Satellite Fault Detection and Exclusion: The Conditions for the Mechanism to Work

Ciel nocturne en longue exposition au-dessus d'un radôme d'aérodrome

About thirty miles before the initial approach fix, the receiver shows nine satellites tracked, and yet the avionics announce that the required accuracy is no longer guaranteed. Nothing is broken on board, the constellation is nominal, no one has entered a wrong procedure. What is missing is not signal, it is exploitable redundancy: four satellites are enough to compute a position, five to cast doubt on it, six to name the culprit — and even then, only if the geometry allows it.

Four unknowns, so four satellites: redundancy begins with the fifth

Satellite position computation solves four unknowns: three coordinates and the receiver clock bias. Four pseudoranges therefore give an exactly determined system. The solution is unique and, above all, it leaves no residual: all the information has been used to adjust the four parameters, so the measurements are consistent by construction. A satellite broadcasting a false clock correction or a stale ephemeris shifts the solution without leaving a trace.

The fifth measurement changes the nature of the problem. The system becomes overdetermined, with one degree of freedom, and the fit leaves a residual vector that stops being zero as soon as the measurements contradict each other. It is there, and only there, that a possible test is born: a statistic is formed from these residuals — or from the gap between the full solution and partial solutions — and compared against a threshold calibrated on the accepted false-alarm rate. This test answers one question, and only one: are the measurements mutually consistent? It does not say which one is lying.

One architectural detail weighs on what follows: an independent barometric altitude measurement plays the role of a pseudo-measurement. It barely improves horizontal accuracy but contributes a degree of freedom; when this input is credited, the required source count drops by one, and it silently rises again when the input disappears.

You need at least four satellites to calculate a position, but with five, you can check if they agree. If they don’t, you know something’s wrong, but you can’t tell which satellite is faulty. With six, you can remove one at a time to see which one fixes the problem.

An A320 on approach to Paris-CDG with six satellites in view. One satellite has a faulty clock. The system detects the inconsistency but can’t identify the culprit. The pilot sees “Required Accuracy Not Available” but the aircraft remains flyable. If a seventh satellite were available, the system could exclude the faulty one and restore full accuracy.

Why exclusion requires one more source, and why this count is not enough

Five satellites, and the test trips: one of the measurements is at fault — can it be named? No, and the reason is purely algebraic. Removing any given satellite brings the system back to four measurements for four unknowns, hence zero residuals. All five sub-solutions are consistent, all five candidates look innocent: detection is achieved, identification is out of reach.

With six, the situation reverses: after removing one satellite, five remain, hence one degree of freedom, hence a test still workable on the surviving subset. Excluding is nothing other than detecting within each sub-solution — one looks for the removal that restores consistency while checking that the others do not. Hence the rule: five to detect, six to exclude.

What four, five, then six independent measurements make possibleA table with three columns, four, five and six measurements, and five rows. The number of degrees of freedom goes from zero to one, then to two. The position is computable in all three cases. Detection only appears from five measurements onward. At five measurements, none of the five sub-solutions missing one measurement is testable; at six measurements, all six are, which makes exclusion possible. A filled disc marks the capability as available, a dash marks its absence.4 measurements5 measurements6 measurementsDegrees of freedom012Position computableDetection possibleTestable sub-solutions0 of 56 of 6Exclusion possible
Each capability is bought with a degree of freedom, not with a satellite. At four measurements, residuals are zero by construction: there is nothing to test. At five, the contradiction becomes visible but anonymous, since none of the five sub-solutions missing one measurement retains a residual to examine. At six, all six subsets remain testable — a necessary condition for exclusion, though not sufficient, since the surviving geometry still has to hold up.

This rule is necessary but never sufficient, and that is where cockpit reasoning goes astray. For exclusion to be available, detection must remain possible in each of the conceivable subsets, including the worst case. Yet a sky with six satellites often owes its horizontal geometry to just one of them, low on the horizon, isolated in an azimuth that no one else covers. Remove it and the remaining five bunch together: the solution stays computable, but monitoring goes blind in one direction. The availability of exclusion is the intersection of the availabilities of all sub-geometries, a product of constraints, not a sum of satellites.

