Li, Hu, Huang

1

1

Data Fusion of Maritime Incident Databases Based on

2

Dempster–Shafer Theory

3 4 5 6 7 8 9

Yi-zhou LI School of Naval Architecture, Ocean and Civil Engineering and Shanghai Jiao Tong University 800 Dongchuan Road, Shanghai, P. R. China

10

Tel: (86)021-62933091, Fax: (86)021- 62933163

11 12 13

Email: [email protected]

14 15

Hao HU* School of Naval Architecture, Ocean and Civil Engineering and

16 17

Shanghai Jiao Tong University 800 Dongchuan Road, Shanghai, P. R. China TEL: (86)021-62933091, FAX: (86)021- 62933163

18 19

E-mail: [email protected]

20 21

Dao-zheng HUANG School of Naval Architecture, Ocean and Civil Engineering Shanghai Jiao Tong University 800 Dongchuan Road, Shanghai, P. R. China Tel: (86)021-62933091, Fax: (86)021- 62933163 Email: [email protected]

22 23 24 25 26 27 28 29 30 31 32 33

*corresponding author Submitted to the Transportation Research Board 93rd Annual Meeting and Publication

34 35

Word Count: Text (4836words) + 3 Figures (750) + 7 Tables (1750) = 7336 words total

36

TRB 2014 Annual Meeting

Paper revised from original submittal.

Li, Hu, Huang

2

37 38

ABSTRACT

39 40 41 42 43

Although there is a huge amount of information in all kinds of marine safety

44 45 46 47

databases, only a small part can be used directly. The disorganized information from different sources leads to a mixture of format and definition. This paper applies Dempster-Shafer theory of evidence (DST) to combine evidence (a piece of information that supports a claim) from different sources. The method is regarded as a generalization of the Bayesian theory, and it can avoid two difficulties in classical probability theory: handling the conflicting information and assigning prior probabilities. The work of data fusion is firstly demonstrated by a decision fusion problem of expert system. DST can provide a comprehensive result to the

48

decision maker by combining different experts’ opinion. Secondly, the fusion of

49 50 51 52 53

two representative maritime incident databases: GISIS (Global Integrated Shipping Information System) and ICC (International Chamber of Commerce) is conducted. Although the records of the databases have some defects such as disorder, error and contradiction, DST is still able to work effectively and calculate an uncertainty

54 55 56 57

interval of incident. Key Words: Data Fusion, Dempster-Shafer Theory, Maritime Incidents Database, Safety.

58 59

TRB 2014 Annual Meeting

Paper revised from original submittal.

Li, Hu, Huang 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109

3

1. INTRODUCTION Maritime safety management is becoming more and more important with the increasing number of ships and maritime accidents. According to statistic, there have been thousands of reported maritime accidents occurred in the last decade, and the real number could be much higher as many accidents are unreported (1). To improve the situation of frequent accidents, many organizations such as the IMO (International Maritime Organization) have made great effort to increase awareness of implementing procedures (2), but how to maintain the reliability of large-scale data (more than 80,000 merchant ships’ navigation and operation information) is a serious challenge. The poor safety management will lead to accidents. A maritime accident usually causes the losses of property and human’s lives, and it also disastrously damages the marine ecology. One of the most famous cases is the Prestige oil spill in 2002, the estimation of losses was 760 million Euros in 2003, and the impacts on seabirds and benthic organisms have been going on for more than ten years (3, 4). How to prevent such disaster from happening again is an emergent issue all over the world. Incident analysis is a major and effective way to help the decision-makers to make hopefully better decisions, and it can also push the authorities to respond to incidents quickly. The evolving information technology and data storage capacity enable people to get detailed data of maritime accidents and activities conveniently so that the analysis process may also benefit. However, a huge amount of information in all kinds of databases cannot be directly used to improve the maritime safety management because many reports are submitted spontaneously. The disorganized information from multisource and multimedia leads to a mixture of format, location, damage and definition. In addition, the underreporting of accidents is also a vital factor which influences the accuracy of probability mass in statistical analysis. Therefore, efficient and modern decision-support system calls for a solution to process imprecise, incomplete, or even conflicting information. Several data-processing techniques such as fuzzy logic, Bayesian theory, artificial neural networks and Dempster-Shafer theory are frequently used in current research. Among these methodology, Dempster-Shafer theory is the most applicable when the different data sources cannot associate a 100% probability of certainty to the output. This technique has the advantage of combining the existing evidence (a piece of information that supports a claim) and handling the uncertainty, which are two key points in accident analysis. The Dempster-Shafer theory (DST) is a promising method of dealing with some problems in data fusion and evidence combination. As a statistically based data-classification technique, it is applied where the evidence is not sufficient to assign probabilities to single events and declare they are mutually exclusive. Also, both inputs and outputs can be imprecise and be defined by sets (5). The concept of DST is relatively simple, and the technique is easily extendable. For marine transport, as an international activity with high risks, new evidence will appear and become available sometime with wars, diplomatic events or other emergency situations. The model based on DST, which allows incremental addition of knowledge, can meet the needs of these conditions. Compared with Bayesian probability theory, DST avoids the necessity of assigning prior probabilities and provides intuitive tools for managing uncertain knowledge (6). This paper presents a data fusion methodology based on DST to analyze the probability mass and uncertainty interval in maritime safety management. The contribution of the analysis is divided into two parts. The first part briefly

