D3.2 - Data and Report Analysis and Processing - FARO
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Ref. Ares(2021)796768 - 31/01/2021 EXPLORATORY RESEARCH D3.2 – Data and Report Analysis and Processing Deliverable ID: D3.2 Dissemination Level: PU Project Acronym: FARO Grant: 892542 Call: H2020-SESAR-2019-2 Topic: Safety and Resilience Consortium Coordinator: CRIDA Edition date: 27 January 2021 Edition: 00.01.00 Template Edition: 02.00.02
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Authoring & Approval Authors of the document Name/Beneficiary Position/Title Date Irene BUSELLI / ZENABYTE Project Contributor 22/01/2021 Carlo DAMBRA / ZENABYTE WP3 Lead 22/01/2021 Luca ONETO / ZENABYTE Project Contributor 22/01/2021 Reviewers internal to the project Name/Beneficiary Position/Title Date Christian VERDONK/ CRIDA Project Coordinator 27/01/2021 Fernando GÓMEZ / UPM WP4 Lead 27/01/2021 Anthony SMOKER/ LU WP5 Lead 27/01/2021 Fedja NETJASOV/ UB-FTTE WP6 Lead 27/01/2021 Patricia RUIZ/ ENAIRE Project Contributor 27/01/2021 Approved for submission to the SJU By - Representatives of beneficiaries involved in the project Name/Beneficiary Position/Title Date Christian VERDONK/ CRIDA Project Coordinator 29/01/2021 Icíar GARCÍA-OVIES/ CRIDA WP2 Lead Fernando GÓMEZ / UPM WP4 Lead 27/01/2021 Anthony SMOKER/ LU WP5 Lead Fedja NETJASOV/ UB-FTTE WP6 Lead 27/01/2021 Tamara PEJOVIC/ ECTL Project Contributor Patricia RUIZ/ ENAIRE Project Contributor Rejected By - Representatives of beneficiaries involved in the project Name/Beneficiary Position/Title Date Document History Edition Date Status Author Justification 00.00.01 24/11/2020 CREATED Irene Buselli, Carlo TEMPLATE WITH TOC Dambra, Luca Oneto 00.00.02 19/01/2021 DRAFT Irene Buselli, Carlo Completed the UK Dambra, Luca Oneto scenario 2
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING 00.00.03 22/01/2021 Review version Irene Buselli, Carlo Final draft completed Dambra, Luca Oneto 00.01.00 27/01/2021 Final version ready to Irene Buselli, Carlo All the reviews’ be submitted Dambra, Luca Oneto comments included Copyright Statement © – 2021– FARO Consortium. All rights reserved. Licensed to the SESAR Joint Undertaking under conditions 3
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING FARO SAFETY AND RESILIENCE GUIDELINES FOR AVIATION This deliverable is part of a project that has received funding from the SESAR Joint Undertaking under grant agreement No 892542 under European Union’s Horizon 2020 research and innovation programme. Abstract Deliverable “D3.2: Data and Report Analysis and Processing” is framed within WP3 of FARO project, and it covers the activities contained in the homonymous task T3.2. The objective of this task is to identify and apply suitable data-driven techniques able to gain useful insight from safety reports and contextual data. In particular, this document reports the main results obtained from the application of different statistical, Natural-Language-Processing, and Machine-Learning tools to public Airprox reports from both Spanish and UK airspace, other than some ancillary data sources. This mixture of free-text and structured data from the two scenarios was thoroughly explored, and the promising results show that it is possible to develop data and visual analytics solutions able to extract information and generate knowledge in the safety and resilience scope. The outcome of these analyses will be further investigated during the activities of T3.3, by interacting also with safety experts in WP4 and resilience experts in WP5. Additionally, this document includes a detailed list of the main data requirements emerged during the work on the task - not only in the interest of the current project but also for future developments in this scope. 4
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Table of Contents Abstract ................................................................................................................................... 4 1 Introduction ............................................................................................................... 8 1.1 Purpose of the document............................................................................................... 8 1.2 Document Structure ...................................................................................................... 8 1.3 Acronyms and Terminology ........................................................................................... 8 2 Available Data and Reports ...................................................................................... 10 2.1 Overview ..................................................................................................................... 10 2.2 Spanish Airspace.......................................................................................................... 12 2.3 UK Airspace ................................................................................................................. 14 3 Exploratory Analysis................................................................................................. 17 3.1 Spanish Airspace.......................................................................................................... 17 3.2 UK Airspace ................................................................................................................. 21 4 Natural Language Processing and Machine Learning ................................................ 30 4.1 Spanish Airspace.......................................................................................................... 30 4.2 UK Airspace ................................................................................................................. 39 5 Data guidelines ........................................................................................................ 47 6 Conclusions .............................................................................................................. 50 7 References ............................................................................................................... 51 List of Tables Table 2: Safety Barriers Legend ............................................................................................................. 22 Table 3: Topics description .................................................................................................................... 30 Table 4: Examples of automatic detection of taxonomy factors .......................................................... 38 Table 5: Topic words and labels ............................................................................................................ 40 List of Figures Figure 1: Available data ......................................................................................................................... 10 Figure 2: Distribution of CEANITA reports in time ................................................................................ 12 Figure 3: Location and separation score of CEANITA reports ............................................................... 12 Figure 4: Distribution of ICAO risk categories in CEANITA reports ....................................................... 13 5
