From Recorded Exceedance to Named Hazard: The Interpretive Half of Flight Data Monitoring

A safety analyst in a control room studies flight data monitoring reports, highlighting the interpretive work behind aviation safety.

A flight data monitoring programme produces events. An event is a measurement: a parameter crossed a value the operator chose, in a phase the operator defined, on a flight the operator flew. It is not a hazard, it does not by itself describe a risk, and counting events tells you mostly about the thresholds that generated them. The work that turns a list of exceedances into something a safety action plan can act on is interpretive, it is done by people, and it is the part of the programme that no software licence includes.

An event is an outcome, a hazard is a condition

The safety management vocabulary separates three things that daily language runs together. A hazard is a condition with the potential to cause harm. A consequence is what that condition produces if it is not contained. Risk is the assessed combination of how likely the consequence is and how bad it would be. Flight data records consequences — or more precisely, it records that a parameter reached a value on the way to one.

The distinction matters because interventions attach to conditions, not to outcomes. “High rate of descent below the stabilisation gate at a given airport” is a count of outcomes; nothing can be done to a count. “The published arrival to that runway requires a descent gradient that cannot be flown clean, so aircraft arrive at the gate with energy to lose” is a condition, and conditions can be redesigned, retrained, renegotiated with the air navigation service provider, or accepted with a stated rationale. The whole value of the analysis lies in getting from the first sentence to the second, and nothing in the data makes that crossing automatically.

Flight data shows what happened, but not why. To turn a simple “event” into a real safety hazard, you need to understand the bigger picture—like why a plane descended too fast or why a landing was too long. That’s where human analysis comes in.

On a runway in Paris, an A320 had a long flare and touchdown. The flight data showed the event, but only by talking to the crew did analysts learn that a sudden gust of crosswind, not pilot error, caused the issue. Without that context, the data alone would have been useless.

What an event measures, and what it structurally cannot say
Recorded event What is genuinely established What it does not establish
High rate of descent below a stated height A parameter crossed a threshold, at a place and a time, on a known aircraft Whether the crew was recovering from an ATC descent constraint or from their own energy management
Long flare / long touchdown The aeroplane spent longer than usual between screen height and contact Whether the runway condition, a gusting crosswind, or a technique issue drove it
Late configuration Flap or gear selection occurred below a stated height Whether a speed restriction was imposed, requested, or self-inflicted
Pitch rate at rotation The rate exceeded a value in the take-off phase Whether the aeroplane was near a tail strike or comfortably clear of it — that depends on geometry and mass
Zero events on a route Nothing crossed a threshold Whether the route is safe, or whether the thresholds cannot see what happens there

What the recorder gives you, and what it withholds in silence

Before interpretation begins there is a data quality question that is easy to skip because it never announces itself. A flight data monitoring programme reads a recording it did not specify — a quick-access copy, or the mandatory recording itself — decoded through a frame layout that maps bits to parameters. When that mapping is wrong for the aircraft, the tail number or the software standard, the decoded values are not obviously wrong — they are plausible. A parameter that is silently offset, scaled incorrectly or absent produces either events that never happened or, worse, the quiet disappearance of an entire event class.

Derived parameters inherit every error in their sources. A computed energy or gradient figure built from several recorded values carries the noise and the sampling rate of the worst of them, and a parameter sampled once per second describes a manoeuvre that happens in three seconds rather poorly. None of this makes the programme useless; all of it means that an unusual finding deserves a look at the raw trace before it becomes a statistic.

The trap. A fleet that suddenly shows fewer events of one type has usually not improved. Far more often something changed in the chain: a software standard, a frame layout, a parameter that stopped being recorded, a threshold that was edited, a phase definition that moved. The first question about any favourable trend is what changed in the measurement, and the answer must come from a configuration record rather than from memory.

The threshold manufactures the statistic

Thresholds are the operator’s choice. Neither ICAO nor the European rules specify the value at which a rate of descent becomes an event, and the widely used stabilisation gates that many programmes are built around are industry guidance rather than regulation. That freedom is necessary — a regional turboprop and a long-haul twin cannot share a trigger — and it is also the mechanism by which a programme can fabricate or erase its own findings.

Set the threshold too tight and the programme drowns. The analyst spends the week validating events that carry no information, the review group sees a wall of low-severity noise, and genuine signals sit inside it undistinguished. Worse, the team adapts: the events stop being read individually and start being counted, which is precisely the failure the programme exists to prevent. Set the threshold too loose and the tail of the distribution disappears. The rate looks healthy, the trend looks flat, and the events that would have shown the drift are simply never created.

The subtler problem is the mid-series change. A threshold edited in month seven produces a time series with two different meanings and one continuous line. Any comparison across that point is meaningless, and any conclusion drawn from it — improvement, deterioration, the effect of an intervention — is an artefact. The discipline that fixes this costs nothing: every threshold, phase definition and parameter change is dated and recorded, and every trend chart carries those dates.

