Fast Prototyping of AI-Enhanced Solutions

Responsible: Luis Blanco

Team members: David López, Nikolaos Bartzoudis


Focused R&D area 1

  • Algorithmic design and experimental validation of AI-based optimization techniques for function placement and resource management in 5G and beyond wireless communications systems.
  • Design of new algorithms based on Machine Learning (ML) for the next generation of PHY and MAC layers, with special focus on link adaptation, interference management and radio resource allocation.
  • Research on AI-assisted sensor data analytics and predictive maintenance for IoT wireless systems.
    • Ongoing activities: ML for hardware Trojan detection in integrated circuits based on the near-field EM signature.



Focused R&D area 2

  • ML to solve complex multi-dimensional DPD problems, tune both PA auxiliary waveforms and analog parameters, and apply CFR, to meet the efficiency and linearity goals (new area, started 2019)



Related projects

Developed technologies

  • Remotely accessible Matlab-based experimental demonstrator for  multi-channel PA digital linearization based on artificial neural networks and considering linear and nonlinear crosscouplings.
  • Dataset dimensionality reduction for neural network digital predistorters to lessen the complexity and training time required to linearize wideband PAs with strong nonlinear effects.
    • Application: UHF microcell 5G-like access (downlink), Class-J GaN PA, 80 MHz intra-band contiguous carrier-aggregated FBMC signal.
      • Techniques: Indirect, noise-enhanced indirect, and direct learning architectures employing real-valued time delayed feedforward neural networks. Orthogonal matching pursuit (OMP) and principal components analysis (PCA) are employed for dataset basis reduction while consecutive and sparse batch selection mechanisms (CBS & SBS) are employed for dataset length reduction.

Targeted R&D effort

  • Neural network based linearization of multi-antenna transmitters and multi-band systems with dynamic-supply and dynamic-load modulation PAs (multidimensional DPD).
  • Expand dataset reduction and apply neural network pruning for multidimensional digital predistorters in beyond 5G transmitter architectures.
  • Explore new paradigms placing the DPD neural network training out of the radio unit box.

Selected references