533

1 INTRODUCTION

In order to solve the problem of collision and aground

accidents caused by human factors, it is a feasible

scheme to develop the assistant decision-making

system of ship collision avoidance to reduce the

effects of man-made factors as much as possible. In

view of the collision avoidance problem, which can

not be reappeared, the intelligent collision avoidance

decision-making system should have advanced

humanoid intelligence, that is, surpassing the human

ability to solve complex collision avoidance problems.

The assistant decision-making system of collision

avoidance should be able to provide both safe and

economical (scientific and reasonable) decision

scheme, and it must follow the rules, seafarers'

common practice and good seaman ship, and have the

ability of quantitative analysis and calculation at the

same time.Therefore, the machine must have the

ability of on-line automatic learning based on the

original knowledge, so that the system has the ability

of automatic perception, cognition, analysis and

reasoning, and can make wise choices in complex and

even dangerous environments. Make a successful

decision-making and implementation plan to ensure

the safety of ships to avoid risks. Nowadays, the deep

learning method based on big data needs a large

number of learning samples. However, due to the

non-reappearance and high uncertainty of the marine

encounter situation and the restriction of the

"International Regulations for Preventing Collisions at

Sea" （Hereinafter referred to as COLREGS）, it is

difficult to effectively solve the problem of collision

avoidance decision-making of ships at sea. Other

machine learning methods, such as intensive

learning[1~2] and so on, at present, still in the

exploration stage, no examples of application have

been found.

The Machine Learning Method of PIDVCA

L. Li, X. Wang & G. Chen

Jimei University, Xiamen, China

ABSTRACT: Building a dynamic collision knowledge base of self-learning is one of the core contents of

implementing "personified intelligence" in Personifying Intelligent Decision-

making for Vessel Collision

Avoidance (short for PIDVCA). In the paper, the machine learning method of PIDVCA combined with offline

artificial learning and online machine learning is proposed. The static collision avoidance knowledge is

acquired through offline artificial learning, and the isomeric knowledge representation integration method with

process knowledge as the carrier is established, and the Dynamic collision avoidance knowledge is acquired

through online machine learning guided by inference engine. A large number of simulation results show that

the dynamic collision avoidance knowledge base constructed by machine learning can achieve the effect of

anthropomorphic intelligent collision avoidance. It is verified by examples that the machine learning method of

PIDVCA can realize target perception, target cognition and finally obtain an effective collision avoidance

decision-making.

http://www.transnav.eu

the

International Journal

on Marine Navigation

and Safety of Sea Transportation

Volume 14

Number 3

September 2020

DOI:

10.12716/1001.14.03.02

534

After more than 20 years of research, the R & D

team of Jimei University combined artificial

intelligence ideal with knowledge in the field of

navigation and solved the expression of existing

knowledge of collision avoidance based on artificial

off-line learning. Based on the perceptual system, the

dynamic and static data of the own ship and the

target ship are obtained, and the on-line machine

automatic learning is carried out with a series of

modles and agrithems of the Personifying Intelligent

Decision-making for Vessel Collision Avoidance

(short for PIDVCA) in the static knowledge library.

Realize machine automatic perception, cognition,

analysis and decision-making. A lot of Simulation

results show that PIDVCA machine learning method

achieves the desired results.

2 BRIEF INTRODUCTION OF PIDVCA PRINCIPLE

2.1 The goal and connotation of PIDVCA

In the final analysis, the goal of studying “PIDVCA”

is to solve the automation problem of "intelligent

collision avoidance decision making for ships".

Specifically, it is to apply the "COLREGS" and

knowledge in the field of ship collision avoidance.

With the help of the environment acquired by the

sensors and the qualitative analysis of the experts'

experience knowledge, the deck officer practical

experience knowledge and artificial intelligence

technology are used to carry out the automatic

reasoning, quantitative calculation and evaluation,

with the help of the environment obtained by the field

sensors. Realize the automatic generation and

optimization of PIDVCA scheme.