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 5: Distribution of 3D separation in data ..................................................................................... 14 Figure 6: Distribution of UK incidents in time ....................................................................................... 15 Figure 7: Distribution of UK incidents in months .................................................................................. 15 Figure 8: Location and severity of UK incidents .................................................................................... 16 Figure 9: Distribution of ICAO risk categories in UKAB reports............................................................. 16 Figure 10: Flight Type frequency ........................................................................................................... 17 Figure 11: Frequency of most frequent conclusions ............................................................................. 18 Figure 12: Average separation score per conclusion ............................................................................ 18 Figure 13: Correlations between conclusions ....................................................................................... 18 Figure 14: Pilot contribution ................................................................................................................. 19 Figure 15: ATM contribution ................................................................................................................. 19 Figure 16: Contribution and severity score ........................................................................................... 19 Figure 17: Contribution and altitude of the incident ............................................................................ 20 Figure 18: Contribution and airspace class ........................................................................................... 20 Figure 19: Flight Type distribution in UK incidents ............................................................................... 21 Figure 20: Frequency of each key factor ............................................................................................... 23 Figure 21: Average severity per key factor (A=2, C=0) .......................................................................... 23 Figure 22: Correlation between the key factors ................................................................................... 24 Figure 23: Distribution of contribution ................................................................................................. 25 Figure 24: Risk perceived by the pilot VS actual ICAO risk category ..................................................... 26 Figure 25: Frequency of the main meteorological phenomena............................................................ 26 Figure 26: Frequency of wind directions ............................................................................................... 27 Figure 27: Histogram of incidents in time (drones, balloons, unknown objects) ................................. 27 Figure 28: Distribution of incidents in space (drones, balloons, unknown objects) ............................. 27 Figure 29: Type of objects ..................................................................................................................... 28 Figure 30: Risk categories ...................................................................................................................... 28 Figure 31: Contributing factors frequency ............................................................................................ 28 Figure 32: Text network visualisation ................................................................................................... 29 6
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 33: Example of word cloud for Topics visualisation ................................................................... 31 Figure 34: Topics average prevalence ................................................................................................... 31 Figure 35: Prevalence of topics in terms of number of reports ............................................................ 32 Figure 36: Text Network ........................................................................................................................ 32 Figure 37: Cluster Dendrogram ............................................................................................................. 33 Figure 38: Heatmap of clusters ............................................................................................................. 34 Figure 39: Example of syntactic analysis tool ........................................................................................ 35 Figure 40: Example of syntactic analysis on sentences with "instructs"............................................... 36 Figure 41: Example of syntactic analysis on sentences with "authorised" ........................................... 36 Figure 42: Example of syntactic analysis with noun “load” .................................................................. 37 Figure 43: Example of retracing of sentences with "without" .............................................................. 37 Figure 44: Decision tree......................................................................................................................... 39 Figure 45: Topic average prevalence..................................................................................................... 41 Figure 46: Topic prevalence in terms of number of reports ................................................................. 41 Figure 47: Text Network visualisation of UKAB reports ........................................................................ 42 Figure 48: Clustering dendrogram of UK reports .................................................................................. 43 Figure 49: Heatmap of clusters ............................................................................................................. 44 Figure 50: Prevalence of taxonomy factors in a report......................................................................... 45 Figure 51: Average prevalence of factors in UKAB reports ................................................................... 45 Figure 52: ID connection between free-text and structured data ........................................................ 48 Figure 53: Scanned text (Spain) VS digital text (UK).............................................................................. 48 Figure 54: Evolution of assessment guidelines (UK) ............................................................................. 48 Figure 55: Example of table in CEANITA reports (Spain) ....................................................................... 49 7
