wireless communication
Interference hunting on a shared 470 MHz band: a field method
A 470 MHz link fails intermittently: the sweep, RBW and near-field numbers that find the source, plus the channel-plan decision most hunts get wrong.
Interference hunting on a shared 470 MHz band: a field method
A 470 MHz link that worked all spring starts dropping frames in July. The radio is fine, the antenna is fine, and the RSSI graph shows nothing wrong because the interference is not always on. What follows is the method we use on that band: five calculations done before leaving the office, three measurements done on site, and one decision that most hunts get wrong — whether the problem is a transmitter you can find or a channel plan you have to change.
The band is popular for a simple reason. Between 470 and 510 MHz there is room for a hundred 200 kHz carriers, and both the TP5510 concentrator and the TP5301 transceiver use that exact layout: one carrier, ≤200 kHz, 2.4 / 19.2 / 76.8 kbps. Wide enough to plan around, narrow enough that a single misconfigured device can sit on top of you for months without anyone noticing.
Why a shared band punishes the usual walk-around
Interference hunting on 2.4 GHz is a proximity game: walk toward the noise, watch the level rise, done. At 470 MHz the same walk covers a completely different scale. Free-space loss at 470 MHz is 85.9 dB at one kilometre (ITU-R P.525), and it grows 20 dB per decade of distance. Turn that around: an interferer does not have to be near you to hurt you, and its level falls only 6 dB when you halve the distance.
Two consequences follow, and they decide the whole method.
First, the radius in which an unwanted emitter can desensitise a receiver is measured in kilometres, not metres. Second, once you are inside that radius, every step you take buys very little — the last 100:1 of distance (1 km down to 10 m) is only 40 dB of level change. A receiver with 60 dB of usable display range can therefore cover the whole approach, but only if the analyser is not being pumped by the strong signal you are trying to find.
That is the trap: the closer you get, the more the front end compresses, and a compressed front end shows you a flat noise floor exactly when you need resolution.
Know which of the four interferer classes you are chasing
Almost everything on this band falls into four groups, and each has a signature you can identify from a single screenshot plus a duty-cycle measurement.
Class | Occupied width | Time signature | First test on site |
|---|---|---|---|
Co-channel radio (another 470 MHz device) | 25–200 kHz | Bursts of 20–80 ms, repeating with a polling period | Compare burst length with the victim's frame length |
Switching supply / LED driver | 20–300 kHz wide, comb of harmonics | Continuous, 20–200 kHz repetition inside each burst | Switch the suspect circuit off at the breaker |
Intermodulation product | Narrow, but not on any assigned channel | Present only while two strong carriers are up | Remove one carrier and watch the product vanish |
Broadband hash (bad joint, arcing, PV inverter) | Megahertz wide, ragged | Correlated with load, weather or time of day | Snapshot at two different times of day |
The classes matter because they lead to different fixes. A co-channel radio needs a channel change or an antenna re-aim; a switching supply needs a ferrite and a re-routed cable; an intermodulation product needs physical separation between two innocent transmitters; broadband hash usually needs one connector re-made. If you start with the wrong class you will spend a day measuring the wrong thing accurately.
The last class deserves a warning about baselines. Man-made noise is not one number: ITU-R P.372 keeps separate families of median noise for residential, business and industrial environments, and the recommended values differ between them. A noise floor recorded at a clean office site is therefore not a target for a workshop with welders, and a workshop baseline is not evidence that something has gone wrong later.
Step 1: budget the sweep before you leave the office
A swept analyser sees one narrow slice of spectrum at a time, so its ability to catch a burst is pure arithmetic. Sweep time is the number of resolution bins multiplied by the dwell per bin:
sweep_s = (span_Hz / rbw_Hz) * dwell_s
hit_prob = min(1, burst_s / sweep_s)
N_sweeps = ln(1 - confidence) / ln(1 - hit_prob)Take the numbers we actually use: 40 MHz span (470–510 MHz), 1 ms dwell per bin, and a 33 ms frame — 79 bytes on air at 19.2 kbps, which is a reasonable assumption for a 64-byte payload with preamble, sync, length and CRC.
