-
StatusOngoing
-
Activity Code7B.079
The goal of the activity is to develop an intelligent AI model that improves useful data rate by enabling less complex and less power consuming coding scheme with respect to current implementations. The developed algorithms can compress data using AI by reducing data replicas that are intrinsically present in the information to be transferred and mitigate possible reception errors by exploiting this intrinsic redundancy.
The applications are tested using a demonstrator testbed, showing the benefit of AI-supported data compression, comparing the obtained results with a state-of-the-art communication standard.
The testbed is composed by a Software Defined Radio (SDR) that simulates the CCSDS 131.2-B data link layer. The SDR takes input from the application layer, which is represented by the AI algorithms.
Such structure allows to test the end-to-end transmission chain, evaluating also the benefit of AI smart data compression models against different channel conditions and different use-cases or data types.
Finally, IngeniArs performs a feasibility study for deploying the AI-compressor into the GPU@SAT hardware accelerator.
GPU@SAT is a technology independent GPU soft-core developed by Ingeniars that can be embedded in space-qualified FPGA and used for space applications, including high-reliability space missions (class 1).
The biggest challenges for the MLDR project are:
- AI-compressors: develop an AI model that is able to compress and decompress data while maintaining the best quality possible is a very challenging task. In fact, such AI models deal with different data types and noise conditions requiring a peculiar training phase that create a well-formed and quantised latent space.
- End-to-end communication simulator: build an end-to-end communication chain based on CCSDS 131.2B could require more than 1 year. IngeniArs has a long experience with such type of communication systems and already disposes of a complete CCSDS 131.2-B SDR, limiting the risks related to the simulation environment.
In the first approximation using AI-models improve the useful data rate from 11%, in case of a transfer frame of 1024-byte to 47% in case of a transfer frame of 223-byte.
Moreover, as shown from the results obtained by IngeniArs, an average data compression of at least 30:1 is foreseen, obtaining more than 40% of improvement by using the intelligent compressor.
Another peculiarity of the AEs is the fixed output size.
Standard algorithms, such as CCSDS 123 or CCSDS 122, compress data by exploiting “recurrent patterns” within data itself. The highest the recurrency is, the better the compression.
On the contrary, AI compression systems always compress data by outputting an array of the size of the latent space regardless of the input.
The compression level can be selected during the AI-model design, making it customizable starting from the training phase.
Models with multiple compression levels can be obtained via retraining of some layers, or in case of additional feature, via fine tuning or transfer learning.
The performance of the developed models will not depend on the data nature.
The AI models and consequently the testbed is evaluated against such metrics:
- Compression rate: Indicates how much the model reduces input data size, with higher compression resulting in smaller latent representations.
- Processing time: Indicates how quickly the model can generate results, crucial for live-streaming tasks.
- Reconstruction quality: Assesses how accurately the model reconstructs data, with higher scores for models that retain fine details.
- Number of parameters: Reflects the complexity of the model and computational cost.
- Power consumption: Indicates the ratio between the power consumption with respect to the original power consumption 𝑃𝑐= |𝑃𝑓−𝑃𝑠|𝑃𝑠∗100
- Generalisation capability: Measures the performance of the model on unseen data. This capability will be assessed similarly to reconstruction quality but using unseen data that was not part of the training process.
IngeniArs adopts the CCSDS 131.2-B simulator supported by GPU, which represents a fully functional end-to-end communication system.
Supporting the simulator with the AI-compression algorithm, IngeniArs can easily derive the metrics while computing the error correction rate. The errors are due to channel impairments (AWGN, doppler error, frequency error, timing error, etc.) introduced by the CCSDS 131.2-B simulator.
The two algorithms run on a dedicated computer with GPUs, which can accelerate both the CCSDS 131.2-B data link and VAE model.
N/A
The project has already passed the SRR, and now IngeniArs is working on the preparation of the datasets as well as a preliminary selection of possible models to be adopted as reference for the final implementation. Concurrently, IngeniArs is preliminary developing the testbed. The latter is composed by three different steps:
- Encoding phase: executed by one of the AI models developed selected with respect to the type of data involved in the use-case
- CCSDS 131.2-B end-to-end simulator
- Decoding phase: executed by the decoder of the AI models used for generating the latent space.
To speed up the testing of the AI models, IngeniArs is adopting a CI/CD strategy, which allows to integrate and develop different models in short time.
PDR is forecasted for Q2 2025.