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING 1 Introduction 1.1 Purpose of the document Work package 3 (WP3) addresses FARO Data Transformation and Loading, which is basic for the smooth running of WP4, WP5 and WP6. Within this WP, Task 3.2 (T3.2), addresses the application of data analytics, Natural Language Processing (NLP) and Machine Learning (ML) to safety reports and shows what kind of information is possible to extract and how it can be used to generate potential safety indicators and dimensions [1]. Data used in this task have been treated according to the procedures described in [2]. Deliverable D3.2 reports on the results of the Task 3.2. 1.2 Document Structure This document is structured as follows: • Section 1 introduces the purpose of this document. • Section 2 describes the available data sources. • Section 3 explores safety reports using data-analytics techniques. • Section 4 reports the results of the application of Natural Language Processing (NLP) and Machine Learning (ML). • Section 5 assesses some guidelines and recommendations about data sources. • Finally, Section 6 summarises the main conclusions. 1.3 Acronyms and Terminology Term Definition ATC Air Traffic Control ATM Air Traffic Management ATM Air Traffic Management CEANITA Comisión de Estudio y Análisis de Notificaciones de Incidentes de Tránsito Aéreo / Commission for Study and Analysis of Notifications of Incidents of Air Traffic CWP Controller Working Position DWH Data Warehouse FT Feet ICAO International Civil Aviation Organization LDA Latent Dirichlet Allocation LoS Loss of Separation 8
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Term Definition ML Machine Learning NLP Natural Language Processing NM Nautical Miles RAT Risk Analysis Tool STCA Short Term Conflict Alert TOKAI Tool Kit for ATM Occurrence Investigation UKAB United Kingdom Airprox Board VFR Visual Flight Rule WP Work Package 9
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING 2 Available Data and Reports 2.1 Overview Deliverable D3.2 describes the results of the application of data analytics, Natural Language Processing (NLP) and Machine Learning (ML) to safety reports and contextual data and shows what kind of information is possible to extract and how it can be used to generate potential safety indicators and dimensions. NLP and ML have been applied to two different sets of safety reports (Spanish, to exploit CRIDA/ENAIRE expertise, and UK, since the English language is the most commonly available in NLP tools): • CEANITA Airprox reports (Spain); • UKAB Airprox reports (UK). Figure 1 outlines the available data sources and their connections. Structured data are coloured in blue and free-text ones in yellow. Figure 1: Available data • CEANITA reports: anonymised incident reports about “Loss of separation” events, from January 2018 to July 2019. They are public reports published by the Agencia Estatal de Seguridad Aérea (Spanish Safety Aviation Agency) under the commission CEANITA (Comisión de Estudio y Análisis de Notificaciones de Incidentes de Tránsito Aéreo – Commission for Study and Analysis of Notifications of Incidents of Air Traffic)1.The content of these reports is in Spanish, and a detailed example of it can be found in D3.1. 1 Source: https://www.seguridadaerea.gob.es/es/ambitos/gestion-de-la-seguridad- operacional/ceanita#Informes%20Definitivos 10
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING • Loss of Separation data: A Loss of Separation describes the interval of time where two aircraft are below some specific limits of horizontal (NM – Nautical Miles) and vertical (FT – Feet) separation. For each event the distance at the closest point of approach (with different metrics) and the separation score are indicated, after the application of the RAT methodology (Risk of Collision, Separation). • ATC events data: Interactions of some Air Traffic Controllers with the Controller Working Position (CWP), related to the aircraft. • STCA alerts data: Raw STCA (Short Term Conflict Alert) data, per flight, as provided by the surveillance source. STCA is a ground-based safety net generating timely an alert of a potential or actual infringement of minimal separation. • Sector entry data: All the geographical transitions of aircraft from a sector to another one. • Flights: The invariant information related to each flight. • Radar tracks: A flown flight track with a frequency of 5 seconds. A more accurate description of the Spanish data can be found in the D3.1. Only information about the analysed reports have been considered. • UKAB Airprox reports: publicly available incident reports about Airprox events stored by UK Airprox Board2. This source includes all the Airprox reports for UK airspace from 2010 to 2020, composed by both structured data and free text. The information that can be extracted is very similar to the one relative to Spanish airspace, except for the absence of the lower-resolution data (i.e., data relative to all the flights, radar tracks, alerts) present in CRIDA Data Warehouse (DWH). See Table 1. Table 1: A summary of an UKAB Airprox Report Example Variable Example ID 2019028 Date 15/02/2019 Time 15:25 Latitude 51:07:00 N Longitude 0:39:00 E Altitude 1900 Risk Category C Aircraft Type (1 and 2) VULCAN - P68 / BEECH - 36 2 Source: https://www.airproxboard.org.uk/Reports-and-analysis/Monthly-summaries/Monthly-Airprox- reviews/. 11
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Variable Example Flight Classification (1 and 2) Non-Commercial Operations - Pleasure Airspace Class G Airspace Name London FIR ATS UK FIS - BASIC SERVICE Flight Rules (1 and 2) VFR Separation (V/H) 100ft V/0.1nm H Weather METAR EGKK 151520Z 15005KT CAVOK 14/05 Q1025 2.2 Spanish Airspace 2.2.1 CEANITA reports The considered reports are relative to Airprox incidents happened between January 2018 and July 2019, in different operating environments. In the following, a brief overview of their content in terms of time (Figure 2), space and separation score (Figure 3) is provided. Figure 2: Distribution of CEANITA reports in time Figure 3: Location and separation score of CEANITA reports 89 incidents are reported in this period (incidents that may or may not, after investigation, be considered a breach of separation), involving different operational environments. The temporal distribution can be seen from the histogram (Figure 2)). It is important to notice that incidents reported by CEANITA are a subset of the safety-related occurrences, with higher severity incidents being overrepresented (Section 2.2.2 offers a more detailed characterisation of the subset). 12
