PAGE CONTENTS
Objectives
The project “Demodulator supported by Artificial Neural Networks” aims to evaluate the benefit of an AI-enhanced demodulator compared to a traditional implementation. Our primary objectives include designing and testing an AI model for physical layer processing tasks, such as symbol soft de-mapping and channel decoding, taking as reference the CCSDS 131.2-B-2 standard. This involves a critical assessment of the state of the art and the definition of a new suitable topology for an AI model, as well as adequate training and validation strategies. The performance is then compared with a standard demodulator in terms of cost, computational complexity, power consumption, and bit error rate. The ultimate goal is porting the model into an FPGA-based architecture platform developed by IngeniArs S.r.l., namely GPU@SAT, to demonstrate the capabilities for onboard satellite applications.
Challenges
The following are the key challenges to be addressed:
- designing and fine-tuning an AI model for physical layer processing tasks like symbol soft de-mapping and channel decoding;
- generating a large and diverse dataset that accurately represents the possible different conditions of the communication channel to effectively train the AI model;
- integrating the AI model into GPU@SAT ecosystem, an FPGA-based soft-GPU platform through model optimization and quantization;
- measuring the performance and comparing it to a traditional demodulator while considering various metrics (cost, power consumption, BER, etc.).
System Architecture
The last phase of the project involves setting up and testing a hardware demonstrator based on the GPU@SAT IP core developed by IngeniArs. For this purpose, a software implementation of a CCSDS 131.2-B-2 transmitter and a channel noise emulator are used to generate the input signal. The receiver, on the other hand, is divided into two parts: the first, which addresses synchronization, is implemented in software; the second, addressing symbol soft de-mapping and decoding, is the AI model running on the soft-GPU platform.
Plan
The project starts with the definition of requirements and preliminary AI model topology, leading to the System Requirements Review. Next, the AI model design and dataset generation are planned, followed by the Preliminary Design Review. The project then proceeds with AI model implementation, performance assessment, and Critical Design Review. Upon achieving the required performance, the project then focuses on porting the model to the GPU@SAT platform, with associated testing and validation, leading to the Test Readiness Review. The final phase involves the critical assessment of the collected results and outlining a roadmap to the utilization of this technology in space applications, concluding with the Final Review.
Current Status
The generation of the dataset and its validation has been completed successfully. Some model architectures representing the CCSDS 131.2 de-mapper and decoder are under training.