RBW | Bins | Sweep | Chance per sweep | Sweeps for 90% | Time to 90% |
|---|---|---|---|---|---|
1 kHz | 40,000 | 40 s | 0.08% | 2,790 | 31 h |
5 kHz | 8,000 | 8 s | 0.41% | 557 | 74 min |
10 kHz | 4,000 | 4.0 s | 0.83% | 278 | 18.5 min |
20 kHz | 2,000 | 2.0 s | 1.65% | 138 | 4.6 min |
100 kHz | 400 | 0.4 s | 8.25% | 27 | 11 s |
200 kHz | 200 | 0.2 s | 16.5% | 13 | 2.6 s |

This table explains the single most common field mistake: setting a 1 kHz RBW "to see everything" and then concluding the interferer is not present. At 1 kHz you would need 31 hours of watching to be 90% sure about a 33 ms frame that repeats once a minute. At 20 kHz you need under five minutes. Start wide, confirm the class, then narrow down.
There is a second, harder limit. The frame in this example lasts 33 ms, so no sweep shorter than 33 ms can guarantee capture either. The practical rule is: sweep time within an order of magnitude of the burst length, and dwell long enough that the analyser's own scan does not chop the burst in half.
Step 2: set RBW to the victim's bandwidth, not the span
The display noise floor of any receiver is fixed by bandwidth:
N_dBm = -174 + 10*log10(bw_Hz) + NF_dBRBW | NF 8 dB (receiver path) | NF 18 dB (handheld) | Can it still see -116 dBm? | -111 dBm? |
|---|---|---|---|---|
1 kHz | -136 dBm | -126 dBm | Yes, 20 dB of room | Yes |
10 kHz | -126 dBm | -116 dBm | Marginal | Marginal |
100 kHz | -116 dBm | -106 dBm | No | No |
200 kHz | -113 dBm | -103 dBm | No | No |

Read the last two columns as a hard constraint: with a 100 kHz RBW you cannot see a signal at the sensitivity of the radio you are protecting, no matter how good the analyser is. This is why the published advice from instrument makers starts at 1 kHz RBW for over-the-air work. Anritsu's interference-hunting note quotes a noise floor around -126 dBm at 1 kHz (11410-00972A), and its mobile-hunting companion note recommends 1 kHz RBW with 300 Hz VBW as a starting point, with a preamplifier worth another 10–15 dB (11410-00845C). A handheld with an 18 dB noise figure at 1 kHz lands exactly on the first number, which is a useful sanity check on the formula before you trust it on site.
The compromise most crews settle on: sweep at 100–200 kHz to classify, then park at 10 kHz on the suspect channel and watch the waterfall for a few minutes, and only drop to 1 kHz once you know the burst is periodic.
Step 3: turn the data sheet into a desensitisation budget
Sensitivity is not the level at which interference begins. It is the level at which a 6 dB signal-to-noise ratio still decodes. Interference starts much lower, because the receiver's noise floor is what you are actually protecting.
For the TP5510, -116 dBm at 19.2 kbps over a 25 kHz occupied bandwidth implies a noise floor near -122 dBm and an 8 dB noise figure. In that frame, 1 dB of desensitisation arrives when the unwanted signal is 6 dB below the noise floor: -128 dBm at the antenna port. That is the number the hunt is trying to find, and it is 12 dB below the published sensitivity. The same arithmetic on the TP5301 implies a noise figure near 13 dB at 19.2 kbps and near 11 dB at 76.8 kbps, which is why its desense points sit a few dB higher.
Victim | Sensitivity | Implied noise floor | 1 dB desense at | 3 dB desense at |
|---|---|---|---|---|
TP5510 @19.2 kbps | -116 dBm | -122 dBm | -128 dBm | -122 dBm |
TP5301 @19.2 kbps | -111 dBm | -117 dBm | -123 dBm | -117 dBm |
TP5301 @76.8 kbps | -107 dBm | -113 dBm | -119 dBm | -113 dBm |
The implied noise figures come from assuming a 25 kHz occupied bandwidth for the 19.2 kbps mode and 100 kHz for 76.8 kbps; they are derived, not published, and the derivation is worth redoing if your channel filter differs.
Now run the radius backwards. With a 40 dB excess-loss allowance for plant clutter — walls, racks, containers — the distance at which a 20 dBm transmitter reaches -128 dBm is about 12.8 km in free space terms, or 6.4 km at the 3 dB threshold. A 26 dBm gateway reaches twice as far. Those numbers are not a reason to panic; they are a reason to stop expecting that "walking around with a handheld" will localise anything until you are within a few hundred metres.
