vy = vy, -vy2 _ - (dx • vx + dy • vy) time, CPA vx 2 + , vy 2
To reduce the number of pair-wise collision detections required, the engine could incorporate a sweeping algorithm, which generates areas swept out by vessels based on their motions. If there is a collision between two vessels, the areas swept by them must intersect one another. In other words if a vessel is represented as a rectangular bounding box, an area will be swept out by each vessel when they move. By extrapolating both the areas swept out by the pair of vessels without reference to the time factor, it can be determined whether the two extrapolated areas do intersect. If they do, then the pair of vessels may possibly be in potential collision and will be subjected to the pair-wise collision detection test. Otherwise, collision is impossible (see figure 3 for an example of the algorithm). Note that since temporal (or time) dimension is ignored during sweeping, two vessels with intersecting areas do not necessarily become involved in a collision.
The conflict resolution engine 5 employs a hybrid of artificially intelligent techniques to come up with advisory information for resolving conflicts. Figure 4 shows the preferred components that are used in this engine . Based on the collisions detected in the conflict detection engine, the rule-based inference engine will categorise them into level 1 and 2 conflicts. For level 1 conflicts (early warning), the rule-based inference engine will determine if such conflicts could be grouped to form congested areas which are areas of particularly high sea traffic density, that is, the number of vessels per unit area of sea coverage exceeds a certain well-established or predefined threshold value. If so, the scheduling engine will be invoked to resolve these conflicts. Preferably for level 2 conflicts (close quarters) and conflicts not constituting congestion, the rule-based inference engine uses the manoeuvres chart (figure 5) and the conflict classification table (figure 6), which are derived from the COLREGs, to infer the resolutions for each pair of conflicts. As shown, the manoeuvres chart divides a ship into different sectors and depending on which sector the intersecting line between the centre of a reference ship and that of another ship, the general instructions in the tabular chart should be referenced and used. The classification table summarizes all the possible conflict scenarios with recommended actions and action-ships, in accordance to COLREGs, in terms of the conflicting sectors. Figure 7 shows some examples on how to use the manoeuvres chart and conflict classification table. For example, when ship B is in sector 1 of ship A while A is in sector 8 of Ship B, the two ships are in a conflict classified as head-on. The manoeuvres chart provides resolutions for the vessel that need to take action, known as the action ship (the vessel that need not take action is known as the Right-Of-Way or ROW ship).
Every resolution inferred from the rule-base inference engine is based on sound reasoning guidelines provided by a domain expert, such as COLREGs or experienced mariners. The solutions space includes different combination of these inferred resolutions (or knowledge). The Genetic Algorithm (GA) is an algorithm for searching a local optimal solution given a solution space is used to explore the solution space while the Artificial Neural Network (ANN) is employed to rank each solution. The ANN could be based on a back propagation model with four input nodes, such as the number of potential conflicts, time to first
potential conflict, the computed delay from expected time of arrival (ETA) and expert user's judgement. The ANN provides the means for the system to adapt by "learning" the best solution with inputs from the external environment. To achieve real-time learning, each feedback from the environment is ideally fed through the ANN to allow the adjustment of the weights. The terminating condition for this single-case learning can be capped to, say 100 cycles, of the updating algorithm in the ANN. This is in addition to the threshold terminating condition in a traditional ANN so as to achieve real-timeliness of the system. One problem with this is that the system tends to forget what it has learned. To avoid this, historical knowledge can be kept and a data mining algorithm used to extract good examples to continuously refresh the memory of the ANN. The data mining algorithm can also be used to discover new rules, if any, for the rule-based inference engine such as for example by a Decision Tree Learning algorithm, ID3).
The present invention seeks to provide a method and system for performing sound reasoning and learning, similar to a human cognitive process, to rank solutions encompassing this inferred knowledge.
