Beth Cardier
Trusted Autonomous Systems, Defence CRC
Griffith University, Nathan, Australia
b.cardier@griffith.edu.au
Mathew Hancock
Sustainable Minerals Institute
University of Queensland, St Lucia, Australia
mathew.hancock@uq.edu.au
Abstract
How can we represent the risk that emerges when multiple processes interact? This research considers how safety incident information can be visualized using a narrative modeling method that represents the accumulation of effects through different contexts.
Currently, organizations conceptualize safety incidents in terms of the formats in which they are classified – assets, roles, cost and deviations from procedures. However, some accidents should also be understood in terms of accumulation and context. An example is presented in which a reversing vehicle caused a fatal injury because safety modifications made for one work location actually reduced safety in another.
This paper explores the inclusion of cumulative and contextual information in a workplace safety scenario by presenting it in a narrative modeling visualization. Animated examples will be shown in the presentation; this paper provides still images of those visualizations and links to the animations. The aim is to develop foundations for a formal approach that reveals how interventions can have a ripple effect across a distributed and dynamic system. A long-term goal is to develop an analytical modeling tool that tracks how influence promulgates through complex heterogenous systems, such as collaborative human-machine teams.
Keywords — Visualizing Influence, Contextual Reasoning, Dynamic Knowledge Representation, Workplace Safety, Narrative Modeling.
I. Introduction
Workplace accidents harm both employees and their organizations. In 2020, there were three fatal accidents for every 200,000 Australian workers, and workplace injuries resulted in an average of 6 weeks’ time lost per incident [1]. The full impact of those accidents is orders of magnitude greater because it also includes emotional distress, loss of human capacity, compensation and operational costs.
Organizations strive to make their practices safer by assessing the risks associated with work being undertaken, designing control frameworks and embedding controls in front-line management plans and work procedures. However, these formats still struggle to include cumulative effects in their analysis, such as the way risk can aggregate when multiple processes intersect, or the way safety parameters can change when an employee drives from one work location to another – as occurred in the example discussed here.
Cumulative and contextual information is not usually included in workplace safety models, however. Context and multi-system interaction are famously difficult to formally represent [2][3]. As well, in the domain of workplace safety, cumulative risk is usually associated with environmental and ecological hazards, where repeated exposure to a ‘stressor’ over time is evaluated in terms of concentration and exposure limits [4].
Niemeier et al. note that this kind of combined, cumulative risk analysis is “rarely applied in workplaces” [5]. Work organizations instead frame safety incidents in terms of sanctioned procedures, deviations from those procedures and required modifications – classification schemes such as the Incident Cause Analysis Method (ICAM) [6] or the WHO injury surveillance guidelines [7] inform these conceptualizations.
Current representations of relationships between employees and assets are thus not easy to model as networks of influence that can change depending on context. If this could be achieved, it could directly improve the way safety is managed.
This research uses a dynamic visualization method from narrative analysis to introduce cumulative and contextual information to a workplace safety model. Narrative comprehension has some unique features that support those qualities, such as the ability to combine information from multiple contexts [8] and then follow the accumulation or changes in those conditions as they move among situations [9].
This research offers those knowledge structures to risk analysis in the form of two animations, Animation A and Animation B, which can be viewed using the embedded links. Still images from those models appear in this paper. The method, what it can reveal about this scenario and future applications are discussed.
II. Focus Scenario
A. A Reversing Vehicle Accident
The following example was derived from a collection of Queensland government coroner’s reports, with a focus on [10]. Key details have been changed and added to make the example simpler than a real-world example would actually be. None of the incidents considered were connected to the organizations responsible for this research. A constructed scenario follows.
Two years ago, ACME Ltd won a contract to perform roadworks in an airport tunnel. A site assessment recommended an upper limit of 95 decibels (DB) for vehicles on site, to comply with the noise restrictions of the surrounding urban environment. Ten of ACME’s vehicles were fitted with 95 DB ‘squawker’ reversing alarms and their standard 105 DB alarms were removed.
