The human society has evolved tremendously since the invention of the steam engine. In almost every decade new inventions and technologies have reshaped the way people interact and do business. Such tremendous change has brought comfort and efficiency in various aspects of human lives, yet one important factor still concerns all till today is the safety of socio-technical systems. It is indeed beyond question that new technology brought significant benefits to the society; at the same time, it brought new modes of accidents which were previously unknown. For example, car accidents were not present before the invention of the diesel engine; air crash never happened before the invention of aircraft; spacecraft never exploded before its invention; so many such accidents can be cited as examples.

In an effort to understand the science behind accidents this research paper investigated the existing accident theories and models. A literature review revealed that accident theories and models can be classified in various ways. Some classification examples can be found in Qureshi [1], Khanzode et al. [2] and Awal and Hasegawa [3,4]. However, in this study the following classification of accident theories and models are considered:

1. Statistical analysis and trends

2. Risk analysis

3. Domino theory

4. Epidemiologic theory

5. Control and System theoretic model

The fundamental points of this group of accident theories and models are discussed in the following sections. Later a comparative study is conducted over these accident theories and models. This comparison reveals the strengths and weaknesses which indicate the importance of further research on accident prediction and analysis.

The scientific approach of studying accidents started in England at the beginning of 20th century when Vernon [5] published his extensive study on the causation of industrial accidents. Vernon considered fifty thousand industrial accidents collected from four different factories across the United Kingdom and produced a statistical analysis of hourly variation of distinct categories of industrial accidents. This research showed a systematic approach of identifying various accident causation factors. Vernon classified factors into two main headings: (i) factors of personal origin and (ii) factors of external origin.

Since then significant research on accident studies has been conducted where the statistical tools have been used to identify various correlating factors that contribute towards the causation of accidents. Statistical analysis reveals share or percentage of different factors, the relationship among various factors and overall trends of accident causation against different kinds of timelines. Awal [6], Awal et al. [7] and Hossain et al. [8] may give some insight into this topic. Figure 1 shows an example from Awal et al. [7] where the monthly variation of total maritime accidents in Bangladesh was shown along with ship collision accidents.

Figure 1: Monthly trend analysis of maritime accidents in Bangladesh.

It was interesting to note that the total number of accidents remains below the average during the months of December and January; whereas the collision accidents increase above the average line. Therefore, during these two months, collision type accidents occur more than any other type of accidents. Hence, such a statistical analysis may assist decision makers to allocate resources on collision prevention rather than focusing on other types of accidents. Another example of statistical accident analysis can be found in Massie and Campbell [9] which is shown in Figure 2. In this example, the fatal involvement of men of different age groups in car crashes is shown. It was observed that with the increase in age of drivers the fatal consequences decrease significantly. Therefore, it can be comprehended that measures taken on younger people may significantly reduce fatality due to car crashes. These examples indicate that the goal of statistical analysis is to identify some trend(s) correlating with some probable factors. Various kinds of hypotheses can be drawn and tested if sufficient statistical data is available.

Figure 2: The observed relationship between age and fatality in car crashes in the United States using one year (1990) data.

One of the most applauded branches of statistical analysis of accidents is risk analysis, which gained popularity after the introduction of Fault Tree Analysis (FTA) [10,11] and Event Tree Analysis [12]. Since then a significant number of studies has been conducted on the development and application of risk analysis in solving practical problems. In traditional risk analysis, the principal task is to identify distinctive threats and simultaneously determine consequences of the respective threats if they are to occur in reality. There are two types of risk analysis: (i) quantitative risk analysis and (ii) qualitative risk analysis. Risk analysis is popular in solving real engineering problems and academic exercises because of its technical facility and wide range of applicability.. An example of ship collision risk analysis using fault tree is shown in Figure 3. This example reveals that the probability of ship collision with another ship can be computed in a fault tree when all the probability values of the connected events are known [13]. Risk analysis may also provide a picture of total accident occurrence probability as shown in Figure 4. The figure shows the probability of tanker ship collision taking place in the Gulf of Finland for different types of traffic flow in one year [14]. The results obtained from such analysis can only provide some guidance on the chances of occurrence of an event, here, in this case, the accident event. Similar studies on wide range of scientific and engineering applications can be found (e.g. safety of nuclear power plants [15], underwater tunnel excavation [16], and many more). These studies are useful in the policy level where Risk Control Options (RCOs) can make an impact in reducing the risk values. However, risk analysis cannot provide the advantage of knowing how an accident may take place. For the occurrence of an accident event, it relies completely on a subjective judgment of the expert who employs this analysis.