There is also a risk specific to exclusion: naming the wrong culprit. Discarding a healthy satellite while keeping the faulty one leaves the fault in the solution and removes a useful measurement. The integrity budget therefore reserves a share for this wrong exclusion, in addition to the missed-detection and false-alarm terms: excluding costs more budget than detecting.

Geometry sets the price: slope, protection level, alert limit

The test lives in measurement space, the alert limit lives in position space. What links the two is geometry, and the useful parameter has a name: slope. For each satellite, it expresses how many metres of horizontal error are produced per metre of observable residual. Low slope, the fault is seen before it does harm; high slope, the satellite is nearly redundant with the rest of the sky and can pull the solution very far while leaving only a tiny residual.

The protection level is built on this idea. The worst slope in the current sky is taken, multiplied by the detection threshold, the contribution of nominal noise at the required probability is added, and the result is a radius that bounds the position error for the allocated integrity risk. The function remains available as long as this radius stays under the alert limit associated with the current operation. These limits tighten as one gets closer to the ground: on the order of several nautical miles in cruise, one mile in the terminal area, a fraction of a mile on lateral-guidance-only approach. The value to use is the one from the applicable navigation specification and approved documentation, never a value remembered or read in an article.

The consequence is counter-intuitive: one more satellite in an azimuth already covered improves accuracy slightly and integrity almost not at all, whereas a satellite opening up a new direction can flip availability in one stroke. Dilution-of-precision indicators average things out where integrity looks at the worst case.

Sky view before and after removing the isolated low satelliteTwo sky views in azimuthal projection. On the left, six satellites: five clustered in the same sector and one alone, low on the horizon, in an azimuth that no one else covers. On the right, that satellite is removed and marked with a cross: the remaining five form a cluster, and monitoring loses all sensitivity in that direction.six sources visibleone source removedlow azimuth coveredcluster, blind direction
The availability of exclusion is judged on the worst sub-geometry, never on the satellite count. The low, isolated satellite in the left-hand view single-handedly holds up that azimuth. Once removed, it leaves a cluster: the position remains computable, but an error from a member of the cluster no longer produces an exploitable residual.
Detection alone versus detection with exclusion: what actually changes
Criterion Detection Detection and exclusion
Independent measurements required Five, with barometric altimetry able to count if credited Six at minimum, with detection required to remain possible after removal
Geometric condition Worst-case slope compatible with the alert limit Same condition verified in each subset missing one satellite
What is protected Integrity: no out-of-tolerance position without an alert Continuity: the function survives the fault
Risk budget consumed Missed detection and false alarm Same, plus a share for wrong exclusion
Visible effect in the cockpit Alert, loss of capability, mode fallback Nothing, or a discreet shift in margins

Detecting protects integrity, excluding protects continuity

This distinction is the key to the whole subject. Detection is a safety device: it guarantees that one will not be left navigating on a false position without being told, nothing more. Exclusion is an availability device: it lets the function survive the loss of one source. An architecture that detects without excluding is not less safe, it is less available.

A pilot operates an aircraft cockpit at dusk, with a navigation display showing 'Required Accuracy Not Available'.
Pilot navigating with degraded satellite accuracy, highlighting system failure risks.

In the cockpit, this produces two very different families of events. Losing exclusion capability means seeing nothing as long as no fault occurs, then facing the alert and the fallback if it does. Losing detection capability means immediately losing the right to use the function for the current operation, regardless of any faulty satellite: the displayed position has not moved and is probably still excellent, but the proof that it is good has vanished. Hence the value of a genuinely prepared fallback: one that assumes a lengthy reprogramming or an unbriefed track is no longer a fallback when it is actually needed.

Protection radius at the moment of a fault, with and without exclusionCurve of the protection radius versus time. It rises slowly, then a fault occurs. With exclusion (solid line), the radius makes a discreet jump but stays under the alert limit, before crossing it later under the effect of the sky’s rotation. Without exclusion (dashed line), the radius immediately exceeds the alert limit and the function is lost on the spot.protection radiustimealert limitsatellite faultwith exclusionwithout exclusionalert
Detecting and excluding do not look the same in the cockpit: one shuts the operation down, the other quietly eats into a margin. The solid line shows successful exclusion — a jump in radius, no annunciation, a margin grown thin. The dashed line shows the same fault without exclusion capability: an immediate breach of the limit and loss of the function.