The connotation of "PIDVCA" is to realize the

automatic intersecting relationship between the ship

and the ship according to the object mark (target ship

and obstructed object). Imitating the analysis and

judgment of the surrounding environment and

dangerous situation and the thinking of the collision

avoidance decision made by the experienced deck

officer (collision prevention expert), the collision

avoidance decision, which is both safe and

economical, can be brought forward by the collision

avoidance expert automatically. If necessary, a simple

coordinated collision avoidance process between two

ships can be simulated; When the decision is accepted

and implemented by the deck officer, the ship can be

automatically informed of the ship's intention.

2.2 PIDVCA realization principle

As shown in figure 1, the functions of each

component of PIDVCA program model are as follows:

Figure 1. PIDVCA program model

1 The prediction of ship dynamic risk and the effect

of collision avoidance decision is an evaluation

system for the decision behavior of collision

avoidance. It not only relies on the basic model of

dynamic collision avoidance knowledge base and

related algorithms, but also uses online learning to

generate dynamic risk threshold in real time as the

criterion of machine perception, cognition and

decision-making process.

2 Dynamic collision avoidance knowledge base

consists of database, model base, rule base,

PIDVCA algorithm base and dynamic collision

avoidance information base.

The database is used to represent the factual

knowledge required in the collision avoidance

decision-making process, including from marine

radar, AIS, gyro compass, log, GPS, ECDIS,

Navigation data, meteorological and hydro-logical

information, provided by navigation equipment

such as visibility meters (referred to as

sensors),,basic parameters of the ship and static

data of electronic charts.

3 The model base provides all kinds of collision

avoidance models needed for the quantification of

the concept and decision-making;

The rule base provides the knowledge of the

situation division, the rules of action and the

assignment of the duty to avoid collision in the

form of production rules, as well as the casual

knowledge of the good seaman ship and common

practices;

The PIDVCA algorithm library integrates the

knowledge of database, rule base and model base

in the form of meta-knowledge, and provides a

series of PIDVCA algorithms for machine

perception, cognition and decision making.

The dynamic collision avoidance information base

is used to store and interact the intermediate

dynamic collision avoidance information

generated by the automatic learning process of

machine in the form of array or variable..

4 Inference engine and algorithm Library as an

Organic whole, and the algorithm flow forms the

automatic reasoning mechanism of reasoning

535

machine heuristic rule, analogy (matching), case

and non-monotone reasoning.

5 Machine learning includes knowledge discovery

and approximate reinforcement learning strategy

under the guidance of automatic reasoning

mechanism.

To sum up, the PIDVCA principle takes the model

and algorithm as the main representation form of

machine learning and the basis of prediction and

evaluation system of ship dynamic risk and collision

avoidance decision effect, and adopts process

knowledge representation as the carrier of

heterogeneous knowledge representation integration

method. The automatic inference mechanism is

designed based on the evaluation system of ship

dynamic risk degree and collision avoidance decision

effect prediction and evaluation, the reasoning engine

guides the on-line learning of the machine based on

the original knowledge of the dynamic collision

avoidance knowledge base (the database, rule base

and model base) and the object information (database)

obtained in the field, using the calculation and

reasoning of integrated machine learning strategy,

form the dynamic collision avoidance knowledge

base. Through the PIDVCA program model, the

heterogeneous dynamic collision avoidance

knowledge and inference machine are expressed in

database, rule base, PIDVCA algorithm library and

dynamic collision avoidance information base. The

system of machine learning and prediction and

evaluation of ship dynamic risk degree and collision

avoidance decision effect is integrated into an organic

whole, which realizes the automatic generation,

verification and optimization of target perception,

cognition and PIDVCA implementation scheme.

3 KEY TECHNOLOGIES OF PIDVCA MACHINE

LEARNING

The original knowledge is composed of model base,

rule base, algorithm base and factual knowledge in

database. The construction of PIDVCA algorithm

library is the core technology of machine learning.

Model base and rule base are the basis of constructing

algorithm library. This chapter mainly discusses the

core technology and foundation of PIDVCA machine

learning.

3.1 The Goal and Mechanism of Machine Learning

The title of this paper makes it clear that the goal of

machine learning is to realize the intelligent collision

avoidance decision of ships, that is, to get new

knowledge and new skills to solve arbitrary collision

avoidance problems through on-line self-learning of

machines, and the concrete learning process is the

automatic perception, cognitive and anthropomorphic

collision avoidance decisions of machines.