TRB 2014 Annual Meeting

Paper revised from original submittal.

Li, Hu, Huang 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158

4

introduces the decision fusion of independent expert reports in maritime risk assessment. The second part establishes a fusion framework for multiple incidents databases. The method is validated by selecting and categorizing the latest piracy incidents (from January 1 to June 6, 2013) in two open databases (GISIS and ICC). The results show that more than half of the attacks result in successful boarding and that an alert system is an effective tool to defer an invasion. 2. LITERATURE REVIEW People began to pay attention to the maritime safety at the end of 20th century, as the accident of shipping was equals to a huge economic loss and an irreparable environmental damage. In 1993, the IMO adopted formal safety assessment (FSA) to promote maritime safety. FSA is a rational and systematic process for assessing maritime safety risks and evaluating the costs and benefits of IMO’s options for reducing those risks (7). Following the development and introduction of the FSA method, many researchers focused on policy study and qualitative analysis before 2005. Rosqvist introduced a peer review process and suggested some qualification criteria to assist decision-maker to be confident in applying decision rules in a precautionary way (8). However, the lack of reliable safety data and lack of confidence in safety assessment became two unavoidable problems (9). In 2006, the use of Bayesian belief networks, one of probabilistic methods, was proposed by the IMO (10). Martins applied the Bayesian belief networks to the human reliability analysis of oil tanker operation, and he emphasized that the internal factors, skills and management and organizational factors should receive more attention for risk reduction (11). With the development of information and communication technology, today a lot of maritime accident analysis relies on databases. For the different users, the types of databases can be divided into public databases, commercial databases and governmental databases. The quantitative risk analysis (QRA) technique is a common technique based on database. Ronza analyzed some tens of thousands of hydrocarbon spills records from two vast US federal spill databases (HMIRS, by the Department of Transportation, and MINMOD, by the US Coast Guard) to account for the possible outcomes of hazardous material spill (12). Another frequently adopted statistical technique is regression and clustering analysis. Yip proposed a binomial regression model based on historical accident data of years 2001-2005 to analyze port traffic risks in Hong Kong Harbor (13). Yin conducted a relatively comprehensive quantitative risk assessment for maritime safety management based on the GISIS and world casualty statistics of Lloyd’s Register (14). Every database is incomplete to some extent because of the limited data resources so that the data fusion technique is valuable to finding the conjunction of the events and the associated probability. The U.S. Department of Defense (DOD) defines data fusion as a multilevel, multifaceted process dealing with the automatic detection, association, correlation, estimation, and combination of data and information from single and multiple sources (15, 16). Most data fusion and mining techniques are derived from physical models, feature-based inference, and cognitively based models (17). Fuzzy Logic is a good choice to solve the uncertainty problem if the data is not enough. Huang applied fuzzy technique to assess the risk of maritime passages (18). Li improved the model of fuzzy logic to satisfy the updating of changing hazard factors (19, 20). However, with the

TRB 2014 Annual Meeting

Paper revised from original submittal.

