NTRIP-Analyser

Analyses NTRIP RTCM 3.x streams from GNSS base stations, and answers whether a station is fit to serve RTK.

View the Project on GitHub pe1mew/NTRIP-Analyser

Thresholds — every number that decides a verdict

Both tiers of measurement end in a word: STATION OK, CAUTION, FAILED for the ninety-second check, and STABLE, DEGRADED, UNSTABLE for the stability report. Behind every one of those words is a number someone chose.

This page lists all of them, with the rule each is used in and the reasoning behind the value. It exists for three reasons:

  1. So a verdict can be argued with. “Failed” is not useful unless you can see what it was measured against.
  2. So the numbers can be checked. Several here are starting points rather than settled facts, and this page says which.
  3. So they can be changed on purpose. A control network and a hobby base are not held to the same standard, and neither should inherit the other’s numbers by accident.

Every threshold lives in src/core/ — never in the CLI, the GUI, the service or the Android app. That is what makes a free verdict worth the same as a paid one, and a desktop verdict the same as a phone’s.

Every screen shows the limit it used. The CLI’s --check and --report, and the GUI’s Station Check and Stability windows, carry a limit column beside each value: min 40.0 dB-Hz, max 1.00 /h, min 29. It is set where the verdict is decided, not written into a frontend, so a screen cannot display a threshold the engine is not using — and KPI 5 shows the figure this station was held to, which depends on the constellations it streams.

You can override them. The CLI takes --thresholds <file>, a JSON policy that is partial by design: it carries only what you disagree with, and every key it omits keeps the value below. A bad setting is refused with the field named, and nothing is half-applied — a policy that was only partly accepted would produce a verdict belonging to no stated standard at all.

ntrip-analyser --thresholds-print                 # what am I judging by?
ntrip-analyser --thresholds survey.json --check   # judge by that instead

--thresholds-print lists every threshold in force, whether it came from a file or the build, and what it means; it needs no config and no network. A run under a non-default policy names it, with a fingerprint over the effective values — because once verdicts can be produced under different standards, STATION OK is not comparable between two users unless each says which standard it used.

bin/exampleThresholds.json is a worked example. The GUI and the service still use the built-in values; see the design track.

How to read a rule

Most thresholds have two levels, and the middle band is the useful one:

Band Tier 1 Tier 2 Means
Good PASS STABLE at or better than the threshold
Middle WARN DEGRADED worth looking at; not yet unusable
Bad FAIL UNSTABLE will cost you fixes in the field
No answer PENDING INSUFFICIENT EVIDENCE not enough evidence yet — a state, not a failure

The last row matters as much as the others. A measurement that has not had time to be made is reported as exactly that, never as a pass.

Tier 1 — the acceptance check (src/core/kpi.h)

Eight checks, about ninety seconds, answering is this station fit to serve RTK right now?

1. Connected and producing

   
Threshold KPI_MIN_BYTES_PER_S = 100 bytes/s
Rule sustained throughput at or above it passes

Rationale. A four-constellation MSM7 station at 1 Hz runs about 1.7 kB/s, and even a single-constellation MSM4 base clears 300 B/s. 100 is therefore not a quality bar but a liveness bar: below it the stream is not delivering corrections at all, whatever the socket says. Set low deliberately, because this check exists to separate “nothing is arriving” from every other kind of problem, and a liveness test that also judges quality tells you neither.

2. RTCM 3.x format

   
Threshold none — structural
Rule at least one CRC-valid RTCM 3 frame must decode

Rationale. Not a number: either the bytes frame and pass CRC-24Q or they do not. A caster serving RTCM 2, NMEA or an HTML error page fails here, which is why it is separate from check 1.

3. Reference position (ARP)

   
Threshold KPI_ARP_DEADLINE_S = 30 s
Rule a 1005/1006 with non-zero coordinates must arrive within it; before that the check is PENDING, not failed

Rationale. 1005 is customarily sent every 10 s, so 30 s allows three attempts and one to be lost. Under ten seconds the check would fail stations that are merely between broadcasts — punishing a base for the gap between its own messages.