Online machine learning includes heuristic rule

reasoning, analogical (matching) reasoning, case-

based reasoning and non-monotone reasoning under

the guidance of knowledge discovery and

approximate reinforcement learning strategies.

The approximate reinforcement learning strategy

is a reinforcement learning method based on dynamic

decision making and objective function optimization

for special problem in order to further optimize

collision avoidance decisions on the basis of forming

initial decision making and determining search space.

To process dynamic collision avoidance knowledge in

real time and obtain the optimal collision avoidance

decision implementation scheme, that is, to solve the

problem of environmental awareness of this own ship

(can be called as an agent), how to learn, in order to

determine the optimal collision avoidance decision

scheme which can safely and economically avoid

dangerous target ship, PIDVCA takes the minimum

track deviation after the implementation of collision

avoidance decision as an economic index. The

objective function of dynamic decision optimization is

further established based on the quantitative model of

the rudder timing(Ts), the range of

avoidance(

)and the opportunity

()

T AC

( ) ( ) ( )

min sin

, 0,1,2,...,

rk k

AC A

J s T AC V AC

Sk N

∈

= ××





∀∈ =

(1)

where

( )

TL Ts

−

∆

T∆

is the search step, A is

the behavior set of the ship in the search space

{ }

,T Ts TL

∈

, S is the state set and is the speed of the

ship. Corresponding to each search step can be

obtained, according to which the corresponding ones

can be solved. According to the prediction end time of

all dangerous target ships, the maximum value of (1)

is their maximum value, which is the corresponding

optimal decision in the minimal case.

Machine online learning how to use field

information and original knowledge of collision

prevention experts through automatic reasoning,

learning, optimization to achieve the desired goals?

The key is to get the original knowledge，which they

are a series models and algorithms, the analogical

source of analogical matching reasoning and the case

source of case-based reasoning through off-line

learning. It can be seen that the original knowledge is

the important technical foundation of machine on-line

learning to produce new knowledge. Therefore, the

machine learning of PIDVCA adopts the mechanism

of combination of off-line artificial learning and on-

line machine self-learning, using off-line artificial

learning to realize the formalization of ship collision

avoidance domain knowledge, and to solve the

problem of knowledge representation of collision

avoidance experts. The original knowledge of

collision avoidance is formed in the model base and

algorithm base, which provides the technical

foundation for online machine learning.

Off-line artificial learning is mainly used to

generate analogy source and case source of analogical

matching reasoning and case-based reasoning

learning, and a series models and algorithms to obtain

heuristic information and new knowledge online, to

form static collision avoidance knowledge base. A

heterogeneous knowledge representation with

process knowledge as the carrier is established so that

the machine can acquire new knowledge and new

skills of problem solving in real time based on pre-

built static collision avoidance knowledge.

536

3.2 The Technical basis of Machine Learning

In the previous section, the importance of the

construction and representation of the original

knowledge of collision avoidance to the realization of

on-line self-learning of machines is discussed, in

which the model base is an important foundation of

machine learning, and the PIDVCA algorithm library

is the representation of machine learning. It is also the

core technology of machine learning to construct

dynamic collision avoidance knowledge base.

3.2.1 Modeling of concepts

The modeling of the concept is mainly through the

geometric analytic method of the vector triangle of the

ship's relative velocity (as shown in figure 2). This

paper constructs some important concepts for

quantifying the important concepts in the rules of

collision avoidance, some important concepts

proposed by the PIDVCA method to realize the

"personify intelligence", and various kinds of

PIDVCA mathematical models for obtaining the

required dynamic collision avoidance parameter

information and dynamic avoidance knowledge in

real time.