Li, Hu, Huang 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192

5

normalization of databases and the accumulation of data, Dempster-Shafer theory can provide more reliable solutions in the field of data fusion. The Dempster-Shafer theory was put forward by Dempster to describe uncertainty problems, and the theory applied upper and lower traditional probabilities to combine independent sources of information (21). Subsequently the theory was defined as a generalization of the Bayesian theory (22). Shafer developed the theory systematically and improved uncertain inference, and it was assumed that all belief functions referred to the same problem or alternatively they were false (23). DST is regarded as a promising method of some of the basic problems arising in combination of evidence, and it only requires a minimum familiar with relational models of data (24). Lawrence implemented DST in a decision support system for advanced traffic management, and the field testing demonstrated its effectiveness (25). Yi introduced the basic rough-sets technique, which was able to reasonably and systematically process a large amount of traffic information, as an alternative to relying on the intuition of traffic operators and system managers (26). The mainstream applications of DST can be seen in the fields of travel time estimation (27), image forensics (28) and automatic identification system (AIS) (29). 3. DEMPSTER–SHAFER THEORY 3.1. Basic Formalism of DST The discernment frame of DST is defined as a finite set   x1 , x2 ,..., xn  . A variable X is any possible value of the problem, and a proposition about variable X is any subset of  (the set of all possible propositions is the power set of  ). For example, a tanker can either be attacked by pirates or not during marine oil transportation. In this scenario, people can define a variable P with frame   p1 , p2  where p1 is the proposition “tanker is attacked by pirates”, p 2 is the proposition “tanker is not attacked by pirates”, p1  p2 is the doubtful proposition “tanker is or is not attacked by pirates”. The basic probability assignment (BPA) is defined on a universal set A as a function of the power set  in the interval [0, 1]. The function is written as: (1) m : 2   0, 1 which satisfies the conditions: m   0, m A  0,  m A  1 (2) A

193 194 195 196 197 198

199

200 201

The value of the BPA, also called mass, expresses the amount of evidence supporting the claim that an element of the universal set X belongs to the set A . Set with non-zero mass is called focal element. Continuing the example in the first paragraph of this section, when the tanker drives into a specific region such as the Strait of Hormuz, people can get the BPA from historical information. Supposed the function is: m p1   0.3 m p   0.5  2 m (3) m p1  p 2   0 m   0.2 where the set p1  p2 is not a focal element in Equation 3 for m p1  p2   0 . Usually, only focal elements will be showed in mass assignments. It means that

TRB 2014 Annual Meeting

Paper revised from original submittal.

Li, Hu, Huang 202 203 204 205 206

6

any condition like the third line of Equation 3 will not be appeared in the following context. The belief function and the plausible function are two non-additive evidential measures of the DST, and they can be calculated from the BPA. For any set A, B   , the belief function is defined as:

Bel  A 

207

 mB 

(4)

B A

208 209 210 211 212

Bel  A represents the minimum belief, which summarizes all reason to believe in A with the available knowledge. With the definition of BPA, the function satisfies Bel    0 and Bel   1 . The plausible function is defined as: Pl  A  1  Bel A   mB  (5)



B A

213 214 215 216 217 218 219 220 221

Pl  A represents the maximum belief, which summarizes all reason not to reject in A with the available knowledge. The function also satisfies Pl    0 and Pl   1. According to the definition above, one of the relationships between the belief function and the plausible function is Bel  A  Pl  A , and it is possible to describe uncertainty of evidential measures by the lower and upper bounds of an interval Bel  A, Pl  A. If the interval is reduced to a point, in which case Bel  A  Pl  A , the problem will be converted into a classical probability problem. If the interval is 0, 1, in which case Bel  A  0 and Pl  A  1 , it means a total ignorance about A . Classical Probability Theory

