Scholarly article on topic 'Bayesian Perspective on the Deck Officer's Situation Awareness to Navigation Accidents'

Bayesian Perspective on the Deck Officer's Situation Awareness to Navigation Accidents Academic research paper on "Economics and business"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Manufacturing
OECD Field of science
Keywords
{"Maritime situation awareness" / "Navigation accidents" / "Human factors" / "Probabilistic model" / "Bayesian inference"}

Abstract of research paper on Economics and business, author of scientific article — George Ad. Psarros

Abstract The deck officer of a vessel responsible for navigation relies on situation awareness to make decisions about track changes or course keeping. The culmination of accurate and complete situation awareness is affected by attentional narrowing in which important information about the current marine traffic and how this might evolve in future projections can be missed, ignored or misinterpreted. Therefore, precious time can be lost while ascertaining the best alternative in varying encounters with the risk of conflict between the officer's expectations and what is actually perceived. This may have severe consequences in the vessel's safe operation that could surface as an event sequence and finally materializing to an accident (i.e. collision or grounding). In order to prevent such disaster, it is necessary to establish in a measurable and quantifiable way how the officer's response albeit to the consistency and clearness of the processed information unfolds in parallel with the evolving situation. Such systematic approach has to be probabilistic and is described in the present paper, which extends the results evaluated in a work conducted previously by the author. By using a probability distribution to reflect the response time variability prior to an accident and the likelihood function which accounts for the information update of the data, it is possible to derive objective knowledge about the envelope of the minimum required time related to gathering data, understanding the situation and projecting ahead.

Academic research paper on topic "Bayesian Perspective on the Deck Officer's Situation Awareness to Navigation Accidents"

CrossMark

Available online at www.sciencedirect.com

ScienceDirect

Procedia Manufacturing 3 (2015) 2341 - 2348

6th International Conference on Applied Human Factors and Ergonomics (AHFE 2015) and the

Affiliated Conferences, AHFE 2015

Bayesian perspective on the deck officer's situation awareness to

navigation accidents

George Ad. Psarros*

Strategic Research, DNV GL AS, Veritasveien 1, NO-1363, H0vik, B&rum, Norway

Abstract

The deck officer of a vessel responsible for navigation relies on situation awareness to make decisions about track changes or course keeping. The culmination of accurate and complete situation awareness is affected by attentional narrowing in which important information about the current marine traffic and how this might evolve in future projections can be missed, ignored or misinterpreted. Therefore, precious time can be lost while ascertaining the best alternative in varying encounters with the risk of conflict between the officer's expectations and what is actually perceived. This may have severe consequences in the vessel's safe operation that could surface as an event sequence and finally materializing to an accident (i.e. collision or grounding). In order to prevent such disaster, it is necessary to establish in a measurable and quantifiable way how the officer's response albeit to the consistency and clearness of the processed information unfolds in parallel with the evolving situation. Such systematic approach has to be probabilistic and is described in the present paper, which extends the results evaluated in a work conducted previously by the author. By using a probability distribution to reflect the response time variability prior to an accident and the likelihood function which accounts for the information update of the data, it is possible to derive objective knowledge about the envelope of the minimum required time related to gathering data, understanding the situation and projecting ahead.

© 2015TheAuthors.PublishedbyElsevierB.V. This is an open access article under the CC BY-NC-ND license

(http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of AHFE Conference

Keywords: Maritime situation awareness; Navigation accidents; Human factors; Probabilistic model; Bayesian inference

* Corresponding author. Tel.: +47-6757-9900. E-mail address: George.Psarros@dnvgl.com