SDA

Y (Co)

NRML1

RML1

DCPA

NRML1'

RML2

NRML2

RML2'

RML1'

Rp(xr,yr)

A1'

A2'

ACp1(xp1,yp1)

RCp(xc,yc)

ACp2(xp2,yp2)

Figure 2. Geometric sketch for OS being altered

AC,predicted recovery time and its restoring confine point

The important concepts of "COLREGS" include

safety encounter distance, urgent situation distance

and urgent danger distance, and the important

concepts of PIDVCA method such as the best rudder

time, the latest rudder time, the margin of avoidance

and the prediction of the time to resume navigation,

etc. The machine invokes the corresponding model in

the model base through the algorithm flow sequence,

obtains the heuristic guidance information of the

automatic reasoning, and then uses these heuristic

information to guide the machine to learn online and

discover the machine perception in real time. New

knowledge of cognition and problem solving. For

example, according to the target parameter model, the

threshold value of the recent encounter distance and

the safe encounter distance of the target ship is

calculated. According to the comparison of DCPA and

SDA, the potential collision risk between the target

ship and our ship is identified. Then the execution

path of the next algorithm flow is determined. Figure

2 is a geometric diagram for solving the range of

avoidance and predicting the time of resumption.

3.2.1.1 Quantizing model of PIDVCA scheme

PIDVCA scheme includes the opportunity of

initial steering rudder (stand for T

isr), amplitude of

altering course (stand for AC) and predicted restore

point (stand for Rp) or opportunity (stand for Tr) of

collision avoidance.

Tisr is calculated by the same method as the last

opportunity of steering rudder (Tln) .

If the Tisr has been missed, AC computing model

is derived from the speed vector triangle A1

’

B1C1 as

shown in fig.2:

180

)

)][][sin(][

sin][ −













+−×

−=

ConCtnCrnnVt

anCrnAC

(2)

Tr is the predicted time of dangerous TS1 sailing

from ACp1(xp1,yp1) to Rp(x

r,yr), the predicted

restoring point Rp(x

r,yr) is the crossover point of

NRML1'and RML1' as shown fig.2. Solving the

crossover point for RML1 and NRML1′, then Rp (x

r) is:

[ ]

( )

[ ]

( )

[ ]

( )

[ ]

( )

[ ]

( )

[ ]

( )

sin sin tan tan

tan sin

SDA n SDA n

Crn n Cr n Cr n Crn n

SDA n

Cr n Cr n

  

=−−

  

  

= +

(3)

In formula (2), Cr and Crn are not included with

the 0, 90°, 180° and 270°.

Then, the computing model of Tr is:

[ ]

( )

[ ]

( )

[ ]

Xpn x Ypn y

Tr n

Vrn n

−+ −×

(4)

Here, n is the number of the most dangerous TS.

3.2.1.2 PIDVCA scheme’s verifying models under

multi-vessel encountering situation.

The verifying models of PIDVCA scheme are

composed the computing models of predicting TS’s

parameters and restoring confine time (short for Tc)

for new dangerous TS. As shown in Fig. 2, Tc is the

sailing time that new dangerous TS2 from ACp2 (xp2,

yp2) (altering course point) to RCp (x

c, yc) (restoring

confine point), and its computing model method is

the same as the Tr. Firstly solving the crossover point

of RML2’and NRML2, and RCp (x

c, yc) as following:

( ) ( ) ( ) ( )

( ) ( )

][sin

][

][tan

][sin

][

][tan

][

iCrn

iCPAn

iCrn

iCrnnCriCr

iSDA

iCrn

iCPAn













−÷













(5)

537

( ) ( )

][

][][

][

−+−

iVrn

yiypxixp

(6)

In formula (5), Cr and Crn are not included with

the 0, 90°, 180°and 270°.

3.2.2 Experiential knowledge representation of causality

in production rules

Rule is one of the important contents of the

knowledge of collision avoidance experience. The

method of production rule representation can

conveniently express the knowledge of collision

avoidance with causality. The following is an example

of overtaking situation in the form of production rule

set.