 

P  A

P A  1  PA

0

1 Dempster-Shafer Theory

Bel  A 0

222 223 224 225 226 227 228 229 230 231 232 233

Uncertainty

 

Pl  A   1  Bel A



Bel A

1

FIGURE 1 Comparison between classical probability theory and DST. Figure 1 shows the fundamentally difference between the classical probability theory and the DST. We should highlight that for A   , Bel  A  Bel A  1 , and the uncertainty part is the lack of information about A .



3.2. Combination Rule of DST The combination rule of DST allows people to merge two independent evidence sources into a single one if two BPAs are defined over the same frame. Here the concept of “independent” in DST is not strictly defined. The word just indicates that the different pieces of evidence are determined by different means.

TRB 2014 Annual Meeting

Paper revised from original submittal.

Li, Hu, Huang 234 235

Assume that Bel1  A and Bel 2  A are believe functions over the same frame  with the BPAs m1 and m2 . Then for any X   , the function m12 is defined as:

m12  X  

236 237 238

7

where

K

1 1 K

1

1

(7)

2

m12 is called the orthogonal sum of Bel1  A and Bel 2  A , denoted by Bel1  A  Bel 2  A . The coefficient K implies the conflict between Bel1  A and Bel 2  A . The higher the K , the higher the conflict. Since K is obtained by accumulating the product of masses assigned to sets having empty intersection. Furthermore, the combination rule treats conflict as a normalization factor, so its presence is no longer visible after fusion. The combination rule of DST is commutative and associative, so the following equations hold: Bel1  A  Bel 2  A  Bel 2  A  Bel1  A (8) Bel1 A  Bel 2 A Bel 3 A  Bel1 A  Bel 2 A  Bel 3 A (9) Figure 2 shows that multiple evidence combination can be transformed into the recursion of several dual evidence combinations refer to the properties of combination rule expressed in Equations 8 and 9. Thus it is convenient for people to add new source of evidence to an old system in arbitrary order. Bel1 Bel2

Bel1 Bel2

… Beln

254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270

(6)

2

 m  Am B 

A B 

239 240 241 242 243 244 245 246 247 248 249 250 251 252 253

 m  Am B

A B  X

DST Combination Rule

Bel

Bel3

… Beln

DST Combination Rule DST Combination Rule DST Combination Rule

Bel

FIGURE 2 Transformation of multiple evidence combination. Despite some good properties, DST rule is not idempotent. This means that observing twice the same evidence results in stronger beliefs. Therefore the hypothesis of independent sources is very important. 4. MARITIME SAFETY MANAGEMENT BASED ON DST 4.1. Decision Fusion of Risk Assessment DST can handle the uncertainty caused by the lack of sufficient evidence. The desirable property helps decision-maker to consolidate the conflict opinion of different experts. For example, the discernment frame of the variable R (Risk) is usually assumed as {L, M, H}, presenting three qualitative levels: low, medium and high, respectively. If an Expert A reports that the low risk with a probability of 60 % and medium or high risk with a probability of 40%, while an Expert B reports that the

TRB 2014 Annual Meeting

Paper revised from original submittal.

Li, Hu, Huang 271 272 273 274

8

risk probabilities of low, medium and high are 50%, 30% and 20%, respectively. The decision fusion of two opinions is depicted in Table 1 and Table 2. TABLE 1 Decision Fusion of Risk Assessment with Empty Set mB L   0.5 mB M   0.3 mB H   0.2 mA L  0.6 mL   0.30   0.18   0.12 mA M  H   0.4 mM   0.12 mH   0.08   0.20