Three weeks into the project, one of the modified vehicles drove from the worksite to ACME’s headquarters to drop off paperwork. ACME’s carpark is located next to a quarry which was producing noise of 100 DB that day. ACME does not manage this quarry. As the modified vehicle was reversing out of the carpark, it ran over another employee who was walking past and could not hear the reversing ‘squawker’. The injured employee later died in hospital.
This example illustrates the way “many interacting contextual factors” [11, p. 12] can align to cause injury or fatality in a workplace. This convergence is sometimes described as a ‘Swiss Cheese’ paradigm, in which “a trajectory of accident opportunity” [12, p. 10] emerges when numerous systems intersect. These convergences can include organizational factors, local conditions and task factors [6, p. 12].
A goal of this work is to understand these different domains as ‘contexts’ and represent their intersections so that safety vulnerabilities can be identified and managed.
III. Method
A. Dynamic Narrative Contexts
The narrative modeling method used in this work was originally developed by Cardier to represent how implicit themes and conceptual structures accumulate in relation to an explicit story text. That concern is similar to that required by the example, in which the affordances of a surrounding context inform how an activity unfolds, even if many of its elements are not explicitly identified.
The resulting modelling method has been described extensively elsewhere [13][14][15], as it has been used to visualize the interaction among contexts and cumulative influence in story interpretation and PTSD.
This research imports workplace task and safety information into that visualization approach. The goal is to introduce three different kinds of information to representations of those workplace processes:
- Multiple contexts and the interactions among them.
- Affordances / controls that are specific to particular contexts – for example, an underground tunnel has different acoustics from an open outdoor situation.
- The ability to adjust elements in the model and view how new consequences cascade through the system.
Qualities (1) and (2) are represented in a special layout and operators that represent context. These extend conventional knowledge representation (KR) techniques so that cumulative and cross-contextual effects can be recorded. ‘Animation B’ illustrates (3) by showing an adjustment to ‘Animation A’.
B. Graphical Taxonomy: Layout
This visualization method uses a layout similar to that of business process models, in which multiple contexts are laid side by side, running from left to right. These bands would be called ‘swim lanes’ in a business process model [26]. That format enables interactions to be shown as connected objects, even if they cross or occur within different contexts.
This layout is unlike business process models in a few important respects. The first concerns the bottom section, the ‘Interpretation’ space. That section pulls together information from across the entire field and gathers it into a structured situation.
This ‘summary’ is important because it represents a context-based assessment of the entities it contains, both within and across situations. As activities unfold, these summarizing situations are formed and reformed, enabling contextual states to be compared and linked together.
C. Graphical Taxonomy: Operators
Operators draw structure over the above layout. Nodes and links represent pieces of information and their connections – this is identical to normal KR approaches [23]. Unlike ordinary methods, however, there are also operators that group nodes and links into local situations (eg. ‘acoustics in a tunnel’). These are local ‘contexts’.
They operate as a combined system; for example, in the workplace, this could be a task with its own procedures or an environment with special features. By encircling node-link networks with a context boundary, tasks can be combined while preserving conditions needed for their operation. A new tier of structure is thus added: there are nodes and links, just like traditional approaches, but these are also grouped into local contexts and their connections.
As information emerges across different domains, nodes appear in the layered bands (swim lanes) to signal its role in the process and follow-on effects. Within and across those layers, those networks are further organized into groups. Given context boundaries and a means to cross or connect them, it becomes possible to record when influence and its consequences are relayed through a system.
Graphically, these features are chained together using animation. This produces visualizations of knowledge structures as they interact and change – a dynamism that is key to representing influence as it passes through numerous processes. The way those influences follow each other can be recorded as an unfolding structure, or as colors which run over the top of those structures, or both.
IV. Example Visualizations
A. Animation A: Processes that Lead to Fatality
The two animated models, A and B, enable two forms of influence to be depicted: internal (the cascade of workplace processes) and external (an adjustment made by the analyst).
Both animations visually track how a series of ACME events unfold, along with the contexts that inform those actions. ACME wins a contract to construct an airport tunnel, modifies its vehicles to comply with noise regulations and then allows those vehicles to drive to a different worksite. The visualization also records how state regulations and ACME’s policies are being adhered to during all these processes.