4. Observations flowing

   
Threshold KPI_MSM_MAX_DT_S = 2.0 s (0.5 Hz)
Rule every constellation the station streams must average an epoch at least this often

Rationale. RTK needs corrections at 1 Hz; 2 s is one missed epoch of tolerance, so a station is not failed for a single hiccup while a base that has genuinely halved its rate is caught. Measured per epoch, not per frame — MSM splits one epoch across several frames, so counting frames would report a healthy base as transmitting at a multiple of its true rate.

5. Satellites in view

   
Thresholds KPI_EXPECT_SATS per constellation; KPI_EXPECT_UNKNOWN = 25 as the fallback, which is KPI_MIN_SATS
Rule at or above the expectation passes; at or above half of it warns; below that fails

Per-constellation expectations, above a 10° mask at mid latitude:

GNSS Expected
GPS 8
GLONASS 6
Galileo 6
QZSS 1
BeiDou 8
SBAS 2
NavIC 2

The fallback applies only when neither the sourcetable nor the stream says which constellations to expect — otherwise a station is held to the sum over the constellations it actually streams.

Rationale. A table rather than one number, because a GPS+GLONASS station cannot reach a flat 25 however healthy it is, and failing it for that is failing it for its age. The values are deliberately modest — a number a station should comfortably exceed rather than its best case — because they decide a verdict. Checked against real streams: Kadaster’s APEL00NLD0 delivers 47 across five systems where this table expects 29, and Centipede’s NEAR delivers 40 across five where it expects 30.

The half-of-expected warn band is the interesting part: a station holding 60 % of what it should is usually a siting or antenna problem rather than a fault, and it deserves a different word from one holding 20 %.

6. Median C/N0

   
Threshold KPI_MIN_CNR_MEDIAN = 40 dB-Hz
Rule at or above passes; at or above 90 % of it (36 dB-Hz) warns; below fails

Rationale. 40 dB-Hz is the conventional floor for a healthy geodetic antenna and LNA chain; below it, ambiguity resolution starts costing time. The median, not the mean, so a handful of low-elevation satellites cannot drag the figure down. The 10 % warn band exists because antenna chains age gradually — the useful signal is a station that has moved into the band, which is why tier 2 measures the fall rather than the level.

7. Frame integrity (CRC)

   
Thresholds KPI_MIN_INTEGRITY_PCT = 99.9 % of frames passing; KPI_BAD_INTEGRITY_PCT = 99.0 %
Window KPI_INTEGRITY_WINDOW_S = 60 s, which is KPI_SUSTAIN_S
Rule at or above 99.9 % passes; at or above 99.0 % warns; below fails

Rationale. Reported as the share of frames that passed — 100 % is a clean stream and the figure falls as frames fail — because that is how the question is asked. 99.9 % is the same standard as one error per thousand frames (KPI_MAX_CRC_RATE), and tier 2 uses the identical pair, so the two tiers cannot disagree about what a clean stream is.

On the name. This is frame integrity, the share of frames that passed CRC-24Q. It is deliberately not called FER. Consistent with BER, FER would be Frame Error Rate — the complement of this figure — and in codec usage “erasure” means a frame that never arrived at all, which here is a dropout rather than a corruption and is reported by availability and delivery rate instead. Neither RTCM 10403.x nor NTRIP defines an error-rate metric; the term the field uses for the complement is CRC error rate, which is what the detail lines say when they describe the failing side.

A clean TCP path delivers essentially zero CRC failures, so anything measurable points at the serial link, the radio, or a caster mangling bytes.

Measured over the last 60 s, not over the session. A figure accumulated since the stream opened dilutes in both directions: a burst in the first seconds is washed out by every clean second after it, and one late in a long run is washed out by everything before. Either way the verdict stops being able to move — and a verdict that cannot move makes the sustain clock meaningless, since it would be timing a number that can no longer change. The window needs at least 100 frames before it will report; below that it stays PENDING.

8. Advertised versus actual

   
Threshold none — comparison
Rule a type the sourcetable advertises but never sends is a failure; a type sent but never advertised is a warning

Rationale. The asymmetry is the whole point. A rover configured from the sourcetable will not receive what it was promised — that is a fault. Extra data nobody advertised is data the rover can use; only the metadata is wrong. Constellations are read from the STR nav-system field rather than the 1005/1006 indicator bits, because those bits cover GPS, GLONASS and Galileo only and a BeiDou-capable base cannot declare itself there.

Tier 1 — VRS assertions (src/core/vrs_check.h)

Added automatically when the mountpoint is classified as a network service. They test behaviour a fixed base never claims to have.