K  0.18  0.12  0.20  0.50  275 276

1 2 1 K

TABLE 2 Normalized Probability Masses for Risk Assessment mB L   0.5 mB M   0.3 mB H   0.2 mA L  0.6 mL  0.60 0 0

mA M  H   0.4

0

mM   0.24

mH   0.16

277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297

1 , and K 1 K is calculated in the last line. Table 2 lists the normalized probability masses for the problem, and each mass with non-empty set in Table 1 was multiplied by the 1 . Finally the BPA of the decision fusion of two experts is coefficient 1 K expressed as: mL   0.60  m  mM   0.24 (10) mH   0.16  In addition, if there is an Expert C who reports new opinion, the second fusion can be executed based on Equation 10. Of course, the experts are not allowed to communicating with each other due to the hypothesis of independence described in the end of Section 3. Table 1 is the application of Equation 6 without the coefficient

4.2. Data Fusion of Incident Database There are many marine transportation safety databases with an extremely huge amount of data. While providing the support to maritime safety management, the multisource data with imprecise (or conflict description) also makes analysis difficult or impossible to perform in a logistical straightforward fashion. DST can find the conjunction of the events and the associated probability with an uncertain and incomplete input, which is reasonable for data fusion of maritime incident databases.

TRB 2014 Annual Meeting

Paper revised from original submittal.

Li, Hu, Huang

Discernment Frame Building Useful Categories

9

Data Filtering

Database

Dual Fusion

DB 1

New Data Source Appearing

DB 2 DB A

Database

298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336

DB 3

FIGURE 3 Data fusion framework of databases. As expressed in Equations 8 and 9, any DST fusion problem can be regarded as the recursion of a dual fusion problem. Figure 3 shows the data fusion framework of databases. The first step is to select useful categories of evidence though which to build discernment frame of DST. Then filter the data to remove the redundancy information and to find the useful information. Next step is to use the DST combination rule to execute the data fusion. At last, if there are more than two databases, return to step 2 and continue the circulation until all databases are included. The application of DST data fusion of an incident database will be detailed introduced in Section 5. 5. APPLICATION OF DST IN PIRACY INCIDENT ANALYSIS Many maritime incident databases are available to public so that people can download the data they need via Internet. It is beneficial to the incident analysis and promotes the process of maritime safety improvement. In this section, the application of DST will be conducted by the piracy incident analysis. The reason why we chose the piracy issue as a case study was based on three points. Firstly, piracy and armed robbery against ships remain a real and ever-present danger to those who used seas for peaceful purposes, and the issue is very hot in maritime safety management. Secondly, the piracy incident databases are relatively new and open, which offer the access to the first-hand data. Thirdly, due to the information in the databases was submitted by people having different incident knowledge and different language skill, some problems such as disorder, error and contradiction were inevitable. However, DST has the advantage of solving these representative problems. Generally speaking, the boarding of pirates is a signal that the crew’s lives are under threat. To analyze the boarded probability of pirates and the injury probability of crew, the discernment frame is designed as the set   x1 , x2 , x3 . Three events of the frame are: x1 = pirates having boarded and hurt (or killed) people of the vessel. x 2 = pirates having boarded but have not hurt (or killed) people of the vessel. x3 = pirates attempted to board but failed. The analysis is based on two open databases. One is the Piracy and Armed Robbery Module of Global Integrated Shipping Information System (GISIS) (30), and the other is Live Piracy & Armed Robbery Report 2013 of International Chamber of Commerce (ICC) (31). There were 161 records in GISIS and 137 records in ICC hit by the search limited to date from January 1 to June 30, 2013.

TRB 2014 Annual Meeting

Paper revised from original submittal.

Li, Hu, Huang

10

337 338 339 340 341 342 343 344

However, in these databases, one record corresponds to one report. The information stored in the report is a piece of narration. Many narrations were written without a uniform format, which may lead to the uncertainty of understanding. As a result, the classification of data is equal to reading the narrations of incident reports because the information from two databases cannot be used to support the events directly. The following examples were copied from two reports in the databases. Example 1 contains enough evidence, but Example 2 is perfunctory.

345 346 347 348

Example 1 (Original) Robbers in a small boat attempted to board an anchored chemical tanker via the anchor chain. Alert duty crew spotted the robbers and raised the alarm resulting in the robbers aborting the attempt. Port control informed.