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING From a spatial point of view, reported incidents tend to be quite scattered, as it can be seen from the map in Figure 3 (which also reports their separation score according to the metric used in CRIDA DWH data). In Figure 4 the distribution of ICAO levels of severity (or ICAO risk category) is shown, as reported in the CEANITA reports. In particular, incidents of class D and E are a negligible amount, while the majority are classified as B. Figure 4: Distribution of ICAO risk categories in CEANITA reports As a final observation, as regards airspace class, the vast majority of the considered incidents happened in classes C and D (about 40% each) and about 10% in class A. 2.2.2 CRIDA DWH Since the focus of Task 3.2 is mainly on the analysis of reports – and therefore the other data sources are used as complement to this activity – in this document only some major characteristics of these data are highlighted. For more information on the data sets please refer to Deliverable D3.1. Figure 5 shows the distribution of 3D separation (i.e., calculated with the so-called 3D algorithm) in incidents contained in the entire LoS collection (red) in comparison with the subset described in CEANITA reports (teal). While in the global case the separation is distributed in an almost Gaussian fashion which tends to the boundaries to the separation limits, CEANITA reports focus on more “serious” LoSs, showing in general lower values of this metric, even if minimum and maximum values coincide in the two groups. Some further observations on 3D separation should be done: • This is not the only metric used to determine the distance between the two aircraft; in fact, other than the 3D algorithm, the value calculated by the Percentage algorithm is reported as well. However, the two metrics are really similar, so in the following only the 3D value will be reported. • The separation value does not have an absolute meaning: the same distance value may or may not be classified as a loss of separation depending on the separation standards in force. 13
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING • The correlations between separation, separation score and ICAO category are positive but not high enough to consider them interchangeable when evaluating the seriousness of the incident. Figure 5: Distribution of 3D separation in data 2.3 UK Airspace 2.3.1 UKAB reports UKAB reports differ from CEANITA ones in a number of respects, but their content is in large part comparable. The reports considered are relative to Airprox incidents happened between January 2010 and December 2019. In the following, a brief overview of their content in terms of time, space and severity is provided. Since 2286 incidents are reported in these 10 years, it is possible to say that the average frequency is more than 4 incidents a week, in all operating environments. Actually, as can be seen in Figure 6, the number of analysed incidents has constantly increased during this period: indeed, the number of reports has almost doubled in these years. Furthermore, the distribution during the year is quite uneven (Figure 7) and, not surprisingly, in summer the average amount of analysed incidents is significantly higher than in the other seasons. 14
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 6: Distribution of UK incidents in time Figure 7: Distribution of UK incidents in months Also, from a spatial point of view (Figure 8), the distribution is not homogeneous: by looking at the map it seems that incidents were more concentrated on the part over London and the South East of England. Severity, instead, seems more scattered in the map. The general distribution of ICAO levels of severity (or ICAO risk category) can be seen in Figure 9. In slight contrast with the Spanish counterpart, the percentage of incidents classified as D or E (risk not determinable or not present) is not completely negligible (together they represent more than 10% of the total); furthermore, the most frequent ICAO category is C and not B as in CEANITA reports. However, ICAO categories A and B (which identify the so-called “risk-bearing” incidents) represent more than one third of the total.3 3 For more information about incident statistics: https://www.airproxboard.org.uk/Reports-and- analysis/Annual-Airprox-summary-reports/ 15
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 8: Location and severity of UK incidents Figure 9: Distribution of ICAO risk categories in UKAB reports The main significant difference with incidents contained in CEANITA reports, however, is in the airspace classes: more than 90% of the incidents described in UKAB reports happened in class G. Furthermore, not surprisingly, flight rules in place during the incidents were mostly VFR. 16
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING 3 Exploratory Analysis 3.1 Spanish Airspace The aim of the analyses reported in this section is to identify some potential contributing factors (and therefore contribute to the definition of safety performance indicators) and their relationships with the different features of the incidents. The exploited variables are mostly derived from the combination of both features extracted from reports and structured information from CRIDA DWH relative to the same incidents. In particular, the main results in this context are: • The statistical significance of the abundance of Flight Training and Military flights in CEANITA reports with respect to their standard prevalence in air traffic (Figure 10). • The differences in prevalence and severity of the conclusions reported (Figure 11 and Figure 12), and their correlations (Figure 13). • The different rate of contribution for pilot and ATM in the incidents (Figure 14 and Figure 15), and how the different contribution can lead to different outcomes (Figure 16, Figure 17 and Figure 18). Figure 10: Flight Type frequency Figure 10 shows how the distribution of the flight types (General, Military, Non regular, Regular, Other) is different in CEANITA reports (teal) with respect to the general distribution in flights (red). In particular, the relative frequency of regular flights (S) is significantly lower in the ones involved in incidents, while flight classified as X and M are relatively more frequent. This graphical observation is confirmed by a statistical test: in fact, the Chi-squared test returns a negligible p-value, which means 17
