How the receiver identifies signals

The channel plan

The receiver doesn’t guess frequencies. RailSpotting sends each receiver a channel plan for its area (GET /api/v1/config): every channel’s frequency, railroad, label (road, dispatcher, yard, EOT, HOT), kind (voice or telemetry) and whether it is a proximity channel. The plan also names a reference station (a NOAA weather radio transmitter) used to check that the radio works. The receiver keeps the last plan it got, so it works through a restart without the network, and picks up a new one when RailSpotting changes it. The website’s receiver page lets the owner turn plan channels off or add a few of their own.

Hearing many channels at once

An RTL-SDR dongle receives a slice of spectrum about 2.2 to 2.6 MHz wide. The receiver turns it into a spectrum every 20 ms or so (a 4096-point FFT) and measures the power in each plan channel (±5 kHz). One dongle therefore watches every channel in its window at the same time: a VHF dongle covers all of an area’s road, dispatcher and yard channels plus the weather station; a UHF dongle covers EOT (or HOT).

For each channel the receiver tracks the noise floor (the quiet level over the last 30 s).

• A channel opens when it is at least 6 dB above its floor for two windows in a row.
• It closes after about 0.3 s below that.
• Each open and close is an event, with its start and end time and signal strength. A call is cut at 120 s, a data burst at 3 s.

The receiver keeps the 0.3 s before each opening, so recordings and decodes include the start.

Voice calls

On voice channels the receiver demodulates narrow-band FM to audio (300 to 3400 Hz) for the calls it needs to listen to: defect detector announcements, and, if turned on, other calls for transcription. Every call is reported as an event either way.

Interference that opens many channels at the same instant (a nearby transmitter, a noisy USB port) looks like calls everywhere at once. On channels watched for detectors, a call that opens together with more than one other channel within a second is dropped as interference.

End-of-train (EOT)

EOT and HOT has the full story: the devices, the signal, the packet and its fields. In short, when a burst on the EOT channel closes, the receiver decodes it:

1. FM demodulation, then the two tones of 1200-baud FFSK (1200 Hz for a 1, 1800 Hz for a 0).
2. It looks for the sync word after the alternating preamble, at eight timing offsets.
3. The 63-bit packet is a BCH(63,45) code word: 45 data bits and 18 check bits.
   • A packet that checks is taken as is.
   • One wrong bit is corrected behind a clean preamble.
   • Two wrong bits are corrected only when two different timing offsets read the same packet, which keeps false decodes out of noise.
4. The fields: unit ID, brake pipe pressure (psi), motion, marker light, marker battery, battery condition and charge, valve circuit, confirmation, turbine, message type.

A burst can hold several packets; the event carries the newest good one and the packet count. Every EOT burst is also saved as a recording (the radio signal itself, state/bursts/), so a missed decode can be looked at later. Weak bursts often can’t be decoded: the signal needs to be roughly 9 dB or more above the noise.

Head-of-train (HOT)

HOT transmissions on 452.9375 MHz are detected, timed and recorded, but not decoded. They are heard by a third dongle in the HOT role, or by the scanner. A HOT burst is a strong hint that a lead locomotive is near, and can make the scanner watch EOT for a while.

Defect detectors

Detectors are recognised in layers:

1. By sound. For detectors the receiver has a template for (built from a few recordings), it compares each call’s sound: the opening identifies the detector, the ending must match its “no defects” all-clear, and the rhythm and length must fit. A clear match is an all-clear (method sound).
2. By words. With speech recognition built in (whisper.cpp, an English model of about 265 MB), the receiver transcribes the whole announcement, using the detector style’s voice profile (its wording and how its voice is usually misheard). It reads the milepost from the words: within 7 miles of the template’s milepost it is that detector, otherwise it looks the milepost up in RailSpotting’s list of detectors. Defect words outside the all-clear phrase make it a defect report.
3. Watched channels. For railroads with no template yet, the receiver watches the channels RailSpotting says nearby detectors use, transcribes calls of 5 to 45 seconds, and reports them as transcripts. Their recordings are kept to build a template later.

A detector usually repeats itself 10 to 40 seconds later; the second announcement is marked as a repeat of the first.

Without speech recognition, the old rule applies: a call longer than the all-clear whose ending didn’t match is reported as a possible defect.