The scheduling engine 8 will allocate time based on priorities that will be explained later, for vessels to transit the congested area, requesting them to slow down or even speed up, based on the manoeuvrability of the vessel input by the conflict resolution engine. Hence, new predicted paths returned to the conflict resolution engine 7 are the same path except that the time for some of the way- points are modified. Scheduling will be done based on certain priorities, such as the vessels' sizes, the loads they carried, risk in manoeuvrability, etc.
The route generation module 9 will translate these actions by modifying the original predicted path of the action vessel by changing relevant way-points and timing at each point to follow a recommended action while avoiding any danger circle. The new predicted paths are fed-back to the conflict resolution engine 7 for further analysis. Figure 8 shows an example of how this is done. Essentially, the route generation module will attempt to avoid the danger circle computed in the conflict detection engine taking into account the manoeuvrability of the vessel input by the conflict resolution engine.
Referring to figure 4, the rule-based inference engine starts the reasoning process with facts 11, i.e. SSP and conflicts detected, inputted to the working memory of the system. This reasoning process will create new knowledge in the working memory, based on a set of predefined rules such as the COLREGs, derived from human expertise. The inferred knowledge can be inserted back into the working memory and could invoke further reasoning to infer more knowledge. The result of this process is a set of inferred knowledge that is to be used to formulate the solution.
Each piece of inferred knowledge that could be used to formulate the solution is denoted as a gene. A possible combination of these genes will give rise to a chromosome, which represents a possible solution. For example, in a multiple ship scenario, solution to a pair among the ships concerned is a gene, a chromosome is a combination of pair-wise resolutions. The chromosome is a possible solution to the multi-ship conflict. The solution space may include different combinations, and this will provide the population for the GA. Each solution may be ranked by a fitness function as described earlier, which can be computed by using the ANN. Solutions can evolve by using the basic operations, namely selection, crossover, and mutation, as shown in figures 9 to 11.. Figure 9 shows the selection operation of the GA in the preferred system. The GA starts with an initial set (or population) of solutions (or chromosomes). Each generation of the GA will create a new population by selecting "good" chromosomes from the previous population, based on the fitness function. To avoid being "caught" in a local optimal solution, the crossover (figure 10) and mutation (figure 11) operations are applied on the chromosomes in the new population. The crossover operation involves 2 chromosomes, and choosing a crossover point; the genes after the crossover point in the 2 chromosomes are exchanged, in the hope of deriving a better solution (indicative by the fitness function). The mutation operation involves just 1 chromosome. This will "change" (or flip) some genes in the chromosome to see if it would be a better solution.
A number of ranked solutions can then be presented to the user via the display module. The best solution could then be selected from the influence of the environment, which could be the advice from a human expert. The ANN is then trained based on the best solution, by rewarding the best solution (i.e.
increasing the desired output value) and penalising the rest of the solutions (decreasing the desired output values). In this way the system is able to adapt whilst in operation.
Another unique feature of the preferred adaptive knowledge-based inference engine is the use of an offline data mining algorithm 15, whereby knowledge such as expert user's inputs and weights of the ANN are mined, to enable the ANN to retrain as and when necessary. This requires the system to keep track of historical knowledge 18 as shown in figure 4. The data mining algorithm 15 may be used to automate the detection of outliers in the historical knowledge base.
The adaptive knowledge-based inference engine is proposed to provide a holistic view based on most (if not all) pairs of conflicting vessels so as to evaluate the various avoidance recommendations and come up with the collision avoidance solution that best resolves all the conflicts in the current environment that is, including multiple ship scenarios. The present invention provides a method for knowledge-based system to adapt to the changing environment.
The present invention seeks to provide a method and system for performing sound reasoning and learning, similar to a human cognitive process, to rank solutions encompassing this inferred knowledge. Whilst the method and system of the present invention has been summarised and explained by illustrative examples, it will be appreciated by those skilled in the art that many widely varying embodiments and applications are within the teaching and scope of the present invention, and that the examples presented herein are by way of illustration only and should not be construed as limiting the scope of this invention.