CC BY
10ТЕХНИЧЕСКИЕ НАУКИ UDK 631.4:574
METHODS AND ALGORITHMS OF ENSURING SAFETY OF FLIGHTS IN CIVIL
AVIATION
ISMAYILOV ISMAYIL MAHMUD
professor, department of aerospace information systems National Aviation Academy, Baku,
Azerbaijan
BINNATALIYEVA TURANA VAHID
lecturer, department of aerospace information systems National Aviation Academy, Baku,
Azerbaijan
Annotation. From the point of view offlight safety, the article considered the issues of analysis of flight data, prices for which are fixed during the flight of the aircraft, in order to detect, record and document undesirable events occurring during the flight. The article discusses the methodology of the intellectual analysis of the multiple flight data which were recorded during the flight of the aircraft.
Keywords: Flight data, flight data analysis, flight event, Data Mining, methods of intellectual analysis.
Introduction. The safety of civil aviation is the main goal of the International Civil Aviation Organization (ICAO) and significant progress has been made in this area in recent years. However, there is still a need for further improvement of the measures taken, since any progress in this area has a significant impact on the improvement of aviation security. SA (Situational Awareness) is an important component of human information processing and is of great importance in pilot decision-making processes.
In aviation, the "human factor" is considered to be the most important condition affecting the level and state of flight safety of any kind of aircraft. A human is the most adaptable and important element of the aviation system and, simultaneously, the most vulnerable element in case of unforeseen situations.
The comfort and reliability of aircraft is growing each year, while many flight stages are performed automatically under strict control by the pilot. However, the ICAO reports that every three out of four aviation accidents have been occurring due to the pilot fault for many years up to now. The measures taken and being taken by ICAO contributed to a reduction in the total number of aviation accidents, nonetheless their causality remains the same, i.e., at least 80% of all aviation incidents, accidents and catastrophes still occur due to erroneous and incorrect actions of aviation personnel, both in the air and on the ground [1,2].
During the flight, the aircraft crew (AC), receiving data from the subsystems of the flight and navigation complex (FNC) and from multifunctional indicators, controls the aircraft by deviating the aircraft control sticks [3]. Moreover, this interaction is influenced by the features associated with the psychophysiology of the pilot. Since the crew's ability to deflect special situations arising on board is limited, it is necessary to introduce intellectual and virtual components into the FNC, which accumulate the experience of the behavior of real experts in the field of aircraft navigation and piloting in special situations. This circumstance makes the problem of developing onboard systems equipped with these components urgent, which reduces the psycho-physiological load on the aircraft crew. This allows to conclude that the further development of FNC is closely related to the introduction of crew support systems and tools, situational awareness systems and further intellectualization of aircraft control (AC) [4].
Problem statement. An important condition for the development of civil aviation is the ever increasing requirements for ensuring flight safety. Flight data (FD) processing in civil aviation plays an important role in improving the safety and efficiency of air transport.
Initially, the main purpose of recording FD was to assist aircraft accident/incident investigators in the analysis of accidents, especially in those involving the death of all crew members. Also, the results of the analysis of the recorded FD provide a better understanding of what a safe flight should look like. Regular evaluation of recorded FD provides a wealth of information regarding flight safety, aircraft system and engine performance. The daily recorded accident-free flights' FD are as valuable data as accidents and incidents. In addition, analysis of the FD of these accident-free flights can help predict potential threats to safety before an incident or accident occurs. [6]
The FD is the only objective source of information about the crew's performance throughout the flight. Therefore, the systematic control and assessment of the crew's flight activity based on the processing of the FD ensures a significant increase in the professional training level of the crew. In addition, ground processing of flight information plays a leading role in determining the condition of aviation equipment and solving its diagnosis.
When an aircraft experiences a potentially catastrophic failure in flight, it is not difficult to say that something went wrong that could have caused it. However, even if the aircraft reaches its destination without incident, it may still have violated the boundaries of safe operating procedures. Figuring out what those boundaries are, what normal operation looks like and devising ways to effectively monitor them can be a hard task. As a rule, aircrafts are equipped with various sensors for recording and analyzing flight data.
During each flight of an aircraft a large amount of data is collected: everything from instrument positions to sensors and audio recordings is collected and stored for later analysis.
