During the September 1997 meeting of the Icarus Committee, an expert team affiliated with Flight Safety Foundation, members agreed to form the Safety Index Working Group to develop a safety-metrics model with which an airline can monitor and measure operational safety performance. The model developed into a software system called the Flight Operations Risk Assessment System (FORAS).1
The goal in developing FORAS was to create a quantitative index for proactively assessing aviation risk, focusing on the recognition of risk factors involved in aviation safety instead of emphasizing accident rates. Full technical details were published in 2009.2 FORAS developers adopted a mathematical model to synthesize a variety of inputs (risk factors), including information on crew, weather, management policy and procedures, airports, traffic flow, aircraft and relevant operations.
In FORAS, risk assessment is tackled with a divide-and-conquer strategy. A given risk category is broken down into a small set of sub-risk categories, and each sub-risk is further broken down to a set of risk measures. This process terminates when directly measurable risk factors (operational data) associated with a flight are obtained. This scheme enables knowledge domain experts to deal with a small number of risk factors at a time, reducing the complexity of their work.
The FORAS model contains a great number of risk factors; to illustrate, only a small portion of the FORAS model dedicated to approach and landing risk value (ALRV) assessment is presented (Figure 1). The ALRV assessment is broken down into three sub-groups: crew functionality, aircraft functionality and sector threat.
These sub-risks are further subdivided into more detailed sub-risks. For example, crew functionality includes inter-crew communication, pilot experience and pilot fatigue. As noted, the automated breakdown process continues until all measurable risk factors are obtained — in this example, experience pairing, rank composition and communication proficiency. These data are available from the airline’s crew and roster databases.
FORAS Model Development
In 2005, EVA Airways collaborated with Michael Hadjimichael (then working at the U.S. Naval Research Laboratory) to develop the first practical application of the FORAS model to monitor the approach and landing risk of each flight. An EVA Airways team interviewed pilots, safety managers, dispatchers and maintenance engineers and discussed the risk factors that contribute to approach and landing risk and the relationships among these risk factors. The FORAS model for assessing this risk then was successfully implemented at EVA Airways as an online system.
Currently, the approach and landing risk of each flight is computed by FORAS two hours and 30 minutes before its departure, and the resulting risk value is shown on a Web-based interface. During its years of experience in using the FORAS model, EVA Airways has gained insights about the causal relationships among risk factors and their contribution to approach and landing risk. This experience also identified a need to revise the original model.
In 2009, EVA Air launched a research project with the Department of Information Management of Tamkang University, to revisit the risk factors and develop a new software system that allows users to construct a new model or change an old model easily. The new software system has been implemented online at EVA Airways and offers great flexibility and convenience whenever the system administrator needs to revise the model.
A Two-Part System
The FORAS software model application is composed of two parts, called the back-end system and the front-end system (Figure 2). The back-end system provides a user-friendly interface for the user to construct a FORAS tree as required by the model (Figure 3).
Each node in the tree represents a risk factor and is associated with a set of parameters that specify the node’s characteristics. The information is used in online reporting and risk assessments. The back-end system uses information from the model to identify legal input data (that is, data conforming to software-embedded rules), parameters for which missing input data are acceptable and parameters that are controllable. When missing data are allowed, risk-index computation proceeds with default values. Controllable parameters are variables that may be controlled by flight dispatch or by a scheduling action to decrease flight operation risk. These possibilities include crew and aircraft factors. Weather is an example of an uncontrollable factor.
The risk-assessment functionality in the FORAS model is obtained by a series of inference procedures moving upward from the bottom level of the user-constructed tree to the top. Taking Figure 1 as an example, the risk value of inter-crew communication is inferred from three risk factors — experience pairing, rank composition and English proficiency; in turn, the risk value of crew functionality is inferred from inter-crew communication, pilot experience and stress level.
Finally, the ALRV is inferred by the FORAS model from crew functionality, aircraft functionality and sector threat. This logical inference procedure is based on a conditions-consequence relation, also referred to as a causal relation. For example, in inferring the inter-crew communication risk (a consequence), experience pairing, rank composition and English proficiency are the determining conditions.
In each airline’s FORAS model, the relation between a condition and a consequence is expressed by rules. A rule is used to describe the degree of the resulting risk under various conditions of its causes, and such conditions are assessed in a linguistic manner.
As an example of the simplified user-interface language of the FORAS model, a typical rule — in this case, part of a logical assessment of the inter-crew communication risk — would be, “If T1 is experienced, T2 is ideal and T3 is poor, then C2 is 4.”