Figure 3: Fault tree applied to derive causation probability for ship collision in good visibility.

Figure 4: The probability distribution of the number of tanker collisions in the whole Gulf of Finland in 2015 for the three traffic scenarios "slow", "average" and "strong".

Domino theory is one of the earliest theories of accident causation, which utilized metaphoric example of domino effect. Heinrich [17] utilized the concept of domino effect where a series of domino arranged in such a manner that the fall of the first domino creates a cause-effect sequence which eventually results in fall of the last domino. Therefore, Heinrich proposed that accident occurs in a chain of events, which can be traced back from injury to ancestry and social environment.

Heinrich elaborated that the individual fault can be related to other factors in sequence, just like a domino. There are five dominoes according to this theory as shown in Figure 5. Heinrich explains that undesirable personality traits can be passed along through inheritance or develop from person’s social environment and both inheritance and environment contribute to faults of a person. This can be considered as the first domino. The second domino deals with worker personality traits. Heinrich explains that inborn or obtained character flaws contribute to accident causation. According to Heinrich, natural or environmental flaws in the worker’s family or life cause these secondary personal defects, which are themselves contributors to unsafe acts or the existence of hazardous conditions. The third domino is the direct cause of incidents - the unsafe act. Heinrich [17] defines four reasons why people commit dangerous acts: (i) improper attitude, (ii) lack of knowledge or skill, (iii) physical unsuitability and (iv) improper mechanical or physical environment. Heinrich later subdivides these categories into ‘direct’ and ‘underlying’ causes and concludes that combination of multiple causes creates a systematic chain of events that leads to the accident. The goal of domino theory is to establish a linear cause-effect relationship among various social and individual factors using five metaphoric dominoes. Domino theory has been applied in various industrial applications such as process industry [18], safety of civil engineering construction sites [19] and many more.

Figure 5: The fundamental concept of domino theory.

The application of domino theory in various fields of science and engineering probably has led to the evolution of linear cause-effect analysis of accidents. Such analyses encompass various forms, for example, root cause analysis, fishbone diagram and others. These tools or methods have gained popularity and proved to be useful in the safety analysis of Biogas production [20] and safety of electric transformers [21].

Gordon in 1949 first proposed the epidemiologic theory of accident causation [22]. Gordon considered accidents as an ecologic problem. In this study, it is stated that the causative factors in accidents have been seen to reside in agents, in the host, and in the environment, as shown in Figure 6. The mechanism of accident production is that process by which the three components interact to produce a result, the accident. Therefore, the cause of accident comes from the interaction between the host, agent, and environment. The hypothesis is if home accidents are primarily a public health problem then the problem is reasonable to be approached in the manner and through techniques that have proved useful for other mass disease problems.

Figure 6: Host, agent and environment triad in accident causation.

Haddon [23] in 1968 proposed a 2D dimensional matrix for injury control which was mainly developed for the car crash. Such matrix helps to identify the host, agent and environmental factors in temporal order (prior to the event, during the event and post-event) and helps to identify preventive measures. Haddon later in 1970 introduced ten strategies for reducing losses which were based on the energy transfer concept [24]. The reason behind this energy transfer concept is because that major class of ecologic phenomena in nature involves the transfer of energy.

In 1990, Reason [25] proposed that there are latent human failures which result in accidents without any visible causes. The proposal included the following three concepts: (i) latent failure, (ii) local triggering event, (iii) system defenses. The study discussed the differences between active and latent human failure and a framework for the dynamics of accident causation. Later on, Reason [26] developed and proposed the Swiss cheese model and the organizational Accident Model shown in (Figures 7 and 8). Organizational accidents occur within modern complex technological societies having multiple causes. The accidents involve many people operating at different levels of their respective companies. Hazard cause losses and barriers are there to prevent. Like Swiss Cheese, there are holes in the barriers. When all the holes align then hazards pass through and cause losses. Each barrier represents each level of organizational defenses against losses.