If: wx[i]=1∧B>112.5°∧B<247.5°

Then: Tsea[i]=4（4 Indicates that the target vessel I is a

overtaking ship）

If: wx[i]=1∧V0>Vt[i]∧Bo>112.5°∧Bo<247.5° Then: Tsea[i]=1（

“1” Indicates that the target vessel i is a being overtaking

ship）

If: Tsea[i]=1∨Tsea[i]=4 Then: Overtaking situation

If：Visibility_Level=1∧Tsea [i]=1 Then：Tzt[i]=1

If：Visibility_Level=1∧Tsea [i]=4 Then：Tzt[i]=2

B is the azimuth of the target ship relative to the

ship, Bo is the orientation of the ship relative to the

target ship, Tzt is the target property identifier, its

value is 1 indicating that the target ship is a direct

vessel, and the ship is making way for the ship, if its

value is 2, the value is the opposite. Tsea is the target

attribute identifier (target ship encounter attribute),

whose value is 1 , 2 , 3 , 4 , 5 and 6 respectively, the

target ship is overtaken, coming from the front of the

right forward, from the right cross, overtaking the

ship, from the port side, and from the opposite

state.The “wx” stand for identifier of danger for target

ship, its value equal to 1 or 0 respectively represents

the target ship being danger or not. Visibility_Level is

stand for identifier of visibility, its value equal to 1 or

0 respectively represents the insight of one another or

restricted visibility condition.

3.2.3 Algorithm representation of Meta-Knowledge

—

The PIDVCA algorithm Library

The PIDVCA algorithm representation of meta-

knowledge is the core technology of machine

learning. It mainly solves the leading information

source of heuristic rule reasoning , the analogy source

of analogical matching reasoning and the case source

of case-based reasoning in machine learning process.

The PIDVCA algorithm library is described by

flow chart, decision tree, decision table and

mathematical equation. It is used to realize the

combination of the qualitative analysis based on

“COLREGS” the ordinary practice and fine

seamanship of seamen, and quantitative calculation

processing , in order to obtain real-time dynamic

collision avoidance knowledge.

It includes target rendezvous feature(TRF)

recognition algorithm, potential hazard analysis

algorithm, encounter situation recognition algorithm,

risk degree analysis evaluation, decision scheme

quantification, scheme verification and optimization,

prediction and evaluation of collision avoidance effect

and a series of PIDVCA algorithms. As shown in

Table 1, the TRF recognition algorithm is used as an

example source to identify the TRF.

Table 1. The value of TRF recognition

_______________________________________________

TRF Ct-Co Cr velocity relation

_______________________________________________

1 0°~90° 90°~180° Vo

≤

2 270°~360° 180°~270° Vo

≤

3 0°~90° 90°~180° Vo

≥

4 270°~360° 180°~270° Vo>Vt

5 180° 180° Vo

≥

50 0° 180° Vo>Vt

6 0° 0° Vo

≤

60 C

o 180° Vo=0

7 0°~90° 0°~90° Vo<Vt

8 270°~360° 270°~360° Vo<Vt

10 90°~180° 90°~180° Vo

≤

20 180°~270° 180°~270° Vo

≤

30 90°~180° 90°~180° Vo>Vt

40 180°~270° 180°~270° Vo>Vt

70 90° 90° Vo<Vt

80 270° 270° Vo<Vt

_______________________________________________

Figure 3. Flow chart of case source generic algorithm for

heuristic case-based reasoning

Based on the anti-collision methods of the

experienced captain and the chief mate taken in the

typical examples, in the virtual box of figure 3 is the

flow chart of case source generic algorithm for

heuristic case-based reasoning (CBR) algorithm which

simulates the deck officer's correct assessment of

collision risk, avoidance opportunity, amplitude of

avoidance recovery opportunity and L-PAE. The Trr,

Tcc and L-PAE in fig. 3 are separately stand for the

value of max{Tr[i]}, min{Tc[i] } and level of predicted

anti-collision effect.

3.3 On-line machine self-learning steps

According to the information of the ship and the

target ship obtained from the navigation and

environment sensor, the inference engine calls the

corresponding algorithm of the algorithm library in

538

turn, and realizes the automatic perception, cognition

and anthropomorphic collision avoidance decision of

the machine, through the pre-designed machine

automatic reasoning mechanism, and according to the

information of the ship and the target ship obtained

from the navigation and environment sensors. The

specific steps are as follows:

Step 1: the system adopts online heuristic rule

reasoning, then calls the algorithm of solving collision

avoidance parameters of the target ship(TS stands for

target ship) in the PIDVCA algorithm library and the

target rendezvous feature recognition algorithm in

turn, and obtains the motion elements, collision

parameters and the threshold value of collision risk

evaluation. The rules of analogical reasoning are used

to match and identify the features of target

rendezvous and the perception of target dynamic

information is realized.