# Assertion Threshold Rationale
A1 The caster accepts the GGA VRS_ACCEPT_S = 5 s A network service that objects to a position does so immediately; a disconnect later than this is unrelated.
A2 Corrections start after the GGA VRS_RTCM_S = 10 s A VRS generates a stream for your position on demand. Ten seconds is generous for that; longer suggests it is not generating anything.
A3 The ARP is near the rover VRS_ARP_MAX_KM = 50 km A virtual or nearest-base ARP should be within the network’s cell. Beyond 50 km it is not serving your position, whatever it is called.
A4 The stream holds at the GGA cadence VRS_HOLD_S = 60 s One minute covers several keep-alive intervals, so a stream that only survives while you hammer it is caught.
A5 The gate test VRS_GATE_S = 90 s Stop the GGA and a real network service drops you. Ninety seconds is longer than any caster’s keep-alive tolerance seen in practice. Opt-in: it ends the session it is testing.

A5 not dropping is a classification, not a fault: it means a fixed base behind a network-looking name.

These five are overridable like the rest, under a vrs section of a policy file — accept_s, rtcm_s, arp_max_km, hold_s, gate_s. They are the least likely of the set to need changing, describing what casters actually do rather than what a station ought to achieve, but a network with unusual keep-alive behaviour would otherwise have no recourse at all — and “unlikely to need changing” is not a reason to make something unarguable.

Tier 2 — the stability report (src/core/station_report.h)

Six measurements over hours, answering has this station been fit, and is it staying that way? Same data, different question — and deliberately a different vocabulary, so that a station which is fit now and was unstable all week can say both without contradicting itself.

# Measurement Degraded at Unstable at Measures
1 Availability SR_RECONNECTS_WARN_PER_H = 1.0 /h SR_RECONNECTS_BAD_PER_H = 4.0 /h reconnections per hour
2 Frame integrity SR_INTEGRITY_WARN_PCT = 99.9 % SR_INTEGRITY_BAD_PCT = 99.0 % the lowest share of frames passing CRC in any 10 minutes
3 Signal level SR_CNR_DROP_WARN = 3 dB-Hz SR_CNR_DROP_BAD = 6 dB-Hz the fall from the window’s best mean C/N0
4 Satellites held SR_SATS_WARN = 25 SR_SATS_BAD = 15 the fewest held at any moment
5 Ionosphere IONO_ROTI_UNSETTLED = 0.5 TECU/min IONO_ROTI_DISTURBED = 1.0 TECU/min the worst median ROTI
6 Delivery rate SR_OFFRATE_WARN = 10 % SR_OFFRATE_BAD = 33 % share of samples with an advertised type off its rate

Why these values

1. Availability. One reconnection an hour is a link worth investigating; four is a link nobody can survey on. Both are about the path, not the base — a station on domestic broadband will show this where the same receiver on fibre will not.

2. Frame integrity uses exactly tier 1’s pair — 99.9 % passing to warn, 99.0 % to fail — so the two tiers do not disagree about what a clean stream is.

The lowest reading in any ten minutes, not the figure for the session. This distinction is the whole metric. The rate carried on the statistics snapshot is cumulative since the session opened, and the extreme of a running mean fails in both directions: two errors against a small early denominator read as 0.93 % and are banked as the “worst” for the rest of the run, while a burst of corrupted frames six hours in moves that cumulative figure by hundredths of a per cent and is never noticed. Both were observed on a live stream, in the same session.

So integrity is measured over SR_INTEGRITY_WINDOW_S = 600 s of stream — ten minutes, matching SR_MIN_WINDOW_S, so the first reading arrives exactly when the report first has the evidence to judge anything at all — and the worst of those readings is what the report carries.

The window is a choice with a cost worth stating: a bad minute inside a good ten shows up diluted by ten, so a minute at 99 % reads as 99.9 % across the window. It is caught, but as DEGRADED rather than UNSTABLE. A shorter window is more sensitive to bursts and noisier on slow stations; ten minutes is the length at which a reading describes something a person would call an outage.

3. Signal level is measured as a fall, not a level. An absolute C/N0 says more about the site than about the station: a rooftop in a city is not a field. A drop says something changed — water in a connector, a loose cable, a new obstruction. 3 dB is half the power and is the smallest change worth a message; 6 dB is a quarter and is a fault. A station that falls 7 dB to 41 dB-Hz is flagged although 41 would pass tier 1 outright, and that is the intended behaviour.