At the end of A, a new risk is flagged when a modified vehicle drives to the noisy office site. There were no lapses in prescribed processes, so the risk identified is a result of a systemic failure rather than a human one.
B. Notable Moment: Safe Becomes Risk
The key event is when the modified vehicle drives to ACME’s head office carpark, and the noise from the quarry creates a new risk. This is represented as an encirclement of features (‘ACME vehicle’) being dragged into a new location, ‘ACME office & yard’.
When that vehicle shifts to a new location, risk is identified and indicated by red highlighting and a modification of the text in the nodes.
The unsafe outcome needs to be avoided. Now, imagine the analyst has a tool that allows a new node to be inserted, to reveal how the consequences change with this new condition. This adjustment is represented by Animation B.
C. Animation B: Adjustment by the Analyst
Animation B uses the same structure as Animation A in every respect except for the parts of the diagram that have been deliberately adjusted by the analyst. This asymmetry enables a direct visual comparison of structure between the two graphical models.
By using a different colour (in this case, aqua blue) to track the adjusted structure and its flow-on connections, Animation B indicates which aspects are related to the adjustment. This visualizes the domino effect of intervention, where successive states may be influenced by it.
In this case, the analyst has added a new regulation that requires reversing alarms to sense the sounds in a worksite and adjust its signal to be louder. When the two models are compared, they indicate that the analyst’s adjustment produced a different outcome for the workplace task: the safety risk was eliminated.
The comparison also reveals which combination of factors would be more optimal for a safe workplace. Animation B does not contain serious risk, whereas Animation A does. This is signalled by the presence of the color red in A, and its absence from B, along with relevant text.
V. Surrounding Literature
The graphical taxonomy for this method combines representational features found in multiple domains, to depict simultaneous contexts, conceptual change and cumulative influence. The new addition to this work is conceptualizations of safety information, so that is the focus here.
Methods of representing safety information fall into two primary fields – assessing and managing risk, and the investigation of safety incidents. There are a variety of health, safety and environmental risk assessment methods, ranging from simple pre-task risk assessments to rigorous analyses such as HAZOP and fault tree / event tree studies [16][17].
These methodologies are generally point-in-time studies carried out before work commences, to identify the control strategies and corrective actions required to manage risks. Exhaustive modelling of all potential failures, assumptions and contexts is costly and time consuming. It is generally only warranted for applications that require extreme reliability such as space flight or nuclear power operations. Instead, most risk assessments set the scope and granularity of analysis to suit the decisions that will be made [18].
Incident investigation methodologies seek to explain how an incident occurred – or could have occurred if conditions were slightly different (near misses). There are a variety of approaches in incident investigation which primarily rely on narrative reporting which is accompanied by structured analysis, within a framework that is presented in a narrative or tabular format [19][20].
Representations of information in incident analyses most often focus on causal factors and recommended improvements, with more detailed or mature approaches examining sequences of events, the effectiveness of controls and the influence of human factors [21]. Of note is the government research by Ale et al. [22] which combined insights across many incident analyses to quantify risk exposure based on local conditions and built causal models from them – an example of incident analysis at scale.
These safety information formats are static and generalized. Narrative modeling was able to introduce a different form of information. Knowledge Representation approaches are also static and generalized. In computer science, structures such as ontologies [24] must be generalized to enable the repeatability that comes with automation. Those structures are represented as networks of objects and relations (graphically, these are nodes and links) [13].
Multi-contextual processes present a problem for this particular kind of generalization, however, because paths of influence from different contexts can be dependent in ways that cannot be expressed using a single conventional ontology. This makes it difficult to computationally explore combinations of effects from different contexts. For this reason, features of business process modeling were also used in the design.
Business process models are similarly generalized (and for the same reason) but represent temporal sequences of events and their dependencies. This structure is able to capture an aspect of multi-contextual interaction, which is expressed as parallel layers known as ‘swim lanes’. Taxén [25, p. 23] observes that layers capture simultaneity whilst at the same indicating constraints that organize the field. Our method also uses layered ‘lanes’ to leverage those qualities.