Step 2: according to the result of targets’

perception on step 1, the system calls in turn the

target ship potential hazard analysis algorithm in the

PIDVCA algorithm library, the encounter attribute

analysis and its situation classification algorithm, and

the recognition algorithm of the collision avoidance

attribute of the ship. The heuristic rule matching of

analogy reasoning is used to automatically judge the

potential danger of the TS, if there is no potentially

dangerous TS, turn step 1, otherwise identify the

encounter attribute, and then determine the own

ship(OS stands for own ship) to give-way vessel or

stand-on vessel and its corresponding avoidance

measures, and then according to the action that the

OS should take, By calling the dynamic risk analysis

algorithm, the initial risk evaluation and the most

dangerous ship are given.

Step 3: according to the result of step 2 initial risk

evaluation, the system calls the algorithm of PIDVCA

scheme generation, and determines the initial

PIDVCA scheme which is composed of avoidance

opportunity, amplitude and forecast time.

Step 4: the system calls the potential hazard

analysis algorithm for predicting TS to verify whether

the initial PIDVCA scheme can clear other

obstructions (dynamic TSs and static obstructions), if

feasible, turn to step 6, otherwise turn to step 5;

Step 5: the system calls the PIDVCA scheme check

and optimization algorithm, implements the check

and of the collision avoidance decision scheme,

generates the PIDVCA scheme. For the unreasonable

scheme of special scene, if necessary, the approximate

reinforcement learning algorithm is used to

implement local optimization and turns to the sixth

step.

Step 6: the evaluation and analysis of collision

avoidance decision based on evaluation system of

ship dynamic hazard degree and prediction of

collision avoidance effect. If the effect of avoidance

belongs to the state of safety or secondary, then

output danger early warning and collision avoidance

scheme and the scheme of avoiding collision is

executed; if it is unsafe, the program of coordination

of immediate danger is executed.

Self-learning

Decision-making

cognition

Feasibility verification of

collision avoidance decision

Output danger early warning and

collision avoidance Scheme

Prediction and Evaluation of the

effect of collision avoidance

Decision-making

Verification and Optimization of

Collision Avoidance Decision

Perception:

recognition of moving state

of target ship and feature of

target rendezvous

Encounter situation recognition

and risk assessment

Generating an initial avoidance

decision scheme

Initialization

Navigation and

environment

information

sensor

Evaluation

system of

ship

dynamic

risk degree

and effect

of collision

avoidance

Dynamic

collision

avoidance

knowledge

base

Potential hazard judgment

Figure 4. Machine online self-learning flow chart

4 SIMULATION VERIFICATION ANALYSIS

4.1 SIHC simulation platform

To validate the algorithm of PIDVCA, we developed a

simulation platform with the name of “the ship

intelligent handling control (SIHCS) platform”, which

is based on the ship maneuver simulating model and

electronic chart display and information system

(ECDIS). The SIHCS platform is constructed by

making good use of the technology of ship handling

simulator. It is composed of one console computer

and three own ship (OS) servers and one target ship

(TS) server. The console, OS and TS servers are linked

by local network. The SIHCS platform possesses all

functions of ship handling simulator. The software of

OS server also has an interface for users. The PIDVCA

algorithm is integrate together with the Self tuning

fuzzy PID autopilot algorithm into the SIHCS

platform in the form of dynamic link library. Thus,

the SIHCS platform is used to carry out simulating

test on the process monitor for vessel automatic

collision avoidance. Figure 5 is a simple simulation

test platform composed of console and target ship

server.