349 350 351

Example 2 (Original) A tug towing a barge enroute from Singapore to Kuantan noticed stores and vessel's properties stolen from the barge upon arrival at Kuantan Pilot Station.

352 353 354 355 356 357 358 359 360 361 362

The narrations of Example 1 are no doubt satisfied the proposition of x3 . The information of Example 2 is too simple to get enough evidence. A variety of reasons will lead to the losses of properties, and the proposition here should be either x 2 or x3 .

363 364

x2  x3 (There is a doubt whether pirates having boarded)= 2/8 = 25.0 % TABLE 3 Sample Key Information List # Arms Crew Boarded Alarm 1 knives safe yes raised 2 guns and knives safe yes n/a 3 none n/a no raised 4 heavily kidnapped yes raised 5 n/a n/a doubt n/a 6 knives safe yes n/a Example 1 n/a n/a no raised Example 2 not seen n/a doubt n/a

365 366 367 368 369

Table 3 lists one part of key information of two examples and six other reports in the databases. This table is only used to demonstrate the process of fusion, and the result is not homogeneous for all reports. According to Table 3, x1 (Pirates having boarded and hurt people) = 1/8=12.5% x 2 (Pirates having boarded but have not hurt people) = 3/8 = 37.5% x3 (Pirates attempted to board but failed) = 2/8 = 25.0%

Similarly, we extended the work to all records in two databases, and the categorized information is listed in Table 4. TABLE 4 Numbers and Ratios in Terms of Discernment in Individual Source GISIS ICC Number Ratio (%) Number Ratio (%) x1 22 14 18 13

TRB 2014 Annual Meeting

Paper revised from original submittal.

Li, Hu, Huang

x2 x3 x2  x3 Total 370 371 372 373 374 375 376

377 378

379 380 381 382 383 384 385

386 387 388 389 390 391

11 69 29

43 18

65 29

48 21

41 161

25 100

25 137

18 100

The data of ratio in Table 5 is useful to calculate the probability masses. Similar as Section 5, the data fusion of GISIS and ICC is depicted in Table 5 and Table 6. In Table 6, mx2  x3   0.26 represents the uncertainty interval width of x2 and x3 was both 0.26. TABLE 5 Data Fusion of Incidents Databases with Empty Set mB x2  x3  mB x3  mB x1  mB  x 2  0.13 0.48 0.21 0.18 m A x1  0.14 0.0182 0.0672 0.0294 0.0252 m A  x2  0.43 0.2064 0.0774 0.0559 0.0903 m A x3  0.18 0.0378 0.0324 0.0234 0.0864 m A x 2  x3  0.25 0.1200 0.0525 0.0450 0.0325 K  0.0672  0.0294  0.0252  0.0559  0.0903  0.0234  0.0864  0.0325  0.4103 TABLE 6 Normalized Probability Masses for Incidents Databases mB x2  x3  mB x3  mB x1  mB  x 2  0.13 0.48 0.21 0.18 m A x1  0.14 0 0 0 0.03 m A  x2  0.43 0 0 0 0.56 m A x3  0.18 0 0 0 0.15 m A x 2  x3  0.25 0 0 0 0.26 Finally, the belief function and the plausible function and their ranges are obtained by Equations 4 and 5. They are listed in Table 7. This procedure results in event x2 having the highest mass, indicating that pirates having boarded but have not hurt (or killed) people of the vessel with a probability of 0.56 to 0.82. TABLE 7 Belief Functions, Plausible Functions and Uncertainty Interval m X  Bel  X  Pl  X  Uncertainty x1 0.03 0.03 0.03 [0.03, 0.03] x2 0.56 0.56 0.82 [0.56, 0.82] x3 0.15 0.15 0.41 [0.15, 0.41] 6. CONCLUSION This paper presents a data fusion technique based on Dempster-Shafer theory of evidence, which may help to increase the reliability of incident probability mass function in maritime safety management. The concept of DST is clear and easy to apply. With the application in decision fusion of risk assessment, the method has

TRB 2014 Annual Meeting

Paper revised from original submittal.

