Leveraging Amazon SageMaker to detect impairments in digital RF signals

2 minute read
Content level: Intermediate

Software Defined Radios are an important part of Satellite ground systems within Earth Observation and Satellite Communications. In a typical Radio Frequency system Spectrum Analyzers are heavily used to detect anomalies, however it can be difficult to troubleshoot signal impairments in the digital domain. This article demonstrates a software solution using the power of AWS machine learning to detect interference, phase-noise, compression issues with IQ (inphase & quadrature) constellations.

The modulation and coding schemes for common SDR algorithms such as DVB-S2X, can be represented in an IQ constellation diagram, with "I" being the amplitude of the in-phase signal and "Q" the amplitude of the 90-degree shifted signal. By varying the magnitude of I and Q we create different RF modulations eg QPSK (Quarter Phase Shift Keying), 8PSK etc.

The shape of these plots can indicate different types of signal impairments: - IQ constellations

A solution using Amazon SageMaker machine learning algorithms has been posted here.

It first uses K-Means Clustering to detect each of the blobs within the data sample, then Covariance Error Ellipse to attain metrics such as density, rotation, and ratio of major/minor axis.

Next, we use the Autogluon library to train a tabular classifier on the features which have been extracted during the preprocessing step. Autogluon will train multiple models based on the training data, and use the model with the highest performance and lowest inference latency.

Finally, we load the Autogluon model and run inference on sample IQ Constellation plots.

The final result is a classification of either Normal, Phase Noise, Compression, or Interference per IQ modulation. Those inference insights are published to an Amazon Simple Storage Service (S3) bucket.