Theories of Conceptual Change in cognitive science, such as that of Thagard [26], indicate how transitions between paradigms can be represented. In the original narrative modeling method, these theories were enriched by research into the detection of agency against contextual fields.
Those theorists include Herman [27], who refers to these parameters as foreground-background in narrative, and Einhorn and Hogarth [28], who understand it as a causal agent and field. In formal logic, Barwise and Perry [29] developed a method to handle shifts between contexts, Situation Theory, and Devlin [30] later gave it formal underpinnings, naming it ‘Layered Formalism and Zooming’ (LFZ) – both methods describe how general knowledge is channeled into a ‘focal situation’.
In our case, a focal situation is represented in the ‘Interpretation’ space. Information across the field is consolidated in a local situation model and sorted into qualities of context and agent, so that risk under current affordances can be identified.
VI. Discussion
This visualization method introduces representations of context and influence to the domain of workplace safety. In the example, a vehicle that was modified to reduce risk in one workplace becomes unsafe in another. The method enables risk to be flagged when a context shift activates new risk conditions, and tracks its dependent processes. When the model is adjusted, new consequences can be tracked to see the new outcome.
This ability to track, modify and see new cumulative effects would be particularly valuable if formalized to support a bespoke machine reasoning system. This system could take advantage of these models to evaluate combination of situations and flag risks that emerge. With sufficient complexity and scale, this could even alert analysts to risks not yet encountered.
To achieve this, there are two important steps for future work. The first is to record degrees of risk in the visualization. This would enable a distinction to be made between a minor risk (disturbing nearby residents) and serious risk (fatality), so an analyst could make appropriate decisions when weighing them against each other. Existing graphical devices could represent this by explicitly annotating that distinction, and the way the degree changes as a result of interactions.
Another important step would be to develop a drawing tool that makes it easier to create these models. Drawing these animations by hand in Apple’s Keynote presentation program is a painstaking process. If it was easier to experiment with different versions of structures or produce a variety of comparative examples, this work would be accelerated.
VII. Conclusion
The inclusion of context and aggregative effects in workplace safety models could make visible some of the paths to injury and fatality that currently cannot be seen. We included this new information in a workplace safety model using an approach that was originally designed for narrative analysis.
Our example explores conditions in which a reduction of risk in one workplace situation contributed to a fatality in another. When an alternate combination of factors is explored, the risk is averted.
Given the ability to trace risky outcomes through multiple processes, and explore the impact of new intervention, we conclude that workplace safety policies and regulations would benefit greatly from the addition of cumulative and contextual information to its representational methods.
This will be even more impactful if developed into a machine reasoning system that can scale this forecasting ability, using the combinatorial power of automation to foresee risks that might not have been anticipated by existing methods.
Acknowledgment
The research for this paper received funding from Griffith University and the Australian Government through Trusted Autonomous Systems (TAS), a Defence Cooperative Research Centre funded through the Next Generation Technologies Fund. TAS also receives funding from the Queensland Government.
Finally, this research receives in-kind support and access to subject matter expertise from the Downer Group.
References
[1] Safe Work Australia, “Key work health and safety statistics Australia 2021,” safeworkaustralia.gov.au, Canberra ACT, 2021.
[2] K. J. Devlin, “Modeling Real Reasoning,” in Formal Theories of Information, G. Sommaruga, Ed. Springer, 2009, pp. 234–252.
[3] D. Noble, “Multi-bio and multi-scale systems biology,” Progress in Biophysics and Molecular Biology, vol. 117, pp. 1–3, 2015.
[4] (EPA) USA Environmental Protection Agency, “Concepts, Methods and Data Sources for Cumulative Health Risk Assessment of Multiple Chemicals, Exposures and Effects,” EPA Publication No. EPA/600/R-06/013F, 2007.
