NEMESynth-EMSO Get more information Account
Standards

RF AI/ML (1900.8)

A model that picks out a radar, a drone controller or a jammer from a waterfall is only as good as the data it learned from, and you can't trust its score unless you know how it was tested. IEEE P1900.8, a draft standard, sets out how to build training and test datasets for RF machine learning, and how to judge the models trained on them.

From the air to a trained model

The draft breaks the work into steps. Each step produces its own dataset with its metadata, so anyone can see what was done to the data and repeat it.

The pipeline: data generation, ingestion, pre-processing, training, and test and evaluation Generate record off the air or synthesize raw data Ingest slice into fixed-size pieces of time raw datagrams Pre-process spectrograms, power spectra, other features cooked datagrams Train the model learns to map features to labels model Test on data it never saw, by stated criteria score Augmentation adds realistic variety: propagation channels, interference, noise
Every step keeps its output and metadata, so a result can be traced back to the recording it came from.

How a dataset is put together

A dataset contains records, records contain RF channels, and channels hold RF data and annotations Dataset metadata for the whole set Record one receive port, band, or time slot RF channel how the receiver was set up RF data IQ, spectra Annotations labeled boxes

A record holds what one receiver port heard on one band or in one recurring time slot. Its RF channel says how the receiver was set up. The RF data is the samples or features themselves: raw IQ, power spectra, spectrograms.

Annotations are the labels. Each one is a box in time and frequency, such as 100 to 120 ms and −10 to +10 MHz, with what's in it: a protocol, an emitter, a signal class. Boxes may overlap.

Every signal is marked as the signal of interest, interference (man-made and unwanted) or noise (natural).

A waterfall with annotation boxes for a signal of interest and an interfering signal over the noise floor signal of interest interference noise floor frequency time
Annotations on a waterfall: each label is a box in time and frequency.

The kinds of jobs it covers

Detection

Is a radar, or any signal of interest, present in this band right now?

Classification

What kind of signal is it: which technology, modulation or protocol?

Fingerprinting

Which individual transmitter is it, from the small flaws in its hardware?

Forecasting

What will this band look like in the next few seconds or minutes?

In NEMESynth-EMSO

A label means more when you know what was supposed to be on the air. NEMESynth-EMSO holds the authorized picture, so a detection can be labeled by whether it matches an assignment, and the sensor keeps the original recording at the edge so the full evidence can always be fetched again.

FeatureStatus
A sensor client that detects energy against a learned baseline, keeps the IQ at the edge and sends compact spectra and eventsIn progress
Waterfalls sent as images whose embedded text points back to the original recordingIn progress
Detections compared with the authorized assignments: friendly use or anomalyPlanned
Measurements and labels exported as datasets in this layoutProposed

IEEE P1900.8 is a draft standard in development. Outline only; not quoted from the draft.

Bring your RF models

If you train or field RF classifiers, talk to us about labeled data tied to the spectrum record and about testing on scenarios like Trazore.

Get more information