2351-9789 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license

(http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of AHFE Conference

doi: 10.1016/j.promfg.2015.07.381

1. Introduction

All too often, the human factor is usually behind everything that goes wrong at sea. Rather than assigning a value, it is preferred to adopt a broader view that considers workplace conditions, physical and natural environment, procedures, technology, training, organization and management as well as the seafarer (i.e. health, fatigue, task load, mental and motivational state, etc.). It is accepted that the previous issues affect the tasks of observing and gathering data from various sources of information, understanding, applying meaningful concepts and projecting ahead in order to effectively manage current and future vessel operations. Thus, problems may arise on how the perceived information has drawn upon the assessed situation and on how to respond in light of the planned voyage and the detected obstacles or constraints [1]. Generally, failures in functional reasoning about a causal network of unfamiliar, infrequent and complex situations can lead to unsuccessful performance with respect to decision making and action. Although humans have demonstrated remarkable capacity for flexible reasoning and the ability to process incomplete or inadequate information, their behavior is rooted in a tendency to remain fixed upon their original hypotheses instead of matching the changing environment. Their attention is captured by pre-established routines, heuristics and short-cuts that reflect the recurring patterns of past experience and may constitute the primitives of most errors [2, 3]. They are not unlimited cognitive processors and there can be points in time where they may have realized that the situation was different from what they believed it to be previously. Their situation awareness is influenced by cues and indications from a dynamic environment which implies constantly updating their interpretation of how things would develop [4].

Briefly, situation awareness represents a state of knowledge construct that goes beyond the traditional element of information processing, which demands distributed attention on how the attended cues are perceived and are translated into near future projections. Since in complex and dynamic environments the supply of attention is limited and the prediction of future states imposes heavy mental workload, a lower level of situational awareness could be generated that decreases decision effectiveness and response in a timely manner. This is a common problem of erroneous situational awareness that requires an objective evaluation [5]. Therefore, building on prior work [6], it is the purpose of the current study to address the previous challenge met by vessel bridge officers, which is based on probability theory. This approach is not aimed at debating the complex cognitive processes involved, but it is considered appropriate to account for the uncertainties associated with the human factor in understanding the perceived environment, the integrated bridge and navigation system and the maintenance of present and future conditions. The paper's structure is as follows: section 2 sets the concept of maritime situation awareness and outlines the performed work on its evaluation. Section 3 describes the probabilistic model and presents the results, whereas their implications are discussed in section 4. Finally, conclusions upon the conducted work are drawn in section 5.

2. Maritime Situation Awareness

Maintaining an adequate level of maritime safety and security requires continuous acquisition, interpretation and update of data from the environment in order to form a state of knowledge of the evolved situation. Obtaining awareness in an area and associating information with objects (i.e. vessels, leisure craft, oil platforms, wind mills, etc.) in that area is a tedious task due to high diversity, large amount and real time streaming of data. With the current advancements in commercial vessel navigation systems, it is possible to extract as much information as possible with respect to the monitored environment and its elements. Powerful sensors such as radar, automatic identification system or even infrared cameras are able to detect vessels in the observed area, as well as to learn patterns in vessel movement which in turn help in identifying and preventing incidents. Examples include traffic rule violations because of course and speed, passing offshore installations and areas where marine environmental protection considerations apply from a safe distance, seasonal fishing, routing or adherence to reporting systems and vessel traffic services, reaction of commercial shipping to bad weather or illegal activities (i.e. piracy, smuggling, etc.) [7]. However, inferring critical from non-critical situations and projecting their status in the near future may not be straightforward and there is a need to enhance the deck officer's situation awareness in order to respond more reliably.

2.1. Evaluating the navigation situation awareness of the deck officer

In reality, the prediction horizons of interest to the bridge officer extend beyond those useful for kinematic tracking purposes. Recent literature provides evidence of two or three dimensional concepts that compile meteorological, oceanographic, geospatial data and can integrate them with vessel movements as well as other intelligence information. Essentially, multiple targets can be classified in the monitored area, the extracted vessel traffic and environment observation values can be exploited to perform route prediction at a given time and warnings can be issued when a distance is reached from at least one target that is shorter than a given threshold, which are displayed on a dynamic map [8, 9, 10, 11, 12]. Although these learning systems operate autonomously through the utilization of efficient inference methods in order to recognize situations and discover the critical patterns that may affect safety, it is important to understand how the officer interacts with them. The realization of the officer's objectives in line with the interpreted regulations and prudent seamanship can be established from simulated experiments, where the timing and quality of a maneuver can be observed [13, 14, 15] and from post voyage interviews and questionnaires, where factors leading to the formulation of any decision strategy can be identified [16, 17]. However, the results of the latter approach depend on the assessor, whereas the former lacks from determining when the officer configures increased confidence and satisfaction with the automated system's derivations. This could be seen as a threat to the officer's ability to assess critically the situation with the consequence of acting randomly. Thus, it is necessary to establish a systematic and objective approach where the timeline mapping of situation awareness is modeled in a probabilistic way. This empirical method offers the benefit of modeling the uncertainties between the officer, the system and the environment without necessarily obtaining an understanding of the cognitive processes involved.