The type of installation shifts the whole issue, which is why a reflex learned on one aircraft cannot be transposed onto another. On a stand-alone receiver, the loss of monitoring is immediate and total: no other source bounds the error, the function shuts down and only conventional navigation remains. On a hybridised chain, where the computer blends satellites, inertial reference units and ground aids, the position does not disappear: it is the credited performance level that drops, and the drift of the backup source sets the length of time one can continue before a re-alignment. Two questions therefore arise before the flight, and only once for the whole fleet: which source takes over when satellite monitoring becomes unavailable, and for how long the approved documentation credits it for the operation envisaged.

The trap. Successful exclusion does not announce itself. When the mechanism discards a faulty satellite, the function carries on, no message reports it, and the surviving geometry can end up just under the alert limit. The next satellite to drop below the masking angle, or a turn that occults it, triggers the alert with no warning sign. The same mechanics apply to an availability computed for the planned arrival time: twenty minutes of holding, and the sky is no longer the one that was assessed.

Multi-constellation: the count stops being the constraint, the fault model becomes it

With two usable constellations, the question of the sixth satellite dissolves: the receiver commonly has far more sources than the algebraic minimum. What takes the place of counting is more subtle: the fault model.

Advanced formulations of autonomous integrity monitoring reason by hypothesis: each fault mode corresponds to a partial solution excluding it, and the protection level bounds the gap between the full solution and each of them, weighted by the mode’s prior probability. With narrow faults — one satellite at a time, independent — the number of sub-solutions stays modest. As soon as wide faults are admitted, correlated at the scale of a whole constellation because they trace back to a ground segment or a shared time reference, sub-solutions missing an entire constellation are needed. And the geometry problem returns intact: a sky amputated of one constellation can be a bad sky.

A second, heavier shift: these calculations rest on parameters that are not measured on board — the range accuracy assigned to each satellite, the prior probabilities of narrow and wide faults. They come from outside, through support data maintained at the service level, and the onboard algorithm is never more reliable than these assumptions.

What onboard integrity monitoring covers, and what escapes it
Threat Signature on the measurements Handling by the mechanism
Fault of one satellite (clock, ephemeris) One measurement inconsistent with the others Core of the model: detected, then excluded if the sky allows it
Wide fault across a constellation Several measurements wrong in a correlated way Covered if the hypothesis is provided for and if the geometry survives removal of the constellation
Jamming Degradation then loss of tracking Outside the model but honest: the function degrades visibly
Coherent spoofing Mutually consistent measurements, normal residuals Blind spot by construction: nothing to detect in internal consistency
Common fault on the aircraft side (antenna, receiver) Error common to all measurements Invisible to the consistency test: falls under equipment monitoring

Predicting instead of hoping: what makes a prediction useful

An availability prediction is not a status check, it is a geometry calculation for a place, over a time window, with an assumed constellation — three chances to get it wrong. The window: the sky rotates, a prediction valid at the planned time says nothing about a slot shifted by half an hour, and availability gaps are narrow and mobile. The place: a prediction made for the destination covers neither the alternate nor the missed approach, which is precisely the moment when the function will be needed. The assumed constellation: a satellite announced unavailable by notice to airmen must be removed from the calculation, and nothing guarantees that the tool used does so on its own.

There remains the aircraft, which the calculation ignores. The masking angle used is a design value; the airframe, the antenna, a prolonged bank in a turn or in holding occult low satellites that the prediction counted on. A prediction is therefore optimistic by construction. And if the barometric input is credited, degradation of the altitude source removes a degree of freedom: detection goes from available to unavailable without any satellite having moved.

The right reflex. Treat the prediction as a dated, localised calculation, not as a green light. Redo it for the actual window as soon as a delay sets in, extend it to the alternate and the go-around, and check what your architecture credits: if the barometric input counts, the altimetry chain becomes a condition for navigation availability. Finally, know which annunciation, on your fleet, means « exclusion lost » and which means « detection lost »: the first leaves options open, the second closes the operation.

What to go and find in one’s own documentation

The reasoning above is generic; its application depends on approval choices that are, themselves, written down. Four elements weigh more heavily here than the state of the constellation.