Aggressor | Threshold | Radius, free space | Radius, +40 dB clutter |
|---|---|---|---|
20 dBm, co-channel | -128 dBm (1 dB) | ~1,280 km | 12.8 km |
20 dBm, co-channel | -122 dBm (3 dB) | ~640 km | 6.4 km |
26 dBm, co-channel | -128 dBm (1 dB) | ~2,550 km | 25.5 km |
26 dBm, co-channel | -122 dBm (3 dB) | ~1,280 km | 12.8 km |
The free-space column is there to make a point, not to be used: at 470 MHz the arithmetic of a 142 dB link budget produces absurd ranges, which is why every field number in this article carries a clutter allowance and is labelled as an assumption.
Step 4: use the near field for the last metre
Once you are within a few metres, signal strength stops being useful and field behaviour starts being useful. At 470 MHz the wavelength is 0.638 m, so the boundary between the reactive near field and the radiating far field sits at λ/2π = 10.2 cm.
Distance from the source | Level relative to 10.2 cm | Slope |
|---|---|---|
2 cm | +42 dB | 60 dB/decade |
5 cm | +18.5 dB | 60 dB/decade |
10.2 cm | 0 dB | boundary |
20 cm | -5.9 dB | 20 dB/decade |
50 cm | -13.8 dB | 20 dB/decade |
1 m | -19.9 dB | 20 dB/decade |
Inside 10 cm a small loop or a sniffing probe gains 18 dB every time you halve the distance; outside it, only 6 dB. That is the whole reason a near-field probe finds a noisy cable gland in seconds after an hour of walking with a whip antenna has failed. Use the probe to confirm identity, not to measure compliance: inside the near field the reading depends on probe orientation and coupling, not on radiated power.
Step 5: triangulate from bearings, not from level
Bearings are more robust than levels because they do not depend on transmit power, antenna gain or clutter loss. What they do depend on is geometry, and the geometry has a price:
position_error ≈ d * tan(bearing_error) / sin(crossing_angle)Bearing error | Crossing 90° | Crossing 60° | Crossing 30° |
|---|---|---|---|
±5° | 87 m | 101 m | 175 m |
±10° | 176 m | 204 m | 353 m |
±20° | 364 m | 420 m | 728 m |
At one kilometre, a comfortable ±10° bearing puts you within about 200 m — good enough to drive to, not good enough to open a cabinet. Two consecutive cuts, each halving the distance, converge quickly; a single bearing with a wide crossing angle does not. Keep the crossing angle above 45° by taking the second bearing from a position offset by at least the distance to the target, and log the compass heading rather than the visual landmark.
Step 6: check for intermodulation before blaming a neighbour
Two transmitters that are both compliant can create a signal on a third channel:
f_imd3 = 2*f1 - f2 and 2*f2 - f1
P_imd3 = 3*P_in - 2*IIP3Two carriers at 471.5 and 472.1 MHz, both at 26 dBm, put products at 470.9 and 472.7 MHz. If your victim sits at 470.9, the interference will appear only while both carriers transmit, will vanish when either is switched off, and cannot be found by walking toward either one.

Separation between the two carriers | Level at 5 m | Level at 50 m | Level at 500 m |
|---|---|---|---|
0.6 MHz, 26 dBm each | -41.6 dBm | -101.6 dBm | -161.6 dBm |
The single row hides the important part: the product level falls 40 dB per decade of distance, twice as fast as the carriers themselves (20 dB per decade). At 5 m two 26 dBm radios produce a product at -41.6 dBm in a receiver with 0 dBm IIP3 — 80 dB above the desense threshold, and well into blocking territory. At 50 m the same pair produces -101.6 dBm, which is harmless. Intermodulation is a proximity problem even though the carriers are not.
This is the case where the correct fix is not a filter. Move one of the two transmitters, or move the victim's antenna; a band-pass filter in front of the victim will not help because the product is generated inside the victim's own front end.
Failure modes that cost the most time
- Measuring with the preamp on and the attenuator at 0 dB. The front end compresses, the noise floor lifts by the compression, and a -80 dBm interferer looks like a -95 dBm one. Set the attenuator so the strongest carrier sits below the analyser's 1 dB compression point, then add the preamp only when the carrier is gone.
- Hunting a duty-cycled interferer with a slow sweep. Covered above; the fix is a faster RBW setting plus a waterfall, not more patience.
