DySPAN Standards Committee

IEEE 1900.8

IEEE 1900.8 Working Group on Machine Learning for RF Spectrum Awareness in DSA and Sharing Systems

Purpose

This standard defines a common approach for creating datasets used to train and test machine learning models that detect, classify, characterize, and/or identify radio frequency signals and signal emitters. The standard also defines a criteria for evaluating the performance of the machine learned spectrum awareness models.

Need

There is no existing standard for how to generate high quality datasets in order to train and test Machine-Learned Spectrum Awareness (MLSA) models or how to evaluate their
performance. Openly available MLSA datasets suffer from insufficient data veracity and sample sizes to develop models that meet the performance requirements of real-world scenarios. Available datasets also suffer from insufficient representation of the variability of realistic spectrum operations that include a range of RF propagation channel effects and interference signals – greatly limiting the performance of the MLSA models once deployed in the real-world.

Therefore, MLSA model developers are forced to create their own unique approach for generating datasets and evaluating model performance – greatly increasing the risk of incurring unexpected costs, encountering hidden technical challenges, and creating models that are unable
to meet the performance requirements of real-world scenarios.

We expect that this standard will also be adopted by the research community that engages in the sharing and exchange of machine learning datasets. The open availability of high-quality MLSA datasets will greatly accelerate improvements to the state-of-the- art of MLSA model design, similar to how the computer vision research community rapidly improved the image classification performance of deep neural networks trained on shared image datasets.

Scope

This standard describes how to generate datasets to train and test Machine-Learned Spectrum Awareness (MLSA) models that detect, classify, characterize, and/or identify radio frequency
(RF) signals and signal emitters. The scope of this standard includes:

  • methods for creating training and test datasets for MLSA models that are representative of real-world Dynamic Spectrum Access (DSA) and spectrum sharing scenarios,
  • methods for using data augmentation techniques to introduce sufficient sample variation so that the MLSA model can generalize to real-world scenarios,
  • methods for enhancing the training dataset with RF propagation channels and interference sources that are representative of real-world scenarios,
  • specifications for how to structure and store MLSA datasets,
  • methods for creating secure and performant MLSA models that operate on resource-constrained RF sensors and processors, and finally,
  • criteria for evaluating the performance of MLSA models.

Working group procedures

Working group documents

TBD

Contacts

Alex Lackpour (IEEE 1900.8 Working Group Chair)

Jesse Caulfield (IEEE 1900.8 WG Vice Chair)

Adnan Shahid, PhD (IEEE 1900.8 WG Secretary)