Four documentation elements, what they change, and the cost of ignoring them
Element to locate What it changes in the mechanism What one suffers for not having checked it
Integrity architecture credited by operation type Designates the fault model the function depends on: purely onboard monitoring, geostationary augmentation, or multi-constellation hypothesis-based reasoning In flight, one relies on a capability that is not credited for the operation, or the opposite: one forgoes a minimum that remained available
Barometric input credited or not Varies the number of required sources by one and ties navigation to the altimetry chain An altimetry degradation brings down monitoring even though the sky has not changed, and the diagnosis heads in the wrong direction
Correspondence between annunciations and internal states Translates each message into loss of continuity or loss of integrity One continues on an annunciation that had closed the operation, or discontinues on one that had left options open
Assumptions of the prediction tool used Sets the masking angle used, the assumed constellation, and the handling of unavailability notices A favourable prediction is contradicted in flight, typically when a satellite withdrawn by notice was left in the calculation

The share the mechanism will not take on

Onboard integrity monitoring reasons on the internal consistency of a set of measurements. Anything that is coherently wrong escapes it, and this limitation lies in the definition of the test, not in its implementation. This is why independent cross-checks keep their value — position from ground aids, plausibility of the route and remaining distance, consistency between altitude sources — the only recourse against a class of errors the mechanism does not see.

For anyone preparing a qualification or a proficiency check in performance-based navigation, the subject is rarely decided by the number. Knowing that exclusion requires six sources is only useful if one can say why five is not enough, why six is not always enough either, and what distinguishes an alert caused by a faulty satellite from an unavailability of monitoring. It is this difference in depth that FL330 training courses work on.

The foregoing describes mechanisms and does not constitute instruction. The applicable regulatory text, the certification specifications of the installed equipment, and the operator’s approved documentation prevail in all circumstances over this article, including where they adopt assumptions or minima different from those discussed here.

Questions fréquentes

Why does detection require five satellites and exclusion six?

The position calculation has four unknowns, so four pseudoranges are enough to solve it exactly, leaving no residual to test. The fifth measurement creates a degree of freedom, hence an observable inconsistency: that is detection. With only five sources, removing any one of them brings the system back to being exactly determined, so no candidate can be named; with six, five remain after removal and the test is still workable on the subset, which makes identification possible.

Does having six satellites in view guarantee that exclusion will be available?

No: a count of six is a necessary condition, never a sufficient one. For exclusion to be available, detection must remain possible in each subset missing one satellite, including the worst case. Yet a sky with six sources often owes its horizontal geometry to a single satellite that is low and isolated in azimuth; removing it makes the protection level jump beyond the alert limit. The availability of exclusion is therefore the intersection of the availabilities of all sub-geometries.

What is the operational difference between losing detection and losing exclusion?

Losing detection immediately removes the right to use the function for the current operation, regardless of any actual fault: the displayed position is probably still excellent, but the proof that it is good has vanished, and the mode falls back. Losing exclusion is not noticeable as long as no satellite fails; the day it does, one faces the alert and the fallback instead of continuing. Detection protects integrity, exclusion protects continuity.

Does barometric aiding really replace a satellite?

It contributes an independent measurement, hence one extra degree of freedom to the system, and lowers by one the number of satellites needed for both detection and exclusion when this input is credited by the certification of the installed equipment and by the approved documentation. However, it barely improves horizontal geometry, which remains governed by the azimuthal spread of the satellites: it buys monitoring availability, not lateral accuracy. The consequence to remember: degradation of the altitude source can make monitoring unavailable without any satellite having moved, and the diagnosis then readily starts from the wrong side.

Does onboard integrity monitoring protect against jamming and spoofing?

Jamming shows up as degradation followed by loss of tracking, hence visible behaviour, even though it falls outside the fault model. Coherent spoofing, on the other hand, produces a set of mutually consistent measurements: the internal consistency test has nothing to detect and stays silent. It is for this class of errors that independent cross-checks — positions from ground aids and plausibility checks — keep their full value.

Does an availability prediction made before departure remain valid in flight?

It is only valid for the place, the time window and the constellation assumed at the time of the calculation. The sky rotates: a slot shifted by half an hour is no longer the one that was assessed, and availability gaps are narrow and mobile. The calculation must therefore be redone for the actual window as soon as a delay sets in, extended to the alternate and the missed approach, and it must be checked that satellites declared unavailable by notice to airmen have indeed been removed from the calculation.

Commentaires

Leave a Reply

Découvrez l’écosystème FL330

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