Li, Hu, Huang

12

392 393

proved to provide the decision maker a comprehensive result of the combination of experts’ reports.

394 395 396 397 398 399 400 401 402

The case study of piracy shows that about 56%-82% of the total piracy attacks can board successfully. How to prevent the pirates from boarding is an emergent issue for analysts and experts to discuss. By the way, in the process of categorizing data, we observed that an alert system can effectively expel most pirates (alarm can interrupt most of attack attempt and force about half of the boarded pirates to abandon the attack halfway). This may be the reason why the injured rate of crew is relatively low after the pirates boarded. In most cases, the pirate looks like a thief rather than a robber. However, the relationship between the alert system and seamen’s lives protecting needs further analysis.

403 404 405 406 407

Although DST allows a set as the input, the result is also presented as an interval. The length of the interval reflects the degree of uncertainty. The form of the result allows the decision maker to adopt different decisions with risk aversion degree. More databases should be incorporated to increase the accuracy of probability mass and reduce the uncertainty in future work.

408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438

REFERENCE 1. Huang, D. Z., Hu H., Li Y. Z. Spatial Analysis of Maritime Accidents Using Geographic Information System. Presentation at the 92nd Transportation Research Board Annual, Washington D.C., 2013. Accepted by Transportation Research Record. 2. International Maritime Organization (IMO). Strategic Plan for the Organization (For the six-year period 2012-2017), Resolution A. 1037 (27), 2011, pp. 3-4. 3. Garza-Gil M. D., Prada-Blanco A., Vazquez-Rodriguez M. X. Estimating the Short-term Economic Damages from the Prestige Oil Spill in the Galician Fisheries and Tourism. Ecological Economics, Vol. 58, No. 4, 2006, 842–849. 4. Kingston P. F. Long-term Environmental Impact of Oil Spills. Spill Science & Technology Bulletin, Vol. 7, Nos. 1-2, 2002, 53-61. 5. Sentz K, Ferson S. Combination of evidence in Dempster–Shafer theory. Technical report. Sandia National Laboratories. 2002. 6. Fontani M., Bianchi T., Rosa A., Piva A., Barni M., A Dempster-Shafer framework for decision fusion in image forensics. 2011 IEEE Int. Workshop on Information Forensics and Security (WIFS), 2011, pp. 1–6. 7. International Maritime Organization (IMO). Guidelines for the Application of Formal Safety Assessment (FSA) to the IMO Rule-Making Process, London, April 2002. 8. Rosqvist T., Tuominen R. Qualification of Formal Safety Assessment: an exploratory study. Safety Science, Vol. 42, 2004, 99–120. 9. Wang J. The Current Status and Future Aspects in Formal Ship Safety Assessment. Safety Science, Vol. 38, 2001, 19–30. 10. International Maritime Organization (IMO). Consideration on utilization of Bayesian network at step 3 of FSA, Submitted by Japan, MSC81/18/1, February 2006. 11. Martins M. R., Maturana M. C. Application of Bayesian Belief Networks to the Human Reliability Analysis of an Oil Tanker Operation Focusing on

TRB 2014 Annual Meeting

Paper revised from original submittal.

Li, Hu, Huang 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487

12. 13. 14. 15. 16. 17. 18.

19.

20.

21. 22. 23. 24. 25.

26.