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING the difference in distributions is statistically significant. It is important to notice that the flights classified with X in CEANITA reports are for the vast majority Flight Academy flights; so, to sum up, the prevalence of Military and Flight Academy flights in the reported incidents can be defined significantly higher than the standard proportion in air traffic. Figure 11: Frequency of most frequent conclusions Figure 12: Average separation score per conclusion Figure 13: Correlations between conclusions As regards the causes of the incident, the ones reported in the graphs are retrieved from the conclusive paragraph of each CEANITA report. These causes are not mutually exclusive: in fact, Figure 13 shows how some couples of conclusions are very likely (or unlikely) to appear together. For instance, • the absence of detection is highly correlated with the absence of resolution, 18
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING • while a problem of wrong clearance is unlikely to happen together with a deviation from process. Looking at how the two colours (blue and yellow, corresponding respectively to positive and negative values of correlations) form two separate blocks, it is possible to derive that these two groups of causes roughly correspond to ATM responsibility (e.g., wrong clearance, wrong resolution or coordination problems) and pilot responsibility (e.g., level bust, deviation from process, airspace infringement). Figure 14: Pilot contribution Figure 15: ATM contribution Other than these conclusions, CEANITA reports also include an evaluation of the actual contribution of both ATM and pilot, classifying them as Direct, Indirect and None. As it can be easily seen from Figure 14 and Figure 15: • ATM contribution is direct in the great majority of cases, while the incidents with pilot’s direct contribution are significantly fewer. • It is also interesting to notice that the percentage of indirect contribution is always minor but not negligible. Figure 16: Contribution and severity score 19
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Given these considerations derived from the CEANITA reports, one may be tempted to think that ATM actions are associated to a greater severity; but, by looking at Figure 16 it is possible to observe that, when the direct contribution is of the ATM only, the separation score tends to be lower. Therefore, from a pure data analysis perspective, it is possible to state that ATM responsibility is involved in the majority of CEANITA incidents, but the most severe ones are usually due to pilot contribution. Figure 17: Contribution and altitude of the incident Figure 18: Contribution and airspace class Finally, Figure 17 and Figure 18 show that the contribution is not independent of other variables. In fact, pilot’s contribution seems to play an important role when incidents are at low altitude (possibly during take-off or landing operations), while at higher altitudes the responsibility is almost surely of the ATM (Figure 17). Figure 18 is a mosaic plot: the area of the tiles is proportional to the number of 20
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING observations within that category and the colours correspond to Standardised Pearson Residuals, i.e., a measure of significance of the correlation: blue and red tiles are respectively “too large” or “too little” to be considered random. It can be observed that: • The tile containing incidents in class C and with ATM contribution is the largest one, while the co-occurrence of pilot responsibility and class C is almost negligible. • In classes G and E (where indeed ATC has no authority) the major contribution is of pilots. In the end, it is possible to conclude that the extraction of features from the reports – and the subsequent combination with structured data – highlighted the main causes or major contributing factors of incidents contained in CEANITA reports. Further analyses of these data-driven conclusions are left to the expert interpretation of WP4 and WP5. 3.2 UK Airspace Analogous with the previous section, the objective of the following analyses is to identify the main contributing factors and their relationships with other features, in a purely statistical way. The exploited variables are mostly derived through the extraction from UKAB reports. In particular, the main contents are: • A quick observation about the flight type distribution (Figure 19). • The differences in prevalence of the causal factors reported (Figure 20 and Figure 21), and their correlations (Figure 22). • The different rate of contribution for pilot and ATM in the incidents (Figure 23), which shows a completely different situation with respect to the Spanish one. • A comparison between perceived risk and actual risk (Figure 24). • An overview of the meteorological situation during the incidents (Figure 25 and Figure 26). Figure 19: Flight Type distribution in UK incidents 21
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING A preliminary observation is needed here: contrary to the Spanish case, for UK only data about incidents could be exploited. As a consequence, no significant comparison with “non-incident” conditions can be made. For instance, it was not possible to check if the Flight Type distribution in the reported incidents is statistically different from the norm; however, the high percentage (Figure 19) of military and training flights involved in incidents may suggest that, as for the Spanish airspace, these kinds of traffic tend to be involved in incidents more easily. Considering the assessment of contributing factors, UKAB reports are an accurate source of information: at least in the most recent years, each report contains a conclusive section about “Safety Barrier Assessment”. The factors considered in this analysis can be seen in Table 2, together with the abbreviations used in the following graphs. Please be aware that since this assessment methodology has evolved during the past years, the only way to aggregate this information in a meaningful and coherent way is to consider just reports from 2017 on, so the results about key factors and contribution are only relative to 2017-2019 incidents. Barrier Abbreviation Regulations, Processes, Procedures&ComplianceG Procedures and Compliance Manning & Equipment Manning&EquipmentG GROUND ELEMENTS Situational Awareness of the AwarenessG Confliction & Action Electronic Warning System WarningG Operation and Compliance Regulations, Processes, Procedures&ComplianceF Procedures and Compliance Tactical Planning and Execution TacticalPlanningF FLIGHT ELEMENTS Situational Awareness of the AwarenessF Conflicting Aircraft & Action Electronic Warning System WarningF Operation and Compliance See & Avoid See&AvoidF Table 2: Safety Barriers Legend In the considered reports, each of these key factors is evaluated as effective, partially effective or ineffective. In Figure 20 the frequency of non-effectiveness of each contributing factor is represented, with different colours accordingly to the level of criticality. In particular, it is interesting to notice that the frequency of flight elements is disproportionately higher than the one of ground elements. In particular, the situational awareness seems to be the most critical problem, or at least the most often mentioned factor when listing the “ineffective” ones. Also the ability to “see and avoid” and “tactical planning and execution” are in the top three of the factors defined as problematic. 22