How the threshold setting shapes what the programme can and cannot see
Setting What the analyst experiences What happens to the finding Symptom to watch for
Too tight Hundreds of events, mostly benign Real signals are diluted; individual reading is abandoned Validation backlog growing; severity bands almost entirely low
Too loose Few events, all severe The precursor tail is invisible; only outcomes are seen Events appear without any preceding drift in the same measure
Well set A readable number, spread across severity Precursors and outcomes both visible Trends respond to known operational changes
Changed mid-series Apparent step improvement or deterioration Every comparison across the change is invalid A step change with no operational explanation

Denominators move too

Rates need exposure, and the exposure figure is a choice with consequences. Events per thousand sectors, per thousand landings, per aircraft-month and per airport pairing answer different questions, and a fleet that is growing, shrinking, changing route structure or changing seasonal mix will move all of them without any behaviour changing.

The instructive case is a seasonal one. A network that adds mountainous or short-runway destinations for a season will show a rise in approach-related events that reflects where the aeroplanes went, not how they were flown. Normalising by sector count hides that; normalising by exposure to the relevant condition — approaches to the runways in question — reveals it. Choosing the denominator is part of the analysis, and stating it is part of the report.

The only source of “why” is the crew

Flight data describes what the aeroplane did with complete fidelity and says nothing whatever about why anyone did it. The gap is not narrowed by more parameters or better algorithms; it is closed by asking, and the mechanism for asking is the de-identified crew contact.

The architecture is well established and its protections are not optional. ICAO requires that a flight data analysis programme be non-punitive and contain adequate safeguards to protect the source of the data; the European rules carry the same requirement, and the occurrence reporting regulation adds its own protections for the reporter. In practice a gatekeeper — very often a crew representative rather than a manager — holds the only means of linking a data set to a crew, and the terms of that role are written into an agreement rather than left to custom. Programmes that erode this do not merely become unpopular; they lose the input that makes interpretation possible, and they lose it permanently.

The value of the conversation is rarely a confession. It is usually a piece of context the data cannot hold: a frequency change at a bad moment, an instruction that arrived late, a chart that reads one way and is flown another, a procedure that works on every runway but one. That context is what converts an event into a hazard statement, and it exists in one place only.

The right reflex. Before writing a hazard, ask what a crew would say if you described the event to them. If the answer is obvious to you and you have not asked, you are about to write your own assumption into the safety action plan and then measure it as if it were a finding.

Writing a hazard someone can act on

A usable hazard statement names a condition, bounds it to a context, and stops short of naming the fix. It should be specific enough that a reader can tell whether it applies to a given flight, and general enough that it is not a description of one crew’s afternoon.

The same cluster of events, written three ways
Version Wording Why it does or does not work
A count Twenty-three high-energy approach events this quarter, up from fourteen Not a hazard. Actionable only against the threshold that produced it
A verdict Crews are descending too fast on the arrival Not a hazard either — it names a behaviour and implies a culprit, closing the analysis before it starts
A condition On the arrival to this runway, a track shortening offered below a certain level leaves an aircraft above the profile with limited means of losing energy, and the events cluster on days with a tailwind aloft Names the condition, bounds the context, points at several possible barriers without choosing one

Note what the third version does not do. It does not say whether the answer is a briefing, a procedure change, a conversation with the air navigation service provider, a chart amendment or an acceptance of the risk as it stands. That choice belongs to the risk assessment and to the people who own the barriers, and a hazard statement that has already decided it deprives them of the decision.

Triangulation, barriers, and the honest denominator of confidence

A hazard supported by flight data alone is a hypothesis. Confidence comes from independent sources agreeing: mandatory and voluntary reports covering the same route or runway, normal operations observations, maintenance findings, air traffic recordings, meteorological records. When two sources that cannot influence each other point at the same condition, the finding is worth the cost of acting on it.

Three people in a control room analyze flight data on multiple monitors.
Flight data analysis transforms raw data into actionable safety insights.

The barrier question follows naturally. For the condition described, which defences were supposed to prevent the consequence, and which of them were present, absent, or present but ineffective? A barrier model turns a hazard into a list of things that can actually be changed, and it exposes the case where the only remaining barrier was the crew noticing in time — which is the finding that justifies action even when nothing bad has yet happened.

Finally, the measurement of effect. An intervention can only be evaluated against a baseline that was recorded before it and a measurement rule that did not change during it. That means freezing the threshold, the phase definition and the denominator for the duration, and publishing the result whichever way it goes. A programme that reports only successful interventions is not measuring; it is advocating. Safety managers comparing how other operators structure their review groups and their gatekeeper agreements will find the practical detail on the FL330 forums.

Commentaires

Leave a Reply

Découvrez l’écosystème FL330

FL330, c’est toute l’aviation professionnelle — explorez nos autres espaces.