Figure 7: Swiss cheese model of accident causation.

Figure 8: Reason’s organizational accident model.

In 1970, Suchman [27] proposed the social deviance hypothesis for accident causation. Suchman considered accident as a social control problem and rejection of social constraint is considered as the cause of an accident. Kjellen and Larsson [28] proposed Deviation Concept for occupational accident control. The proposal stated that deviation from the norm causes accidents in a production process. Rasmussen [29] and Rasmussen and Svedung [30] discussed risk management in dynamic society and proposed AcciMap for accident investigation. Rasmussen’s work fundamentally considers accidents as a control problem, which utilizes control engineering metaphors in a social context. Rasmussen presented an interesting control structure which is based on control architecture applied at the different level of social structure. This is shown in Figure 9. Branford’s research [31] can be considered as an example that applied AcciMap and studied loss of control at various levels of the Uberlingen mid-air collision.

Figure 9: Rasmussen’s concept of control of hazardous activity in nested social structure.

Perrow [32] proposed the term ‘Normal Accident’ which became very popular in various high-tech industries. According to Perrow, a normal accident is a characteristic of a system. Given the characteristics of the system, multiple and unexpected interactions of failures are inevitable. The interactive complexity and tight coupling the system characteristic inevitably produces an accident; therefore, it is called ‘Normal Accident’ or ‘System Accident’. The premises of this idea are: (i) people make mistakes, (ii) big accidents begin from small beginnings and (iii) many failures originate due to organization/technology. Hollnagel and Goteman [33] proposed Functional Resonance Accident Model (FRAM). The idea suggested that accidents occurred due to functional resonance within a system. The basic concept of FRAM is shown in Figure 10. Leveson [34] introduced System-Theoretic Accident Model and Process (STAMP) which is essentially a model of control system applied in the social structure. She suggested that accidents occur when there is a lack of constraints. In general, control and system theoretic models attempt to discover weaknesses within the concerned system and prescribe various control parameters to prevent accidents.

Figure 10: Basic of Functional Resonance Accident Model (FRAM).

In this paper, five groups of accident theories and models were studied. The paper systematically presented the characteristics of each group and discussed their attributes with necessary figures. In order to summarize the findings of this study, a comparative sum-up is produced in this section. Three different questions were asked of each group of accident theories and models and they are analyzed in tabular forms as shown in (Tables 1-3).

Table 1: Comparison of accident theories and models: What causes the accident?

Table 2: Comparison of accident theories and models: How the analysis is done?

Table 3: Comparison of accident theories and models: How to prevent an accident?

The first question asked to the group of accident theories and models in ‘what causes the accident? This is discussed in Table 1. In answer, it was found that each group is significantly different from the other. The statistical analysis and trends group is unable to provide any simple answer. It has been found that according to this group, an accident can be caused by various factors and the purpose of this group is to identify the correlation between the accident and the accident causing factors. Similarly, risk analysis group is unable to provide any simple answer; rather this group of theories and models can only compute risk value and assist decision makers in controlling risk and thereby allocating necessary resources. The domino theory provides a simple answer. Accidents occur due to faulty act and this faulty act is one of five dominos which fit into the concept of linear cause-effect relationship. Interestingly, the cause of accident according to epidemiologic theory is the interaction of host, agent, and environment. The accident comes out of this interaction. Lastly, the cause of accident according to control and system theory group is the lack of control and lack of understanding of the system dynamics. This group is comparatively new and requires further developments.

The second question asked to the accident theories and models is ‘How the analysis is done?’. This is discussed in Table 2. In answer, it is found that statistical analysis and trends group requires a bulk amount of data for various statistical modeling. Therefore, in cases where the data is rare, such kind of theories and models are less useful. Similarly, the risk analysis group requires sufficient amount of data in order to utilize various methods of risk modeling. Once the risk model is constructed, Risk Control Options (RCOs) are applied and compared with a base value to undertake preventive measures. The domino theory group requires subjective judgment and analysis to define the dominoes. This theory/model is not handicapped by lack of numerical data. The epidemiologic theory group necessitates subjective judgment to identify the host, agent, and environment. The interaction among these three is also a matter of subjective analysis. Finally, the control and system theoretic group needs similar to the previous two groups, i.e. subjective judgment. The analysis identifies system components and role of control over the system.