At this time, it is assumed that the more values of flight data the expert has, the more different conditions will be recorded in the database, as a result, more flight events will be detected and the expert will be able to more accurately assess the situations that lead to these events. Although increasing the number of flight data records and recorded flight data sets theoretically increases the accuracy and reliability of flight condition expertise, in practice, the systematic analysis of all flight events detected by an expert requires both the computational power required for the analysis of FD values and the workload of the expert. It means significant expenses that manifest themselves. On the other hand, as a rule, only those sets of FD and those flight events are analyzed, the analysis of which is required according to the regulatory documents on flight safety. Such experts form their own assessment of flight conditions using only a very small number of flight events that are pre-recorded in the database and represent only a small fraction of flight events. But this "unrecorded" portion of flight events, which may seem insignificant to the expert, can potentially reveal deeper future problems, which are not visible in unsystematic expert reports of all recorded values of flight data.
With so much data, there is no good way for problems to stand out unless a commercial company already knows what those problems look like. Says Captain Jeff Hamlett, the director of flight safety at Southwest Airlines Co., "We have mounds of data; the big request is always, 'Tell me something I don't know.' We have to start with something specific, like an issue we discovered in a pilot report, and then we can search through the system and discover the breadth and depth of the issue." [7] We explore different methods of analysing data in aviation and their effectiveness. The aviation industry is always looking for new ways to improve flight safety. However, due to the large amount of flight data collected daily, it is not possible to analyze it all manually. Therefore, problems are discovered during accident investigations.
So existing methods of analyzing flight data sets recorded during flight of an aircraft provide the possibility of comparing flight events or dynamics for identical flight conditions. However, it would be more appropriate for the expert to have access to all flight events, allowing detection based on the recorded values of the flight data. With data mining, a new area of information technology,
we can analyze unmanageably large amounts of information to find patterns and anomalies that indicate potential incidents before they happen.
On January 31st, 2000 a plane travelling from Puerto Vallarta, Mexico to Seattle, Washington dove from 18,000 feet into the Pacific Ocean, losing 89 lives. The cause of this accident was found to be "a loss of airplane pitch control resulting from the in-flight failure of the horizontal stabilizer trim system jackscrew assembly's acme nut threads. The thread failure was caused by excessive wear resulting from Alaska Airlines' insufficient lubrication of the jackscrew assembly"[7]. The cause of this accident could be predicted by analyzing the FD. There are many other incidents that can be prevented by analyzing FD.
Data mining is a broad field of data science designed to make future predictions based on patterns found in collected FD. Due to the large amount of data collected every day, it is impossible to manually find events in flight data. Data mining has been able to start addressing this problem. There are various methods and algorithms of data mining. Although they are not yet optimized for mining of aviation data in their current state, some common data mining methods, such as Hidden Markov Models and Hidden Semi-Markov Models, are being explored.
Problem solution. To understand Hidden Markov Models (HMM), we must first understand Markov chains. A Markov chain is a sequence of states in which the probability of transition from one state to another depends on the current state. An example of a Markov chain in aviation would be the probability of transitioning from one maneuver to another (eg, a pilot is more likely to stop an aircraft after landing than after takeoff).
HMM is a hidden Markov chain (level x in Figure 1) that allows inference about observable states (level o in Figure 1) and the most probable state of the Markov chain. In the aviation, the Markov chain is implicit because we are measuring the aircraft's maneuvers (xt) based on the pilot's decisions (ot). For example, the pilot turning the rudder to the left, pulling back the rudder slightly, and applying left rudder would be the observable grass state, while the latent state would be the aircraft turning left, xt. For example, the pilot turning the yoke to the left, pulling the yoke back slightl and applying left rudder would be the observable state, ot,while the hidden state would be the aircraft turning to the left, xt.
Figure 1: An example of a Hidden Markov Model [8].
Hidden Semi-Markov Models (HSMM) is an HMM that takes into account the variation of probabilities between states over the lifetime of the state. This is necessary because the duration between movements can be classified as anomalous. An abnormally length of time between landing and stopping is an example of a sequence that is anomalous due to duration. (Figure 2)
Figure 2: An example of a Hidden Semi-Markov Model [8].
Data Mining FD using HMM and HSMM. The article analyzes HMM and HSMM methods to measure their effectiveness in finding anomalous patterns in FD. As mentioned earlier, the HMM method has a disadvantage compared to the HSMM method because the HSMM method has the ability to estimate the duration of states.