3. Model formulation

Although an evaluation of the minimum required time T for the execution of navigation activities was studied in previous work [6], it is necessary to provide an insight about its population spread. This can be evaluated from a Bayesian perspective, where a prior distribution is assigned for the unknown parameters (i.e. the spread) and is combined with a likelihood function that reflects information (evidence) about the parameters. In this way, the posterior distribution is produced which contains objective knowledge about the parameters, after the data have been observed. Let's assume that the bridge officer's response times yi before an accident occurred are drawn from a sample y1, y2,.. .yn, normally distributed as N(9,a2), with mean 9 and variance a2 unknown.

Given the quantities y,, the prior distribution p(9,a) is introduced in order to express a state of knowledge about 9 and a2 before the data are obtained, hence the distribution p(y|9,a2) can be established. Therefore, it is possible to calculate the posterior distribution p(9,a2|y) of 9 and a2 given the data y. From the sample, we know the value of T and the variance s2 with v = n-1:

L fc - t )2

s2 = --(1)

By applying Bayes' theorem and given the sample, the posterior distribution of 9 and a2 is:

p(0,CT|y) = p(0,CT)/(0,CT|y) = p(0,CT)p(m^2)p(s2 | CT2) (2)

Under the assumption that 9 and a2 are independent, i.e. p(9,a) = p(9) p(a), the posterior distribution is [18]:

p(0,a | y) = q exp^ (vs2 + n (0-T)2)j (3)

Where q is the appropriate normalization coefficient. Since Eq. (3) is a function of two parameters, a realization of its behavior can be understood by plotting probability density contours in the (9, a) plane, with each contour being represented by a curve. After taking logarithms from Eq. (3), a density contour can be defined as [18]:

-(n + l)ln <r--^(vs2 + n (d-T)2) = c (4)

By differentiating Eq. (4) the mode (i.e. highest point) of p(9,a|y) is obtained as:

= T and a = \ 12 (5)

n +1 1

By substituting Eq. (5) in Eq. (4), the characteristic contour yields:

c0 =-(n +1)^ lna+ 1 j (6)

If we integrate Eq. (3) within the region defined by a contour curve (Eq. (4)), it is possible to have an indication of the posterior probability as [18]:

-(n + 1)lnCT --L- (vs2 + n(0-T)2)= c0 -1 (2,a) (7)

2 <r 2

With x2(2,a) the upper 100a% of a chi-square distribution with two degrees of freedom and encloses a region whose probability content is approximately (1-a). Three contours at 50%, 75% and 95% are suitable to provide an adequate impression of the distribution. From tables it is found that %2(2,0.50) = 1.38629, %2(2,0.25) = 2.77259 and X2(2,0.05) = 5.99147. From published work [18], the following can be noted:

• It is concluded that a is distributed as -Jv s x^ where v s2 = ^^(yj -T)2 and v = n-1 degrees of freedom. The

limits of the 95% interval can be obtained from tables of the double-tailed chi-squared distribution corresponding to a = 0.05 and v = n-1 degrees of freedom. Thus, the lower and upper limits of a are respectively:

—2 , vs2

and ^12 (8)

2 (v) * ¿(v)

• It is shown that 9 follows Student's t distribution with v = n-1 degrees of freedom, which is symmetric and

centered at T with scaling factor—s= . Further, the limits of the (1-a) intervals of t{T,s2 -Jn,v) aregiven by:

"ifa6* (9)

3.1. Results

The sample statistics of n navigator's response times are given in Table 1 for collision and grounding respectively.

Table 1. Sample statistics [6].

n T (seconds) Standard deviation (seconds)

Collision 180 299 188

Grounding 146 520 208

The contour plots illustrated in Figure 1 show that the response time is framed by a wide envelope and the minimum required time T appears to be uncorrelated. This means that the inference about particular values of T does not have a relationship to the inference of standard deviation values. Furthermore, the observed slight flattening is attributed to the fact that the reference distribution is centered at the sample midpoint. From Eq. (5), (8) and (9) the following values are obtained (See Table 2). For collision, the minimum required time can be 299 seconds plus or minus 28 seconds, with standard deviation 188 seconds, but could be as low as 171 seconds or as high as 210 seconds. For grounding, the minimum required time can be 520 seconds plus or minus 34 seconds, with standard deviation 208 seconds, but could be as low as 188 seconds or as high as 237 seconds.