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 20: Frequency of each key factor Figure 21: Average severity per key factor (A=2, C=0) The difference in frequency, however, does not completely overlap with the difference in average severity. Figure 21 shows the average severity of an incident given that a particular factor is non- effective; the numerical severity is computed posing risk category C equal to 0 and A equal to 2, excluding classes D and E from the computation. A first observation is that the difference between these values is not particularly large (for instance if compared with Figure 12 which shows the Spanish equivalent). Nevertheless, the most evident observation however is the discrepancy in ranking between this graph and the frequency one: for instance, problems with electronic warning system at ground level are the least common but the most critical in terms of average severity; on the contrary, the loss of situational awareness is the most frequent problem but its impact on safety is just in the middle of the ranking. 23
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 22: Correlation between the key factors The way factors correlate with each other, instead, is not particularly surprising: flight elements are in general positively correlated within themselves and negatively correlated with ground elements, which in turn are positively correlated with each other. This is not completely systematic though: for instance, “WarningF” is negatively correlated with “Procedures&ComplianceF” and “TacticalPlanningF”, and slightly positively correlated with ATM situational awareness. Indeed, as shown in Figure 23, it is quite common (more than 30% of times) that contributing factors are both ground and flight elements. However, in the majority of incidents (more than 60%) the highlighted contributing factors are identified as flight elements. The difference with the Spanish situation is remarkable; nonetheless, this is in line with the different proportion of airspaces classes: most UK incidents are reported to happen in class G, where the responsible of the separation among aircraft are the pilots. As a last remark on this, contrary to the Spanish scenario, further analyses on the different outcomes generated by different contribution did not lead to significant results. 24
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 23: Distribution of contribution The last observation about human factors and, in particular, about human biases, regards the discrepancy between the risk of collision perceived by the pilot and the actual risk. Figure 24 shows how the perceived risk is often really different from the one defined by the risk category (please be aware that the size of the cells on both sides of the chart is just due to their proportion in the reports): • “High” risk is associated just in a minor part to category A, while about half of the times the incident ends up being classified as C or even E (i.e., no risk at all). • “Medium” risk is associated largely with C and E, which again are not risk-bearing incidents. • “Very high”, “Low” and “None” seem more coherent with the final classification, even if not completely. In general, the tendency suggests that pilots tend to overestimate risk most of the times. Potential factors that might be behind the difference are: • First, pilot estimation of risk is often a judgement formed “in the heat” of the action, when their own safety might be at risk. • Second, they are often unaware of surrounding factors that might result in a lowered risk score, such as ATC giving adequate evasive instructions to the other aircraft involved. 25
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 24: Risk perceived by the pilot VS actual ICAO risk category The last part of this section is a brief overview about the meteorological conditions reported during the incidents. Indeed, contrary to CEANITA reports, in the UKAB ones, weather is reported both when a particular phenomenon is in place and when no significant phenomenon is observed. As a result, it is possible to state that more than 95% of the considered incidents (Figure 25) happened in the absence of significant meteorological phenomena. As regards wind direction (Figure 26), the most frequent ones are W and SW, which is in line with the standard conditions in UK: again, this suggests atmospheric conditions do not play a major role in incidents. Figure 25: Frequency of the main meteorological phenomena 26
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 26: Frequency of wind directions 3.2.1 Airprox with drones, balloons and unknown objects Airprox reported by UKAB include a non-negligible percentage of problems with drones, balloons and unknown objects. Figure 27: Histogram of incidents in time Figure 28: Distribution of incidents in space (drones, (drones, balloons, unknown objects) balloons, unknown objects) A first overview of the data about this kind of incidents reveals that the trend in time and space is similar to other kinds of Airprox (Figure 27and Figure 28): peaks are in summer and near London and the South East of England. 27
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING The most frequently involved objects are drones, while balloons represent only about 4% (Figure 29). Looking at risk categories (Figure 30), the situation is partially different from standard incidents: the distribution is almost one third for each category A, B and C, while D and E are a negligible percentage. Figure 29: Type of objects Figure 30: Risk categories Reports about these incidents are shorter and less detailed than the other Airprox reports, but there are some interesting contents: • A schematic part describing the main contributing factors (Figure 31), which suggests that lack of situational awareness and deviations from procedures are the most common causes. • A free-text part describing the incident, which is summarised here with the text network in Figure 32. It is interesting to note that the core of the network (bottom left), which form the very common sentence “the incident is best described as the drone has flown in conflict with the aircraft” has two main “branches”: one points out the vicinity to an airfield approach path, the other one is about the infringement of height limit of 400 ft. Figure 31: Contributing factors frequency 28