4. Satellites held uses tier 1’s fallback as its warn level so the tiers agree; 15 is roughly the point below which RTK stops resolving in reasonable time. This is the minimum over the window, so a single bad moment is visible — which is why the report ignores the first 30 s (see below).

5. Ionosphere reuses core/iono.h’s own scale rather than inventing one: space weather does not mean something different here than it does there. Under 0.5 TECU/min is quiet, over 1.0 is disturbed. This is the one measurement on the page that is nobody’s fault — it exists so that poor RTK during a storm can be attributed rather than blamed on the base.

6. Delivery rate counts samples in which any advertised type was arriving off its advertised rate. A tenth of the window is a station that stutters; a third is a station whose sourcetable entry is fiction.

Evidence rules — not thresholds about the station

These decide how much proof is needed before any verdict is offered. They are not judgements about a station’s quality, and mixing them up with the ones above is how a report starts grading an hour’s question on a minute’s data.

Rule Value Why
KPI_SUSTAIN_S 60 s Tier 1’s verdict must hold unchanged for a full minute before it is reported. A station that flickers cannot pass by being healthy at the right moment — and a number that keeps moving cannot be quoted in a handover.
Tier 1 ceiling 300 s A run that cannot conclude must still end. A mountpoint absent from the sourcetable leaves check 8 unable to judge — rightly, since “could not check” is not a pass — and without a ceiling the run would never finish.
SR_WARMUP_S 30 s Tier 2 ignores the first 30 s of a session. sats_total describes the last five seconds, so a sample taken while the first epoch is still arriving sees a partial constellation. Found the hard way: a station that never dropped below 39 satellites was reported as holding 9.
SR_MIN_WINDOW_S 600 s No tier-2 verdict at all below ten minutes.
SR_MIN_SAMPLES 10 And not without ten samples, however long the window claims to be.
SR_STALE_S 120 s Wall-clock silence of the stream clock after which tier 2 offers no verdict at all. Every window here is stream time, which is what makes a replay reproduce a live run — and also means that when a stream stops, its window stops with it: the samples keep arriving, the numbers never change, and the report describes a period that ended. One station published STABLE over 1.7 h for fourteen hours after its last observation on exactly that arithmetic. Two minutes sits above stall_timeout_s (60 s) so the session reconnects first; what reaches here is the case reconnecting cannot fix.
KPI_INTEGRITY_WINDOW_S 60 s Stream each tier-1 integrity reading covers, and deliberately KPI_SUSTAIN_S: a check has to be able to change within the run, or the sustain clock is timing a number that can no longer move. At least 100 frames must arrive in it.
SR_INTEGRITY_WINDOW_S 600 s Stream each tier-2 integrity reading covers, matching SR_MIN_WINDOW_S so the first arrives when the report can first judge.
Sampling cadence 1 s of stream Tier 2 samples once per second of stream time, in every program. A replay at disk speed therefore produces the same number of samples as the live run did.

Tier-2 windows are measured in stream time, taken from the observation epochs rather than the host clock. A dropout counts, because the epochs on either side say so; an NTP correction cannot distort a running window; and a captured file is judged over the window it holds rather than the seconds it took to read.

What is not a threshold

Numbers that look like thresholds and are not:

Which of these are settled, and which are guesses

Stated plainly, because a page like this reads as more authoritative than it is:

Confidence Thresholds
Well founded — conventional, or checked against real streams KPI 6 (40 dB-Hz), KPI 5’s per-constellation table, KPI 7 (1 in 1000), the ROTI scale
Reasoned, not measured — defensible, but no study behind the exact value KPI 1 (100 B/s), KPI 3 (30 s), KPI 4 (2 s), the whole of tier 2’s six pairs, all five VRS assertions
Learned from a defect — the value exists because something was wrong without it SR_WARMUP_S (30 s), KPI_SUSTAIN_S (60 s), SR_STALE_S (120 s), the 300 s ceiling

The middle row is the honest answer to “are these correct?”: they are plausible, they behave sensibly against the stations this project has measured, and none of them has been derived from a population of stations. That is the strongest argument for making them overridable — and for showing them on screen beside the verdict they produced.

See also