Table 2. Parameter values.

0 (seconds) a (seconds)

Interval Interval

Value Lower Upper Value Lower Upper

Collision 299 271 327 188 171 210

Grounding 520 486 554 208 188 237

The above numbers indicate that the bridge officer may have only 100 or 298 seconds prior to a collision or grounding respectively in order to obtain a quick perception of the overall situation and distances between vessels or coast, whilst it is uncertain if any response can be provided promptly. These values correspond to 33% and 57% of the minimum required time for collision and grounding respectively. However, recognition of potential dangers, translation of the comprehensive visualizations from displays into the current and future vessel operation and reconstruction of the traffic scene may take as long as 537 or 791 seconds for collision or grounding respectively. This entails that passive system observation and absence of manipulation controls have the consequence to evaluate less critically possible encounters, to loose proficiency in task completion and overreliance on system automated modes. Although the aforementioned bring out higher requirements on prudent seamanship (vigilance, skills and experience) as well as training; they can be associated with failures into identifying, understanding or treating information and judging future behavior of the vessel and her systems.

Collision (a) 50% posterior interval

Grounding (b) 50% posterior interval

2(0 268 276 284 292 300 .108 .116 .124 332 MO Q

■ Mill

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Collision (c) 75% posterior interval

Grounding (d) 75% posterior interval

268 276 284 292 300 30S 316 324 332 340

¿®U ¿OB ¿11) 1 ¿irz

4TD 480 490 500 510 520 530 540 550 560 570

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

ÜLULULLLUi

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.0 1

Collision (e) 95% posterior interval

Grounding (f) 95% posterior interval

Fig. 1. Contours of the population unknown parameters (reference to Eq. (7))

4. Discussion

Despite the fact that modern vessel navigation systems are designed to promote superior performance through information automation and the capability to suggest a maneuver alternative with the bridge officers retaining the authority for executing it or choosing another, still a loss into their situation awareness may be experienced. Taking into account the wide envelope of response time prior to an accident, a possible assertion could be unfolded into two components. From one hand, the navigators' confidence in their own ability to control the situation is greater than the possessed trust to critically assess the system's automated advisories. This overestimation of own competence may hide the possibility that their diagnosis could be incorrect and faces the risk to act fast without considering the devastating consequences that may follow. On the other hand, the additional time required by the navigators to decide how to respond to an automated situation assessment may impose an unsafe delay, albeit to deficient monitoring of the information sources feeding the system. Hence, it could be argued that the system might be enhanced by intervening when navigators are detected to be in overload, low to respond, or bypass warnings (adaptive automation). This means that action implementation tasks can be delegated to the system whenever navigators fail to respond reliably. Essentially, whether a task should be automated or not depends on the amount and type of information that needs to be tailored to the available decision time [19, 20].

It needs to be stressed that this automation enhancement capability should be based on circumstances on the moment and should be altered dynamically on the premise to boost navigators' situation awareness. This is a complex problem that requires the determination of an automation level in which part of the task is allocated to the system and the rest to the navigator. Accepting that automation should be pursued whenever navigator's situation awareness is likely to be poor or not timely, the decision to delegate automation could be based on a threshold time value (i.e. the lower limits estimated previously). Of course, for such automation that supports action implementation, a critical boundary exists where the designed system provides effortful challenges and tasks related to the navigator's engagement and monitoring of the situation. To this end, the risk of out-of-the-loop unfamiliarity as well as return-to-manual mishap is reduced, the navigator gets involved into the tasks served by automation and overall productivity of the human-automation system is improved [21, 22].

5. Conclusions

The current paper provides evidence that by employing measurable and quantifiable ways based on probability theory, it is possible to determine the variability in the response time of the vessel's bridge officer before a situation evolves into a casualty. In this formulation, a probability distribution function is firstly assigned to describe the uncertainty of the data itself (the response time) and secondly a likelihood function is utilized to reflect the knowledge coming from the developing situation. When the two functions are combined, an objective evaluation of the officer's ability to gathering data, understanding the situation and projecting ahead can be offered. The method may be adapted into improving the capabilities of current vessel navigation systems as well as training simulators and further supporting the officer's situation awareness.