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 32: Text network visualisation 29
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING 4 Natural Language Processing and Machine Learning 4.1 Spanish Airspace 4.1.1 Unsupervised 4.1.1.1 Topic Modelling Topic Modelling is an NLP unsupervised-learning text analysis technique which enables the discovery of the main themes that pervade a collection of textual data. Thanks to this technique, a text collection can be organized according to the discovered themes and its information can be therefore made quantitative, since the probability of finding a topic (or prevalence) can be associated to every report, becoming a numerical feature describing the document. The technique applied here is Latent Dirichlet Allocation (LDA), which is one of the most used in literature ( [3] [4]). In particular, the application of LDA lead to the identification of 12 main topics which aviation experts from CRIDA considered significant and coherent. The words or bigrams describing these topics are reported in Table 3. WORDS/BIGRAMS LABEL helicopter drop water fires extinguishing coordination drop area emergency / fire load work high alone workload instructions previous tactical workload aircraft to take off departure / initial departure to take off runway rate they are climb aircraft climb wind tail down-wind leg wind leg right tail runway wind / traffic circuit adverse meteorologic weather adverse detours due to weather thunderstorm weather weather conditions aircraft runway go around go around to take off to land go around established sector frequency sectors high coordination transfer limit sector operations aircraft sector communication / answer received finally decided they saw communication visual contact traffic information course aircraft to descent sector to aircraft to clearance rate descent descent descend rate descend maintain to confirm maintaining received coordinating confirming sector informs receipt coordination receipt formation activation alert early early alert activation function alert function alert function military military military formation defence air defence main centre military formation aircraft Table 3: Topics description 30
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 33: Example of word cloud for Topics visualisation For a more clear and intuitive representation, topics can also be visualised via word clouds, as in 33. Topic modelling allows to obtain a sort of higher-resolution insight with respect to simple exploratory analysis: in fact, while in Section 3.1 just the major conceptual problems are identified (e.g., wrong clearance, airspace infringement etc.) now some underlying factors can be pointed out (e.g., it can be discovered that the ATM error was due to excessive workload or to an emergency situation). In Section 4.1.1.3 the relationships between topics and main causes are analysed by applying a clustering technique. Further investigations about relationships between topics and previously identified variables will be considered in T3.3. In Figure 34 the average prevalence of each topic (i.e., how prevalent it is inside each report) is shown; this value is interesting in terms of “relative” comparison (i.e., the fact that top topics are relative to communication, alerts, wind and workload), but to have an idea of how frequent each topic is, we have to look at Figure 35. Figure 34: Topics average prevalence 31
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 35: Prevalence of topics in terms of number of reports 4.1.1.2 Text Networks The second unsupervised technique applied to the free-text reports is Text Network Visualisation [4] [5] . The result (which can be seen in Figure 36) gives an insight which is partially overlapped with the one produced by Topic Modelling, but not completely. Indeed, other than identifying correlated groups of words, similarly to LDA, it also outlines some bigrams or trigrams which are naturally connected in the report language, giving some hints on the fixed formulas which are often used in the text. In the picture, the groups of words circled in blue are the ones similar to LDA topics. Figure 36: Text Network Some examples of information retrieved from the network are: 32
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING • A group of correlated words like helicopter, fire, extinguishing, water etc., which interestingly are connected to Llutxent (Valencia), where a serious fire caused severe problems in summer 2018. • A group of words like defence, military etc. • A group of words like alert, warning, activation etc. • A group of words describing aerodrome traffic circuit (downwind, circuit, airport). • A group of words (circled in yellow) which form sentences like “the rest of the documents do not contain other relevant information for the analysis”. • Some potentially interesting bigrams, to be further analysed by safety and resilience experts: rotation-shifts, early-alert, progress-strip, workload, opposite-direction etc. It is important to consider that these are not the most frequent groups of words, but just the most relatively correlated ones, which is a different kind of information. In general, this approach does not add much information to the one presented in the previous sections. 4.1.1.3 Similarity Clustering The last unsupervised technique is the Clustering one. In particular, hierarchical clustering with Ward distance is applied on a dataset where the reports are described in terms of topic probabilities, contribution (ATM vs Pilot) and presence of the main conclusions (Section 3.1). The dendrogram (Figure 37) allows us to identify 8 possible clusters, of different sizes, which group the reports and, therefore, the exploited variables, according to the heatmap in Figure 38. Figure 37: Cluster Dendrogram 33
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 38: Heatmap of clusters As it is possible to observe from the heatmap, which reports normalized values of the relative frequency of the features, the eight clusters are quite different. In detail: • Two very small subgroups (clusters 6 and 8) are identified as particularly different from the norm: cluster 8 is composed of incidents where the main topic is “emergency” (indeed, they are the reports referred to Llutxent fire in summer 2018); cluster 6 instead contains the three incidents caused by level bust (and quite expectedly, according to the heatmap, the main contribution was of the pilot and the other main conclusion was “deviation from process”). • The largest group (cluster 5) is mainly composed of wrong-clearance and late-detection incidents, with clearly the highest frequency of ATM contribution and an interesting high prevalence of “descent” topic. • Cluster 4 contains incidents mainly caused by “wrong resolution” by ATM, with high prevalence of topics related to go around, departure and weather. • Cluster 7 is composed of incident caused essentially by transfer problems or coordination problems; the most frequent topics are “sector” and “military”. • Incidents in cluster 3 are essentially due to pilots’ errors, in particular to airspace infringement and unfulfillment of VFR process. • Cluster 2 is characterised by incidents due to pilots’ deviations from process, especially in the landing phase (topic wind/traffic circuit). • Cluster 1 is composed of incidents due to ATM’s inability to both detect and resolve, with interesting high values of “alert” topic. A further interesting fact is that some topics (workload, communication etc.) are almost homogeneously distributed in all the clusters, without high or low peaks. 34