In plain English, that means, “While planning a specific flight, if the airline safety specialists rank the condition called flight crewmember experience pairing (T1) as ‘experienced’ (from their predefined scale of possible ratings); they rank the condition called flight crew rank composition (T2) as ‘ideal’; and they rank the condition called flight crew communication proficiency (T3) as ‘poor,’ then the FORAS model must use the value “4” wherever the inter-crew communication risk value is required in the model’s trees and algorithms. In this way, that risk value — selected by the specialists using a predefined scale from 1 to 10 (the greater the number, the higher the risk) — will be applied consistently by the FORAS model, representing the specialists’ overall perceived risk value for inter-crew communication.
The formulation of such rules is based on experts’ knowledge and group decision making. Rules of this type have the advantage of being easy to express and understand for the knowledge-domain experts from whom the knowledge base is derived. In particular, evaluating conditions in a linguistic manner alleviates the difficulty of quantifying an uncertain or subjective judgment that doesn’t lend itself to precise numerical expression.
The back-end system contains a rule-setting module, where the setting of rules can be in either a rule format or a table format (Figure 4).
Rules Into Equations
The last thing needed to make such rules work together in risk assessment is to define mathematical equations that quantify the linguistic terms in the rules. This definition of equations is based on the values of the conditions, which the FORAS model’s designers call the membership function of the risk factor. Membership functions interpret the airline’s plain-language linguistics terms (such as “high experience”) as specific numerical input values.
The back-end system in the FORAS software also contains a module for membership function–setting, where typical functions are provided and presented in a graphic form (Figure 5).
After a FORAS tree has been built by the airline using the back-end system, it is “published” to a central database. The front-end system then retrieves the tree and its associated parameter settings (input data) from the database to compute the risk value of a flight. This risk assessment of a flight is computed two hours and 30 minutes before its takeoff.
In a snapshot of the online risk report of actual flights (Figure 6), various “traffic light–style” signals clearly indicate the risk status of a flight (a green light means Normal, yellow means Warning and red means Alert). If desired, the user can click on the computer display of the flight to read a detailed risk report that lists the risk values of all nodes in the tree associated with that flight. On this screen, users can also request further analysis, including a drill-down analysis, a trend analysis of the risk of interest and a critical risk factor analysis that identifies which factor contributes most to the risk of interest (Figure 7).
EVA Airways’ Risk Assessment
EVA Airways so far has constructed two FORAS models to construct two risk assessment models, one for routinely assessing approach and landing risk and one for assessing departure risk. Both models are run online for about 200 flights every day worldwide.
Based on FORAS reports, safety managers evaluate the overall level of risks for these aspects of their operations, and analyze the effects of management decisions on this risk level. With the trend analysis function of the system, managers can track various risks over time. The critical, risk factor–identification function assists the safety managers to identify the risk factor that contributes most to the risk of interest. Theoretical concepts and implementation issues of this system were presented at an FSF International Air Safety Seminar in 2011.3
The EVA Airways’ software application of the FORAS model is installed on a cloud-computing platform to enable sharing FORAS models within the airline industry in the future. The plan is to establish a community-computing, cloud-based system, run by a third-party non-profit organization, with a multi-tenant infrastructure shared among several organizations with common computing interests.
The authors envision customized FORAS models constructed and maintained on the cloud platform, and that individual airlines will access the risk-assessment service by securely sending encoded flight data to the cloud platform. For data security, the cloud platform would not keep the flight data or their computational results. Users could customize their risk models (including trees) by setting model parameters via a Web interface to the front-end system. For airlines that intend to operate the FORAS model on their own, a private cloud-computing platform can be built based on the same architecture.
FORAS development was originated by Flight Safety Foundation, and originally was funded by the U.S. National Aeronautics and Space Administration, the U.S. Naval Research Laboratory and EVA Airways.
Currently, EVA Airways plans to share the latest FORAS model version with interested airlines to promote this proactive and quantitative safety management concept and tool. This promotion of the FORAS model will be non-commercial, in which airlines that are interested in acquiring the system will pay an amount to the FORAS Association based on their fleet size.
Plans call for the FORAS model fund to be managed by a committee of trustees. The fund will be utilized to set up FORAS scholarships and FORAS awards, and to sponsor the future development of the FORAS model, as well as other aviation risk management initiatives.
- Hadjimichael, Michael; Deborah M. Osborne; David Ross; Diana Boyd; and Barbara G. Brown. “The Flight Operations Risk Assessment System: Proceedings of the 1999 ASE Advances in Aviation Safety Conference.” Daytona Beach, Florida, U.S., pp. 37–43.
- Hadjimichael, Michael. “A Fuzzy Expert System for Aviation Risk Assessment.” Expert Systems with Applications, 36, pp. 6512–6519. 2009.
- Ho, D.C.; H.-J. Shyur; C.-B. Cheng; W.-H. Yeh; S.-C. Kao. “The Enhancement and Implementation of the Flight Operations Risk Assessment System (FORAS).” In Flight Safety Foundation, Proceedings of the 64th International Air Safety Seminar, Singapore, 2011.