The third question asked to the accident theories and models is ‘How to prevent an accident?’ The statistical analysis and trends group and risk analysis group require subjective judgment and decisions to apply preventive measures. The decisions of preventive measures are taken based on the correlation and risk values obtained from the analyses. The domino theory group suggests removing any of the dominoes so that the cause-effect chain is no longer continued. In many real-life examples this theory, however, is proven impractical. The epidemiologic theory group suggests stopping the necessary interaction among the host, agent and the environment in order to prevent accidents. In many organizational accidents, this group of theory appears to be successful to explain the accident phenomena. Finally, the control and systemtheoretic group of accident theory and models impose constraints to prevent accidents. According to this group keeping necessary control is essential in preventing accidents. Table 3 briefly discusses these issues.

The accident theories and models studied in this paper are fully dependent on the subjective judgment of human experts and in many cases are unable to deduce how an accident may take place. The current need in the discipline of safety science is to develop new techniques that can predict ‘how’ an accident may take place with reasonable accuracy. Such deficiency is identified in previous studies by Awal [35], Awal and Hasegawa [36-41]; these studies developed a new concept called Logic Programming Technique (LPT).

LPT is a method of logical deductions which utilize heuristics to search through given knowledge and attempts to discover ‘how’ accident may take place. The fundamental principle of LPT is logical deductions using deductive arguments. Propositional logics are utilized to develop logic worlds and find out how accidents take place within the world. Propositional logics are simple sentences which have one or more premise(s) and one conclusion. An example is shown in Figure 11. This simple format allows modeling systems of various disciplines. This is a significant advantage while studying accidents in the multidisciplinary platform.

Figure 11: An example of a propositional logic.

Awal and Hasegawa [38,40,41] discussed the advancement of LPT into the agent-based perception-action technique. In this technique, all agents search through respective perceptions and actions. Thereby, logical deductions reveal all the possible sequence of events that may take place. Some of these events may lead to accidents which are of particular interest to achieve safety. Hence, Logic Programming Technique (LPT) can automatically deduce ‘how’ accidents can take place. Such deductions are generally difficult to produce by human judgment or subjective analysis alone because of the complexity and size of the problem space. Traditional accident analysis also becomes very difficult when there is a change in the system; whereas logic programming can handle system changes easily. For example, an addition of an event or an agent is simply done by adding arguments in the logic program; while in traditional accident analysis (e.g. fault tree analysis) such change complicates the total structure (tree in this case) and requires reconstruction of the structure again. Also in traditional accident analysis, the (e.g. fault tree) the combination of events is known. However, in Logic Programming Technique (LPT) the combinations of events are deduced automatically. This gives Logic Programming Technique (LPT) a significant advantage over other accident analysis techniques. Figure 12 shows this concept graphically.

Figure 12: Difference between traditional accident analysis and Logic Programming Technique (LPT).

This paper discussed and revealed the underlying science of accident causation in complex engineering and socio-technical systems. A number of accident theories and models have, therefore, been presented and categorized in this study. The study reveals that that over the years, particularly the past century, researchers from various fields, including engineers, medical doctors, psychologists, and other professionals have studied accidents and attempted to theorize the science of accidents. It has been found that the accident theories have transformed from a single root cause to linear cause-effect models, from epidemiologic concepts to organizational accident models and evolution of control and system theoretic concepts. This transformation and evolution of a broad range of accident theories/models suggest that understanding the science of accidents requires a wide perspective and vast knowledge domain. The study revealed that the problem space of safety science is diverse and incorporation of knowledge from multiple disciplines is necessary.

The application of LPT in complex industries such as automobile industry can be highly beneficial and it appears that the potential of application is immense. For example, LPT can be applied in automobile production facility where it can deduce possible event sequences that may lead to production failures. Also, LPT can be applied in the most modern driverless automobiles to deduce the errors in software that may lead to possible catastrophes. Having said the advantages and future potentials of LPT, there are some shortcoming of this method which needs to be studied further in order to utilize it in solving accident problems in complex socio-technical systems.