Adding deeper machine learning to the RAN

DeepSig targets open vRAN, waits for Massive MIMO acceleration.

DeepSig, a company that developed advanced RF monitoring and sensing software for test and measurement applications, says that it can apply its algorithms to enhance “upper Layer One” RAN functions, achieving superior baseband processing performance.

It is launching OmniPHY-5G in 2021, and expects to have software in its customers’ labs and products before the end of the year. It says that could provide a performance boost to 5G NR equipment, and it especially sees application within Open RAN platforms that some say cannot currently match the performance of integrated systems provided by “legacy” vendors.

DeepSig says that it uses machine learning to apply artificial intelligence to radio systems, benefitting L1 and system performance beyond traditional approaches. 

Its publicity says, “Systems are learned from wireless channel measurements and directly optimised for real-world hardware and channel effects using end-to-end performance metrics and feedback.

“DeepSig’s revolutionary approach to RF signal processing applies machine learning to time-series radio samples and channel measurements, allowing it to learn from data. By creating algorithms that learn signals and effects from the I/Q representation, DeepSig’s systems achieve better performance than traditional simplified analytic models or feature-based methods.”

It says it will create reference implementations to allow vRAN and other 5G-NR base station integrators to understand and quantify the value of embracing the data-centric RAN algorithms provided by OmniPHY-5G

David Oberholzer, VP Business Development, told TMN that the company sees opportunity in the open vRAN space, attracting vendors in that area with its promise of improved performance and spectral efficiency: “Physical layer OmniPHY 5G has better algorithms in the upper Layer One than the standard 3GPP algorithms. We are targeting the Open RAN community first because that allows us to go to multiple partners with the same implementation. 

“We haven’t found anyone yet who’s actually doing deep learning enhancements in that upper L1. We find people doing AI work in areas such as scheduling and beamforming and of course a lot with network management operations – analytics and SON – the kind of things that might become good apps or xApps on a RIC platform. But we haven’t found someone else yet that’s doing specifically what we’re doing in having deep learning based estimation that reduces the computational load of a RAN.”

Oberholzer added:5G is very complex, computationally, especially when you look down at the DU where you’re adding all those radio elements even for standard MIMO. But getting up into massive MIMO it’s a huge compute load. So when we can maybe cut that in half, that’s a real opex and capex benefit to an operator. And we’re doing it by dropping in a set of code into the existing software architecture, it’s not a rip and replace.”

DeepSig’s claim is that via machine learning it can handle randomness and real world data better than 3GPP standard algorithms. Oberholzer cannot state which chip companies DeepSig has been working with to prove out its code, but it includes major OpenRAN and vRAN ecosystem players, and he says that this work has demonstrated meaningful improvements in that upper L1 channel estimation. That results in benefits such as reduced computational load, with the knock-ons of reduced power and hardware demands, and also better cell edge performance.

Oberholzer said that integrating DeepSig’s technology could enhance the competitiveness of vRAN providers, as well as their operator customers.

“I can’t tell you where it came from, but we had a technology planning team at an operator that said, ‘If you can improve my cell edge link budget by 0.7dB, we’re very interested in talking to you’. We’re just in our standard mode and are testing it to a 2dB improvement. So when you bring that kind of benefit we can then take that to a vendor and say, ‘Hey, there’s an operator over there that would find your software very attractive if you had this feature, if you had this capability.’”

Oberholzer said that the company had expected massive MIMO to be adopted quicker than it has been, so did not grow business as quickly as it expected in 2020. Adoption of m-MIMO, where DeepSig thinks it can show the most benefit, is about a year behind the company’s expectations.

The company currently has 30 people on staff, featuring a founding leadership with background in military communications applications. It raised $10 million in a funding round in February 2020.

You can access its scientific publications on AI in communications systems here.