The two methods use a data set of "110 landings in normal flight conditions" from a flight simulator to determine normal operation. This data was obtained from a flight simulator called FlightGear. 12 discrete pilot commands were recorded for these simulations. 5 different types of anomalous landings were then created using FlightGear:
1. Throttle is kept constant and flaps are not put down, the rest of operation is normal.
2. No initial throttle increase, the rest of operation is normal.
3. The flight is similar to normal, except that the flaps are not put down.
4. At the end of the flight the brakes are not applied, the rest of operation is normal.
5. Pilot overshoots the airport runway and lands some- where behind it.
Each of those scenarios was replicated 10 times for 50 abnormal scenarios
The log of the probability of a sequence divided by the length of the sequence was then found to determine the likelihood of a the sequence. If a sequence of states were found to be anomalous, the probability of each state, given the sequence of states before it, was used to find the anomalous state [6]
A simple set of synthetic data was used to check that the HSMM was able to detect anomalous state durations and HMM was not. This data set had 25 sequences with normal duration between states, and 25 of the same sequences, but with abnormal duration between states. The ability of HSMM to detect anomalous state durations can be seen in Figure 3.
To interpret a Receiving Operating Characteristic (ROC) curve, one must know that as the line is followed from (0, 0) to (1, 1), the threshold is being changed for how the data is classified. The knotted line depicts the ability of HSMM to detect anomalous state durations. Since the area under the knotted line approaches the coordinate (0, 1), we can see that HSMM has a threshold value that will produce mini- mal false positives and catch most true positives. However, the solid line shows that HMM is fairly unreliable at any threshold level.
False Positive Rate Figure 3: Detection of anomalous state duration of HMM and HSMM [6].
Conclusion. Poor quality or incomplete data can affect forecast accuracy, which in turn can result in the inability to completely eliminate the event and potential safety risks. The relationship between data quality and risk is illustrated in figure 4. In recent years, the statistics of machine learning algorithms have shown that the transparency and mathematical validation of the used algebraic methods are not fully ensured. The "black box" characteristics of machine learning algorithms show similarities to a number of human skills. But the physical models of the activity are almost not understood by the majority.
Algorithms I nreliable Bad Potential
prediction» maintenance safety risk
Wrong 1 nulla Ые Bad Potential
algorithms prediction* maintenance safely risk
Figure 4. The relationship between information quality and risk.
Although currently Data mining methods are not optimized for FD analysis, some methods such as kernel methods, Hidden Markov Models and Hidden Semi-Markov Models are being studied. Kernel techniques are mainly developed based on discrete or sequential information. This limitation makes it unsuitable for use in combined discrete and continuous data collected in aviation. Hidden Markov Model is limited to analyzing sequences without considering the duration of actions. Aviation incident reports often contain a small amount of information per report, whereas existing text classification methods require large amounts of descriptive information. Although these approaches are not optimal for aviation data, we can use these insights to develop new approaches for data mining.
REFERENCES
1. Human Factors Training Manual, Doc 9683-AN/950 First Edition 2008. (Russian)
2. Flight safety management manual, Doc 9859 AN/474, International Civil Aviation Organization, Third Edition 2013, 300 p. (Russian)
3. Zemlyaniy, E.S. (2016). Piloting and navigation complex with intelligent support for the aircraft crew. Dissertation for the degree of PhD in technical sciences. Moscow State Technical University named after N.E. Bauman. Moscow, (pp. 45-48) (Russian)
4. Ismayilov I.M. (2018). Expert system for intelligent pilot support in on-board complexes and its software. News of ANAS, Problems of Information Technologies, 2, 18-27 (Russian)
5. Flight Data Analysis Software Manual, Doc 10000-AN/501 First Edition 2014. (Russian)
6. https://spinoff.nasa.gov/Spinoff2013/t_3.html
7. S. Das, B. L. Matthews, A. N. Srivastava, and N. C. Oza. Multiple kernel learning for heterogeneous anomaly detection: algorithm and aviation safety case study. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 47-56. ACM, 2010.
8. I. Persing and V. Ng. Semi-supervised cause identification from aviation safety reports. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP: Volume 2-Volume 2, pages 843-851. Association for Computational Linguistics, 2009.