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING 4.1.2 Supervised The aim of supervised analyses is to take a further step towards the identification of recurring patterns and contributing factors: indeed, the objective is not only to detect underlying factors, but also to link them with safety indicators and with TOKAI taxonomy (currently implemented only in the UK scenario due to the language). 4.1.2.1 Syntactic Analysis for contributing factors This section reports the main results relative to the syntactic analysis of the reports. This tool (Figure 39) is based on a state-of-the-art methodology [6], which in this framework can be used for three main analyses: • Starting from a particularly significant verb (e.g., “instructs”, or “authorized”) the related event is automatically retraced (e.g., who instructs the aircraft to do what) possibly providing some hints about the contribution of the action to the incident (Figure 40 and Figure 41). • Starting from a particularly significant noun (e.g., “load”), the related adjective and nominal modifier (if present) can be retrieved, possibly providing a more useful information about the object or fact (Figure 42). • Starting from some particular words which may indicate something happening in an incorrect or unconventional way (e.g., “without”, “couldn’t”, etc.), possibly providing some hints about what went wrong during the incident (Figure 43) and possibly allowing to associate each problem to a factor in taxonomy (Table 4). This could be useful not only in a qualitative way (i.e., improving knowledge from a human point of view) but also in a quantitative way (i.e., generating some features describing reports by the presence or absence of these factors in an automatic way). Figure 39: Example of syntactic analysis tool As can be seen in Figure 39, through syntactic analysis it is possible to identify subject, main verb, object and clause; furthermore, we can retrieve specific adjectives or complements, or search for a particular lemma (e.g., if you want to find all the possible derivations of “instruct”, “instructs”, “instructed”, “instructing” etc., you can just search for the lemma “instruct”). Figure 40 and Figure 41 shows the most probable subjects and clauses related to “instructs” and “authorized” contained in the entire collection of CEANITA reports. 35
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 40: Example of syntactic analysis on sentences with "instructs" Figure 41: Example of syntactic analysis on sentences with "authorised" 36
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Figure 42: Example of syntactic analysis with noun “load” Figure 42 is an example of syntactic analysis applied to nouns, in particular to the noun “load”. First of all, it is interesting to see that the four possible nominal modifiers suggest four different situations: traffic load, workload, water load (in case of fire extinguishing) and load of communications. These are four different scenarios and for each of them we can retrace the seriousness by retrieving the associated adjectives, which in CEANITA reports vary from “reasonable” to “excessive”, identifying different level of severity. Figure 43: Example of retracing of sentences with "without" 37
D3.2 – DATA AND REPORT ANALYSIS AND PROCESSING Factor Example A. Personnel A-1. Perception A-1-2; A-1-4. Detection Sector authorised without detecting aircraft 2 A-1-5; A-1-6. Perceive Crew may not have seen visual/auditory information A-2. Memory A-2. Recall information Aircraft2 was authorised without remembering presence [of Aircraft1] A-4. Action A-4-5. Convey / record Sector indicated without specifying information (completeness) level A-5. Conformance A-5. Conformance [with Aircraft 2 did not comply with the rules or procedures] instruction B. Interaction with Environment B-2. Pilot/Controller B-2-11. Phraseology They instructed without using Communications (correctness) phraseology B-6. Weather B-6-6. Thunderstorm They could not comply due to activity thunderstorm Table 4: Examples of automatic detection of taxonomy factors 4.1.2.2 ML models for contributing factors Another possible way to identify contributing factors and to assess their relationships is building a Machine Learning model. ML is often used with predictive aims, but the objective can actually be (like in this case) not strictly predictive: it is possible to just simulate a prediction of the severity of incident given some features in order to understand the roles of them in generating a more or less severe issue. Clearly, a situation in which we already know an incident will take place in few minutes and we want to know how severe it will be is completely unrealistic. However, if the model is white-box and easy to interpret, it can really help in discovering the contribution of each factor to the severity of the incident. The success of tree-based models in literature is essentially due to their readability: indeed, even without a particular knowledge about ML, “decision trees” like the one in Figure 44 are quite intuitively interpretable: starting from the top, the reader just has to follow the “branches” until he reaches the “leaves” at the bottom. For instance, taking a right he can read: if the flight type of one of the aircraft involved is M or X, and the topic about workload is present, the most probable outcome in terms of separation score is 10. Furthermore, the leaf also indicates that the probability of this outcome given these features is 83% (while just 8% of outcome 1 or 3 and zero probability of outcome 7) according to the past. The last information contained in the leaf is that these features (flight type M or X and presence of workload) characterise the 17% of the reports corpus. By iterating this reading mechanism, the tree interestingly suggests for instance that the most severe incidents typically involve Military or Flight-school flights (in CEANITA reports the X-type aircraft are usually the “flight-school” ones), that incidents containing “alert” topic are usually the less problematic, and that workload is always a very worsening factor. 38
You can also read