4.2 Simulation example analysis

In this paper, give a typical training example of a

marine radar simulator that meets multiple ships, and

the PIDVCA algorithm is integrated into the target

ship server with the help of the ship intelligent control

simulation test platform, such as figure 5, which is a

simple test platform composed of the console and the

539

target ship server. When the target ship is selected as

an intelligent state, it means that the target ship has

the functions of automatic generation of PIDVC

scheme and automatic collision avoidance. In this

example, the target ship 1 is chosen as the intelligent

own ship (OS), with a heading of 0.0 °and a speed of

15.0 kn ,and TS2,TS3 and TS4 are unintelligent TSs, as

shown in figure 6 (a), the initial state of the simulation

instance, as shown in Table 2. Figure (b) and (c) are

the simulation results of the automatic monitoring of

the process of the OS being turned and clearing the

target respectively. The PIDVCA algorithm of OS is

executed according to the online machine self-

learning sequence:

Figure 5. A simple sinmulation test platform

(a) (b) (c)

Figure 6. The automatic anti-collision process for simulating

example

Table2. Situation for OS relative to TS2, TS3 and TS4

_______________________________________________

No B R Ct Vt Cr Vr

(Deg) (nm) (Deg) (kn) (Deg) (kn)

_______________________________________________

TS2 21.6 3.99 225.0 14.0 200.1 26.7

TS3 49.6 11.2 224.3 10.0 197.3 23.2

TS4 151.8 4.93 115.6 12.0 152.1 23.0

_______________________________________________

The relevant information is calculated according to

step 1 and step 2, as shown in Table 3, as the DCPA

between the OS and TS2 is 0.01nm, that is

DCPA<SDA (risk judgement threshold) it constitutes

a potential collision hazard and is determined to be

the most dangerous TS.

Table 3. Information for TS2, TS3 and TS4 relative to OS

_______________________________________________

No TRF Tsea TCPA DCPA SDA DCPAn

(min) (nm) (nm) (nm)

_______________________________________________

TS2 20 1 8.93 0.01 1.37 -1.54

TS3 40 - 24.0 6.00 1.20 0.76

TS4 30 5 12.9 0.19 1.28 2.4

_______________________________________________

Table 4. L-PAE: Level of Predicted Anti-collision

_______________________________________________

L-PAE Connotation

_______________________________________________

Safety Class: 0 Only changed the course by OS（Own

ship) according to COLREGS or

ordinary practice, satisfy with

DCPA>=SDA.

Secondary Safety Only changed the course by OS

Class ：1 according to COLREGS or ordinary

Close Quarter satisfy with SDAmin<DCPA<SDA

Situation practice, just

Safety Class: 2 Only changed its original anti-collision

(Only for course direction by OS , according to

multi-vessels) ordinary practice in multi-vessels,

satisfy with DCPA>=SDA

Secondary Safety Only changed its original anti-collision

Class ：3 course direction by OS, according to

Close Quarter ordinary practice in multi-vessels just

Situation satisfy with SDAmin<DCPA<SDA

(Only for

multi-vessels)

Urgent Class: 4 Only changed the course by OS can’t be

Immediate Danger avoided collision with DCPA<SDAmin.

Situation

_______________________________________________

Step 3. Based on the encounter relationship

between OS and TS2, according to the rules applicable

between the two ships, the uasual practice of

seafarers', the OS is required to fulfill the obligations

of the responsible ship, and by calling the Crn (Crn

represents target’s new relative course after turning

AC) model and formula (2), (3) and (4) the initial

PIDVCA scheme was obtained as follows: Tisr: 0 , AC:

49° and Trr: 9.3 min, that is, to turn Stardboard to 49°

immediately, to predict that 9.3min can be resumed.

Here Trr is the maximum value of the OS's predicted

recovery time.

Step 4 check the initial scheme. Table 3 shows that

the predicted results are in the DCPAn (0.76nm) <

SDA (1.2nm) of TS3. Is it possible for the OS to pose a

new potential collision risk with TS3? By calling the

generic algorithm from case source for heuristic case-

based reasoning(as shown in the imaginary box in

figure 4), first determines whether the TCPA (24min)

of the new dangerous target TS3 is less than the latest

time (TL) of the OS turning course? Because of the

correlation model calculation, it is obvious that the

TL=2.18min, judgment result is negative; Then

continue to calls (3) and (5) to calculate the predictive

reversion limitation time of TS3. The Tc is 22.2min.

obviously, the Trr (9.3min) < Tcc (22.2min) of OS

shows that the potential collision danger between OS

and T2 will disappear naturally after OS restored. As

shown in figure 3, the initial decision scheme: (T

isr: 0,

starboard AC: 49 °, T

rr: 9.3min) is effective. According

to the result of calculation, the DCPA of OS and TS2 is

more less than SDA(1.28n.m) after OS recovering. it

can be seen that there is no collision risk between OS

and TS3. Because it is feasible to verify the result of

the initial scheme, go directly to step 6. Based on the

analysis of the evaluation table 4 of the effect of

collision avoidance prediction, according to the

implementation of the initial scheme, in the case of

the TS with guaranteed direction and speed, the

prediction safety level of avoidance effect is 0, That is,

"safe".