Acknowledgements

The work reported in this paper has been carried out under the DNV GL Strategic Research programmes. The opinions expressed are those of the author and should not be construed to represent the views of DNV GL Group AS.

References

[1] S. Kristiansen, Maritime Transportation - Safety Management and Risk Analysis, Elsevier Butterworth-Heinemann, Oxford, 2005.

[2] J. Rasmussen, J. Occup. Accid. 4 (1982) 311-333.

[3] J. Reason, Int. J. Man-Mach. Stud. 27 (1987) 463-470.

[4] S.W.A. Dekker, J. Safety Res. 33 (2002) 371-385.

[5] M.R. Endsley, Hum. Factors. 37 (1995) 32-64.

[6] G.A. Psarros, Comparing the navigator 's response time in collision and grounding accidents. Paper presented at the 34th ASME International Conference on Ocean, Offshore and Arctic Engineering (OMAE2015-41001), St. John's, Canada, June 2015.

[7] M. Glandrup, in: P. van de Laar, J. Tretmans, M. Borth (Eds.), Situation Awareness with Systems of Systems, Springer Science+Business Media, New York, 2013, pp. 21-38.

[8] A.C. Bukhari, I. Tusseyeva, B.-G. Lee, Y.-G. Kim, Expert Syst. Appl. 40 (2013) 1220-1230.

[9] Y. Fischer, J. Beyerer, Int. J. Adv. Syst. Meas. 6 (2013) 245-259.

[10] G. Pilato, A. Augello, M. Missikoff, F. Taglino, Integration of Ontologies and Bayesian Networks for Maritime Situation Awareness. Paper presented at IEEE Sixth International Conference on Semantic Computing (ICSC 2012), Palermo, Italy, September 2012.

[11] C. Ray, R. Goralski, C. Claramunt, C. Gold, in: V. Popovich, C. Claramunt, T. Devogele, M. Schrenk, K. Korolenko (Eds.), Proceedings of the 5th International Workshop on Information Fusion and Geographical Information Systems - Towards the Digital Ocean (IF&GIS2011): Lecture Notes in Geoinformation and Cartography Vol. 5, Springer-Verlag, Berlin Heidelberg, 2011, pp. 161-175.

[12] A.H. Razavi, D. Inkpen, R. Falcon, R. Abielmona, Textual Risk Mining for Maritime Situational Awareness. Paper presented at IEEE International Inter-Disciplinary Conference on Cognitive Methods in Situation Awareness and Decision Support (CogSIMA 2014), San Antonio, TX-USA, March 2014.

[13] C. Chauvin, J.-P. Clostermann, J.-M. Hoc, Safety Sci. 47 (2009) 1222-1231.

[14] K.S. Gould, B.K. Reed, E.-R. Saus, V.F. Koefoed, R.S. Bridger, B.E. Moen, Appl. Ergon. 40 (2009) 103-114.

[15] T. Okazaki, A Study of Situation Awareness of Evading Navigation Using Marine Radar. Paper presented at IEEE 13th International Symposium on Consumer Electronics (ISCE '09), Kyoto, Japan, May 2009.

[16] M. Grabowski, S.D. Sanborn, Int. J. Hum.-Comput. Stud. 58 (2003) 637-670.

[17] K.I. Overgard, L.J. Sorensen, T.J. Martinsen, S. Nazir, in: N. Stanton, S. Landry, G. D. Bucchianico, A. Vallicelli (Eds.), Proceedings of the 5th International Conference on Applied Human Factors and Ergonomics (AHFE 2014): Advances in Human Aspects of Transportation Part I, AHFE Conference, Louisville, KY-USA, 2014, pp. 240-251.

[18] G.E.P. Box, G.C. Tiao, Bayesian Inference in Statistical Analysis, John Wiley and Sons Inc, Toronto, 1992.

[19] J.D. Lee, K.A. See, Hum. Factors. 46 (2004) 50-80.

[20] R. Parasuraman, C.D. Wickens, Hum. Factors. 50 (2008) 511-520.

[21] L. Onnasch, C.D. Wickens, H. Li, D. Manzey, Hum. Factors. 56 (2014) 476-488.

[22] T.B. Sheridan, R. Parasuraman, Hum. Factors. 42 (2000) 403-407.