13

Collision Accidents. Reliability Engineering and System Safety, Vol. 110, 2013, 89–109. Ronza A., Vilchez J. A., Casal J. (2007) Using Transportation Accident Databases to Investigate Ignition and Explosion Probabilities of Flammable Spills, Journal of Hazardous Material, 146, 106-123. Yip T. L. Port Traffic Risks: a study of accidents in Hong Kong waters. Transportation Research Part E: logistics and transportation review, Vol. 44, 2008, 921-931. Yin J. B. Quantitative risk assessment of maritime safety management. Doctor thesis, the Hong Kong Polytechnic University, 2010. White F. E., Jr. Joint Directors of Laboratories Data Fusion Subpanel Report: SIGINT Session. Technical Proceeding of Joint Service Data Fusion Symposium, Vol. 1, DFS-90, 1990, 469–484. Data Fusion Development Strategy Panel. Functional Description of the Data Fusion Process. Office of Naval Technology, Washington, D.C., 1991. Klein L. A. Sensor and Data Fusion Concepts and Applications, 2nd Edition, SPIE Press, Bellingham, Wash., 1999. Huang D. Z., Hu H., Li Y. Z. Application of Fuzzy Logic to Safety Risk Assessment of China’s Maritime Passages. Transportation Research Record: Journal of the Transportation Research Board, No. 2273, Transportation Research Board of the National Academies, Washington, D. C., 2012,pp. 112– 120. Li Y. Z., Hu H., Huang D. Z. Dynamic Fuzzy Logic Model for Risk Assessment of Marine Crude Oil Transportation. Transportation Research Record: Journal of the Transportation Research Board, No. 2273, Transportation Research Board of the National Academies, Washington, D. C., 2012, pp. 121-127. Li Y. Z., Hu H., Huang D. Z. Huang, D-Z. Developing an Effective Fuzzy Logic Model for Managing Risks in Marine Oil Transport. Int. J. Shipping and Transport Logistics, Vol. 5, Nos. 4/5, 2013, pp.485–499. Dempster A. P. (1967). Upper and Lower Probabilities Induced by a Multivalued Mapping. The Annals of Mathematical Statistics. Vol. 38, No. 2, 325-339. Dempster A. P. (1968). A Generalization of Bayesian Inference. Journal of the Royal Statistical Society Series B. Vol. 30, No. 2, 205–247. Shafer G. (1976). A Mathematical Theory of Evidence. Princeton, N J: Princeton University Press. Zadeh L. A. A Simple View of the Dempster-Shafer Theory of Evidence and its Implication for the Rule of Combination. AI Magazine, Vol. 7, No.2, 1986, 85-90. Klein L. A., Yi P., Teng H. Decision Support System for Advanced Traffic Management Through Data Fusion. Transportation Research Record: Journal of the Transportation Research Board, No. 1804, Transportation Research Board of the National Academies, Washington, D. C., 2002, pp. 173-178. Yi P., Lu H., Zhang, Y. Rough Sets and Probability Masses for Dempster– Shafer Data Fusion at a Traffic Management Center. Transportation Research Record: Journal of the Transportation Research Board, No. 1836, Transportation Research Board of the National Academies, Washington, D. C., 2003, pp. 151-156.

TRB 2014 Annual Meeting

Paper revised from original submittal.

Li, Hu, Huang 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504

14

27. Faouzi N. E., Klein L. A., Mouzon O. Improving Travel Time Estimates from Inductive Loop and Toll Collection Data with Dempster–Shafer Data Fusion. Transportation Research Record: Journal of the Transportation Research Board, No. 2129, Transportation Research Board of the National Academies, Washington, D. C., 2009, pp. 73-80. 28. Fontani M., Bianchi T., Rosa A., Piva A., Barni M. A Framework for Decision Fusion in Image Forensics Based on Dempster–Shafer Theory of Evidence. IEEE Transactions on Information Forensics and Security, Vol. 8, No. 4, 2013, 593-607. 29. Talavera A., Aguasca R., Galvan B., Cacereno Andres. Application of Dempster–Shafer Theory for the Quantification and Propagation of the Uncertainty Caused by the Use of AIS Data. Reliability Engineering and System Safety, Vol. 111, 2013, 95–105. 30. Global Integrated Shipping Information System (GISIS). Piracy and Armed Robbery Module. http://gisis.imo.org. 31. International Chamber of Commerce (ICC). Live Piracy & Armed Robbery Report 2013. http://www.icc-ccs.org/piracy-reporting-centre/live-piracy-report.

TRB 2014 Annual Meeting

Paper revised from original submittal.