Signals that behave alike
Every circle is a group of signals with a similar spectrum, envelope and phase behaviour. The signal type was never shown to the grouping. Red circles are groups that turn out to be mostly one signal family. Select one to see what gives the family away.
Cluster Hierarchy
Thinned for PreviewFrom raw records to groups you can explain
A broad, known mix of signals
The signals are generated with TorchSig, so the true type of every one is known. They span 69 types in 19 families: PSK, QAM, FSK, OFDM, AM and FM, chirps, ADS-B, LoRa, Bluetooth, Wi-Fi and more, each with channel impairments applied.
150,000 signals, 4,096 samples each, signal-to-noise ratio between 5 and 30 dB.
Separate the signal from the noise beside it
For each capture the noise floor is estimated and the signal is filtered to the band it actually occupies. Everything measured afterwards describes the signal, not the empty spectrum around it.
Bandwidth and centre frequency vary from signal to signal and are not given to the tool.
Describe behaviour, not content
Each signal is reduced to how its energy is shaped in frequency, how its envelope fluctuates, how its frequency moves, and whether squaring it or raising it to the fourth power exposes a spectral line, which reveals the constellation order whatever the carrier offset.
22 features per signal; 14 used for clustering.
Let the signals sort themselves
HDBSCAN groups signals by these features alone. Signal type and signal-to-noise ratio are attached afterwards, only to check the result.
89 clusters; 45 at the finest level.
Explain every group
Each cluster reports which features set it apart: a chirp cluster stands out on linear frequency sweep, an ASK cluster on its power-2 line, an OFDM cluster on its flat-topped spectrum.
Six ratio features and five scale features charted per cluster.
Triage in the browser
Compare an unfamiliar group with known ones side by side, recolour the hierarchy by any feature, and drill into individual signals.
All of it runs in the browser against the live service.
Families found without a classifier
28 of the 45 finest clusters are at least 60% one signal family, covering 10 of the 19 families. The table lists the largest cluster for each. Select a row to open it.
| Signal family | Signals | Share of cluster |
|---|---|---|
| Loading… | ||
Four views of every signal
None of these needs to know the carrier frequency, the symbol rate or the protocol in advance.
Spectrum shape
Occupied bandwidth, how flat the occupied band is, how prominent a carrier is, and how evenly energy is spread.
Envelope
Peak-to-average power and how much the amplitude fluctuates: constant for FSK and FM, noise-like for OFDM.
Frequency behaviour
How the instantaneous frequency is distributed, and whether it sweeps linearly as in chirps and LoRa.
Constellation order
Spectral lines in the squared and fourth-power signal, and fourth-order cumulants, which depend on the constellation and not on the carrier.
The demonstration dataset
Generated for this demonstration with TorchSig, a signal-processing toolkit. The signals are synthetic: realistic waveforms with simulated channel impairments, not recordings. That is what makes the true signal type of every record known.
Dataset Profile
Bring your own collections
DataExplorer reads recorded IQ as readily as generated signals. Talk to our team about an evaluation on your data.
Request a Briefing →