One-stage blind source separation via a sparse autoencoder framework

Author

Jason Anthony Dabin, New Jersey Institute of Technology

Document Type

Dissertation

Date of Award

5-31-2022

Degree Name

Doctor of Philosophy in Electrical Engineering - (Ph.D.)

Department

Electrical and Computer Engineering

First Advisor

Alexander Haimovich

Second Advisor

Osvaldo Simeone

Third Advisor

Joerg Kliewer

Fourth Advisor

Ali Abdi

Fifth Advisor

Hongya Ge

Abstract

Blind source separation (BSS) is the process of recovering individual source transmissions from a received mixture of co-channel signals without a priori knowledge of the channel mixing matrix or transmitted source signals. The received co-channel composite signal is considered to be captured across an antenna array or sensor network and is assumed to contain sparse transmissions, as users are active and inactive aperiodically over time. An unsupervised machine learning approach using an artificial feedforward neural network sparse autoencoder with one hidden layer is formulated for blindly recovering the channel matrix and source activity of co-channel transmissions. The BSS sparse autoencoder provides one-stage learning using the receive signal data only, which solves for the channel matrix and signal sources simultaneously.

The recovered co-channel source signals are produced at the encoded output of the sparse autoencoder hidden layer. A complex-valued soft-threshold operator is used as the activation function at the hidden layer to preserve the ordered pairs of real and imaginary components. Once the weights of the sparse autoencoder are learned, the latent signals are recovered at the hidden layer without requiring any additional optimization steps. The generalization performance on future received data demonstrates the ability to recover signal transmissions on untrained data and outperform the two-stage BSS process.

Recommended Citation

Dabin, Jason Anthony, "One-stage blind source separation via a sparse autoencoder framework" (2022). Dissertations. 1600.
https://digitalcommons.njit.edu/dissertations/1600