A complex-valued autoencoder neural network ca-pable of compressing & denoising radio frequency signals with arbitrary model scaling is proposed. Complex-valued time sam-ples received with various impairments are encoded into an embedding vector, then decoded back into complex-valued time samples. The embedding and the related latent space allow search, comparison, and clustering of signals. Traditional signal processing tasks like specific emitter identification, geolocation, or ambiguity estimation can utilize multiple compressed embed-dings simultaneously. This paper demonstrates an autoencoder implementation capable of compression by a factor of 64 that is still resilient against RF channel impairments. The autoencoder allows individual scaling by network depth, width, and resolution or in a compound sense to target both embedded and data center deployments. The common building block is inspired by the fused inverted residual block (Fused-MBConv), popularized by EfficientNetV2 & MobileNetV3, but with kernel sizes more appropriate for time-series signal processing. A complex-valued PyTorch implementation is available along with a pre-trained model, at https://github.com/the-aerospace-corporation/glaucus.
Glaucus: A Complex-Valued Radio Signal Autoencoder
2023-03-04
1783658 byte
Conference paper
Electronic Resource
English
Complex-Valued Radio Signal Loss for Neural Networks
IEEE | 2023
|Complex-Valued Signal Capsule Network for Automatic Modulation Classification
Springer Verlag | 2022
|Improvements of the traffic signal control by complex-valued Hopfield networks
Tema Archive | 2007
|Springer Verlag | 2025
|