[5] T. Niemeier, P. Williams, A. Rossner, J. Clougherty, and G. Rice, “A Cumulative Risk Perspective for Occupational Health and Safety (OHS) Professionals,” International Journal of Environmental Research and Public Health, vol. 17, p. 6342, 2020.
[6] J. De Landre, “About ICAM Australia,” ICAM Australia, 2022.
[7] Y. Holder, M. Peden, E. Krug, J. Lund, G. Gururaj, and O. Kobusingye, “Injury Surveillance Guidelines,” Centers for Disease Control Prevention, World Health Organization, Atlanta USA, 2001.
[8] D. Herman, “Genette Meets Vygotsky: Narrative Embedding and Distributed Intelligence,” Language and Literature, vol. 15, pp. 357–380, 2006.
[9] B. Cardier, “Narrative Causal Impetus: Situational Governance in Game of Thrones,” in Intelligent Narrative Technologies 7, Palo Alto, CA, 2014, pp. 2–8.
[10] J. McDougall, “Inquest into the death of James Leon Short,” Queensland Courts, Queensland Government, Southport, 2012.
[11] J. Reason and A. Hobbs, Managing Maintenance Error: A Practical Guide. Boca Raton, FL: CRC Press, 2003.
[12] D. R. Smith, D. Frazier, L. Reithmaier, and J. Miller, Controlling Pilot Error. New York: McGraw-Hill Professional.
[13] B. Cardier, “Unputdownable: How the Agencies of Compelling Story Assembly Can Be Modelled Using Formalisable Methods From Knowledge Representation, and in a Fictional Tale About Seduction,” University of Melbourne, 2013.
[14] H. T. Goranson, B. Cardier, and K. J. Devlin, “Pragmatic Phenomenological Types,” Progress in Biophysics and Molecular Biology, vol. 119, pp. 420–436, 2015.
[15] B. Cardier and H. T. Goranson, “System and Method for Ontology Derivation USPTO 14/093,229 (20140164298),” 2014.
[16] J. Joy, “Occupational safety risk management in Australian mining,” Occupational Medicine, vol. 54, no. 5, pp. 311–315, 2004.
[17] NSW Department of Industry and Investment, “Minerals industry safety and health risk management guideline,” Mine Safety Operations Branch, NSW Government, 2011.
[18] ISO Central Secretariat, “ISO 31000 – Risk Management,” International Organization for Standardization, 2021.
[19] E. Hollnagel and J. Speziali, Study on Developments in Accident Investigation Methods. International Atomic Energy Agency, 2008.
[20] P. Katsakiori, G. Sakellaropoulos, and E. Manatakis, “Towards an evaluation of accident investigation methods,” Safety Science, vol. 47, no. 7, pp. 1007–1015, 2009.
[21] P. Dodshon and M. Hassall, “Practitioners’ perspectives on incident investigations,” Safety Science, vol. 93, pp. 187–198, 2017.
[22] B. Ale et al., “Accidents in the construction industry in the Netherlands,” Reliability Engineering & System Safety, vol. 93, no. 10, pp. 1523–1533, 2008.
[23] A. Adamski and R. Westrum, “Requisite imagination,” in Handbook of cognitive task design, CRC Press, pp. 193–220.
[24] T. Gruber, “Towards Principles for the Design of Ontologies Used for Knowledge Sharing,” International Journal of Human-Computer Studies, vol. 43, no. 5–6, pp. 907–928, 1995.
[25] L. Taxén, “An inquiry into Knowledge Integration and Innovation Management,” presented at the 2nd Advanced KITE Workshop, Linköping, Sweden, 2010.
[26] P. Thagard, Conceptual Revolutions. Princeton University Press, 1992.
[27] D. Herman, Story Logic: Problems and Possibilities of Narrative. Univ of Nebraska Press, 2002.
[28] H. Einhorn and R. Hogarth, “Judging Probable Cause,” Psychological Bulletin, vol. 99, pp. 3–19, 1986.
[29] J. Barwise and J. Perry, Situations and Attitudes. Cambridge, Massachusetts: MIT Press, 1983.
[30] K. J. Devlin, Logic and Information. Cambridge University Press, 1995.