An artificial intelligent paradigm for systems safety in the cockpit

This paper discusses the use of an artificial intelligent paradigm consisting of three intelligent autonomous agents whose purpose is to monitor the flight crew and the aircraft. Two of these agents will monitor the path of the aircraft, one armed with prior knowledge of how planes tend to land at a particular airport, the other with the ability to extrapolate forward from the plane’s current position in order to identify potential dangers. The third agent will monitor the flight crew’s behaviour for potentially dangerous actions or inactions. In effect when combined with one of the other agents it will operate as a situation awareness agent. This paper highlights recent developments in the situation awareness agent. Introduction Over the last five years the civil aviation industry has experienced a large reduction in aircraft accidents. This period has probably been the safest on record. However, that statement needs to be qualified; because it really depends on how one defines and measures safety. The risks have not disappeared and there may have been as many ‘mistakes’ as normally occurs in a 5 year period. It may be that risks are being managed better or it may just be a statistical variation in the accident rate. However, it can be generally said that over the last few decades, with the introduction of technically innovative designs of on-board systems and crew training in team management, air travel has become safer. But, except for the last five or six years, there appears to be a small, but significant, remnant of air accidents. Helmreich and Foushee, (1993) have suggested that 70% of these remnant air accidents are due to flight crew actions or in some case inactions. This is despite the fact that pilots are extremely technically competent and well trained in crew resource management (CRM). Flight crews undergo regular flight checks to determine their competency in both the human and technical areas of line operations. As a consequence airline flight crews are highly skilled to operate in the modern cockpit environment. This includes both the technical and human environments. This raises the question as to why these accidents happen and, perhaps even more disturbingly, why they continue to happen, albeit at a very low level of incidence. This paper discusses the problem of contemporary aircraft approach-and-landing accidents and proposes a paradigm consisting of three intelligent software agents whose task it is to jointly monitor the aircraft and its flight crew (Thatcher, Jain & Fyfe, 2004a, 2004b, 2005). The trio of intelligent agents within the proposed paradigm are organised as follows: • Two agents (the Pattern Matching Agent and the Prediction Agent) will be physically situated onboard the aircraft. • The remaining agent (the Anomaly Detection Agent) will be physically situated at the destination aerodrome. On board the aircraft the Prediction Agent will be able to predict the aeroplane’s future position using its current three dimensional position and its vertical and horizontal velocity variables. The predicted future position of the aircraft will be used to identify potential terrain threats on the approach. The other on board agent, the Pattern Matching Agent will monitor the flight crew’s behaviour and determine if the flight crew are losing situation awareness on the landing approach. The interaction between the two onboard agents will form, in essence, a situation awareness agent. The agent positioned at the airport will monitor the flight path of the aircraft as it commences its approach. This agent, the Anomaly Detection Agent will have knowledge of typical aircraft approach profiles for that particular aerodrome. This paper will outline the proposed knowledge-based intelligent landing support paradigm with particular emphasis on the agents on board the aeroplane. Combined these agents form the Situation Awareness Agent. Controlled flight into terrain and approach and landing accidents A Flight Safety Foundation (FSF) report concluded that from 1979 through 1991 Controlled Flight Into Terrain (CFIT) accidents and approach-and-landing accidents (ALAs) accounted for 80% of the fatalities in commercial transport-aircraft accidents (Flight Safety Foundation, 2001) The FSF Approach-and-landing Accident Reduction Task Force Report (Khatwa & Helmreich, 1999) concluded that the two primary causal factors for such accidents were ‘omission of action/inappropriate action’ and ‘loss of positional awareness in the air’. Positional awareness or as its more commonly termed situation awareness is defined by Endsley (1987, 1988) as “the perception of the elements in the environment within a volume of time and space, the comprehension of their meaning, and the projection of their status in the near future.” One solution to the problem is to increase the level of automation onboard the aircraft. However, automation has associated human factors problems and may lead to a decrease in situation awareness amongst the flight crew. We may simply be trading one type of error for another. Endsley and Strauch (1997) maintain that ‘despite their high reliability, accurate flight path control, and flexible display of critical aircraft related information, automated flight management systems can actually decrease’ a flight crew’s ‘awareness of parameters critical to flight path control through out-of-the-loop performance decrements, over-reliance on automation, and poor human monitoring capabilities.’ Further, pilots can in some respects configure the Flight Management System to present a view of the physical world which supports their interpretation of the world or their mental model of the current operating environment. In this three agent paradigm it is proposed to reduce both the human error rate and the automation induced error rate by incorporating intelligent agent systems. However, it must be acknowledged that this may induce other forms of error into the system. It must also be considered that if the agents are truly intelligent there may be similar errors to that of the flight crew. Enhanced ground proximity warning system Following a series of Controlled Flight Into Terrain (CFIT) accidents such as the American Airlines Boeing 757 that struck a mountain while on descent for a landing at Cali, Colombia, on December 20, 1995 the Flight Safety Foundation CFIT Task Force recommended that early model Ground Proximity Warning Systems (GPWS) be replaced by Enhanced GPWS (EGPWS) or Terrain Awareness and Warning Systems (TAWS) which have a predictive terrain hazard warning function. (Khatwa & Helmreich ,1999). In response to the report the FAA mandated in 2001 that all heavy transport aircraft be fitted with EGPWS and further, that all turbine aircraft with 10 or more passenger seats be fitted with EGPWS from 2003. The American Airlines 757 aircraft had been fitted with a functioning GPWS however, it did not prevent the accident. The reason that the GPWS had not prevented the accident was the lateness of the warning to the crew due to the lack of a predictive element in the software. In comparison the EGPWS compares the aircraft’s position and altitude derived from the Flight Management and Air Data computers with a 20MB terrain database stored in the EGPWS. In the terrain database the Earth’s surface is reduced to a grid of 9x9 km squares, each with an individual height index. In the vicinity of airports the grid resolution is increased to squares of 400m x 400m. The height index and the aircraft’s predicted 3 dimensional position 20 to 60 seconds into the future are compared to see if any conflict exists. If it does the EGPWS displays an alert or warning to the flight crew. Other than to initially alert the pilots of ‘TERRAIN’ up to 40-60 s before impact or warn the pilots to ‘PULL UP’ up to 20-30 s before impact it does not offer any other solution to the potential terrain conflict situation. This research aims to extend the EGPWS by using three intelligent software agents, described below, which can plot a course around, or over, possible conflicting terrain and present a solution to the pilot on the cockpit display system (or, perhaps controversially, as input to the autopilot). Intelligent agents Wooldridge (2002) describes an intelligent software agent as a computer program that performs a specific task on behalf of a user, independently or with little guidance. It performs tasks, tailored to a user’s needs with or without humans, or other agents, telling it what to do. To accomplish these tasks, it should possess the characteristics such as learning, cooperation, reasoning and intelligence. By analogy, a software agent mimics the role of an intelligent, dedicated and competent personal assistant. In this application Thatcher, Jain and Fyfe (1994a) propose developing three agents, one ground based and the other two aircraft based, which will aid pilots during the critical approach and landing phase. In effect the two onboard agents will act as another flight crew member who has situational awareness.