TS4

540

5 CONCLUSION

In this paper, a learning mechanism combining offline

artificial learning and online machine self-learning is

proposed. Based on off-line artificial learning, the

empirical knowledge of experts in the field of

collision avoidance is regularized, conceptual

knowledge modeling and the algorithm

representation of procedural meta-knowledge are

presented. The static collision avoidance knowledge

base and automatic reasoning mechanism are

constructed to provide the technical basis for online

machine self-learning. A lot of simulation results of

intelligent ship loaded with PIDVCA algorithm show

that:

1 Under the guidance of predefined reasoning

mechanism, the machine can reasoning and

calculation from the field information provided by

the perceptual system and the corresponding

knowledge provided by the static collision

avoidance knowledge base, and learn the new

knowledge to solve the collision avoidance

problem in any encounter scene.

2 The machine has the ability of perceiving TS,

cognitive TS scene and making scientific and

reasonable collision avoidance decision plan.

3 The machine has the ability to simulate people's

thinking mode in solving complex collision

avoidance problems and the ability to deal with

them.

ACKNOWLEDGMENT

Project supported by the General Program for the National

Natural Science Foundation of China (Grant no. 51879119).

REFERENCES

Chenbo,Wang & Xinyu, Z. et al. 2018.Research on intelligent

collision avoidance decision-making of unmanned ship

in unknown environments. Chinese ship research13(06):72-

Zen, Chi. 2016.Research on Autonomous collision avoidance

method of Unmanned vehicle based on Machine

Learning. Harbin University of Engineering.

Tom M. Michell. Machine Learning. 2005.China Maine press.

Jing, Shen. 2007. Theory and method of hierarchical

reinforcement Learning. Harbin Engineering University

Press.

Lina，Li.& Conggui,Chen. 1996. Research On Multi-ship

Anti-collision Intelligent Mathod (ACIM) At Widely

Area. The Proceedings of ANTI-COLLISION'96

CONFERENCE. Chiavari PublishingVol.2:214-220.

China:Dalian

Lina, Li. & Yang, S.H. et al.

2009. Study on the Theoretical

famework of Personifying Intelligent Decision-making

for Vessel Collision Avoidance. Navigation of China,

(6):30-34.

Lina, Li. 2002. Determination of the Factors about Safe

Distance of Approach and etc On the Research of Ship

Automatic Avoidance Collision. Journal of Dalian

Maritime University (3): 23-26.

Lina, Li et al. 2002. Method for Building and Optimizing to

the Intelligent Decision of Single-ship Anti-collision[J].

Navigation of China (2): 49-52.

Lina, Li et al. 2006. Study of Simulation Platform

Construction for Automatic Ship Collision Avoidance

and Its Test Method. Navigation of China (3): 47-50.

Lina, Li. & Zhennan,Xiong. et al. 2003.Method for Building

and Optimizing to the Decision of Multi-ship Intelligent

Anti-collision. Information and Control (2): 189-192.

Lina, Li. & Guoquan,Chen.et al. 2011. Construction of the

PIDVCA system and its evaluation standard. Journal of

Dalian Marine College. 37(4):1-6

Guoquan, Chen. Yong,Yin. & Lina, Li. Li et al. 2010.

Mechanism and Simulation of Personifying Intelligent

Decision- making for Vessel Collision Avoidance. The

proceedings of 2010 International Conference on

Computer Application and System Modeling V 4681-V4686

Guoquan, Chen Yin Yong, & Lina, Li. Li et al. 2015. Design

and realization of the intelligent navigational simulator.

Journal of Dalian Maritime University