- Assuming the loudest signal is the culprit. The loudest signal is usually a legitimate high-site transmitter. The culprit is the one whose duty cycle matches the frame loss pattern.
- Ignoring the antenna's own pattern. A whip on a metal cabinet has deep nulls; walking in a circle with the same antenna can produce a direction that points at the cabinet wall rather than at the source.
- Changing channel without checking the neighbour's bandwidth. A 200 kHz carrier moved 200 kHz away overlaps its own old channel's skirt, and on a band shared with other users that is how a new interference case is created.
A worked case: 32 nodes, one blocked link
A site with 32 TP5301 endpoints polling one TP5510 concentrator every 60 s has a comfortable margin on paper. Each exchange is a 33 ms frame plus a 20 ms acknowledgement, so the concentrator is busy for 32 × 53 ms = 1.7 s out of every 60 s — 2.8% channel occupancy. At 19.2 kbps the rest of the time is silence, which is exactly what makes the link fragile: there is no continuous carrier to measure, and no LBT to protect the frames.
Item | Value | Source |
|---|---|---|
Endpoints | 32 | Site plan |
Frame + ack per exchange | 33 + 20 ms | 79 bytes at 19.2 kbps, ack assumed |
Occupancy | 2.83% | 32 × 53 ms / 60 s |
Interferer duty cycle | 2.5% (50 ms every 2 s) | Measured on site |
Expected collisions per cycle | 0.80 | duty × 32 frames |
Residual loss after 3 retries | 1.3 × 10⁻⁵ | 0.025⁴ × 32 |
The arithmetic says something counter-intuitive: a 2.5% interferer does not look like a 2.5% outage. It looks like an occasional late packet that the application layer hides with retries — until the interference rises to 10% duty, at which point the residual loss after three retries jumps to 3.2 × 10⁻³ per cycle and the polling schedule starts to slip. That slip, not the packet loss, is what operations notices first.
Field verification: what to record on site
Record enough to reproduce the measurement, or the next crew starts from zero:
date/time local | site id | victim product + data rate
span | RBW | VBW | dwell | preamp | attenuator | antenna + height
noise floor measured at the victim channel (dBm)
burst: start time, duration (ms), period (s), peak level (dBm)
class: co-channel / switching supply / intermod / broadband
bearing 1: position, heading, distance estimate
bearing 2: position, heading, crossing angle
action: source identified? channel changed? retest resultTwo checks close the job. First, re-run the 90% detection calculation with the burst length you actually measured — if the numbers say you could not have seen it, your first sweep was not evidence. Second, re-measure the noise floor at the victim channel after the fix and confirm it came back to the value predicted by -174 + 10log10(BW) + NF, within a couple of dB.
FAQ
Can I hunt interference with one of the radios themselves?
Only for the last step. A TP5301 or TP5510 will tell you that the link is degraded and roughly how often, which is enough to confirm a fix, but neither has a spectrum display or a narrow enough filter to distinguish a -128 dBm interferer from its own noise. Use the radio as the victim and a separate analyser as the instrument.
Why does the interference disappear when I arrive?
Because you are the wrong instrument for that signal, or because the source is load-dependent. If the burst is shorter than your sweep time, moving around does not help; if the source is a switching supply, it may only run when the load runs. Leave a logging analyser or an SDR capture running for 24 hours before you walk the site.
Is a band-pass filter worth fitting?
For adjacent-channel interference from a single strong carrier, yes. For co-channel interference, a filter cannot separate two signals on the same frequency. For intermodulation, a filter at the victim is the wrong end of the problem: the product is generated either in the aggressor's output stage or in the victim's front end, and separation is the only reliable fix.
How wide should the channel plan be?
At 19.2 kbps the occupied bandwidth is about 25 kHz, so a 200 kHz carrier spacing leaves ample guard room. The failure we see most often is not tight spacing but mismatched spacing: one link planned at 200 kHz and a neighbour at 100 kHz, so the second one's skirt lands inside the first one's filter. Measure the occupied bandwidth of every device on the site before deciding the plan; the products page lists the channel bandwidth for each model, and the field notes on this band cover the trade-offs in more depth.
If you are planning a site rather than fixing one, the cheapest hour you can spend is the one before installation: pick the channels so that no two devices share a carrier, check the nearest high-power transmitter, and write the plan into the commissioning record. Interference hunting is what you do when that hour was skipped.



