5G is a ‘brute force’ tech without machine learning – in the industrial sector, at least
The telecoms industry must get a handle on machine learning if it is to make advanced LTE and 5G communications feasible in industrial settings. This was the conclusion of a presentation by Cambridge Consultants at URLLC 2018 in London last week, an event focused on ultra-reliable, low-latency communications (URLLC), the most advanced strand of the new 5G system.
“There are amazing new technologies available today. 5G and IoT are two. These technologies will radically transform industries that make deep use of them,” said Derek Long, head of telecoms and mobile at the UK-headquartered consultancy, which is focused on making latest technologies feasible for commercial application.
The rising interest in the industrial sector from technology companies, including traditional IT firms, OT automation specialists, and, belatedly, a telecoms industry seeking use cases for 5G technologies. A glut of industrial IoT platforms are now available to industrial sectors, notably manufacturing, energy and utilities, and transport and logistics.
Industrial IoT for such applications as asset monitoring, predictive maintenance, and plant optimisation will make use of the full 5G system, as defined in the latest 5G NR specifications, comprising enhanced mobile broadband (eMBB) and massive machine type communications (mMTC), as well as URLLC.
The traffic generated on these new industrial IoT platforms, at the crossover of IT and OT operations, is “quite similar” to that in wide-area networking, noted Long. “It might be more granular, but it is similar – with large file downloads and large numbers of devices, or sensors, providing small amounts of data.”
The difference comes with the crucial control information in industrial set-ups, around robots and devices on the factory floor. This cannot be served by traditional wide-area networking, said Long. “The one-size-fits-all approach to the communications network is not the optimal solution.”
The safety and security of communications networks, and the data they carry and systems they control, is critical. To date, this has been resolved to an extent with the application of “brute force”, he said, through techniques like deterministic scheduling and routing, and by employing large numbers of high-powered cell sites with.
But such measures have been taken case-by-case, and have tended to require significant resources in terms of energy and expertise. “The question is how scalable that approach is, really? Because it takes lots of power and equipment, and the guarantee that’s given is a planned guarantee, rather than an actual guarantee,” said Long.
This is where machine learning comes in, particularly with the kind of big data volumes served up in URLLC scenarios. Long outlined six areas for machine learning in telecoms: in customer care, radio network optimisation, core network management, network security, operations management, and application support.
He focused on its application in the radio access (RAN) and core network at URLLC 2018. The complexity of the radio network, in particular, is spiralling out of control; performance parameters at 5G cell sites are in the “thousands”, he said, compared to just “hundreds” with old 2G stations.
The challenge for industrialists is to guarantee a “minimum level of quality of experience” in the network, rather than a “best-effort performance,” which is possible with machine learning techniques.
At the same time, the core network is under a “huge amount of disruption” from initiatives like network slicing to virtualisation. “There are a huge number of things intended to break up the core network and provide a more active amount of functionality,” said Long.
While network virtualisation increases complexity on an industrial level – as network software functions are “broken up among many vendors”, raising the need for cooperation, integration and testing – it also presents a “great opportunity to fine-tune the network” to the meet the specific requirements of the industrial sector, operative and application.
“It has to be done on an automatic basis for this to be intelligent. An example with WAN is to decide where to locate the application functionality –- on the device at the edge, or centrally in the cloud – and balance latency and throughput against cost,” said Long.
The point, in the end, is URLLC networking in industrial scenarios will provide the foundation for efficient machine learning implementations, and machine learning is the only way for industrialists to make sense of the chaos of URLLC in terms of control data.