AI-Powered Industrial Automation and Private 5G
Historically, private wireless networks operating in the unlicensed frequency bands had two main goals: to drive throughputs higher and lower costs. In today’s industrial environments, wireless connectivity requirements have grown more complex. With the emergence of AI-driven automation and mobile robotics, expectations have risen exponentially across key performance indicators (KPIs) such as throughput, latency, reliability and security.
This blog explores the transformative potential of artificial intelligence (AI) for non-public 5G radio networks. It delves into the emergence of AI-powered radio frequency (RF) sensing and signal processing, which is revolutionizing spectrum sharing and taking the performance of 5G-connected industrial automation systems to a whole new level. We also look at the major industry forces shaping next-generation standards and offer a vision for AI/ML in the 3GPP- and Open RAN-driven evolution towards 6G.
Unlocking 5G Growth
Despite recent doom and gloom headlines, global cellular network traffic growth is projected at 20% CAGR through 2029[1], driven by new AI applications, growing 5G penetration and emerging segments, such as fixed wireless access and Private 5G. The original promise of 5G–to create new enterprise demand for novel use cases built on ultra-reliable, low-latency broadband connectivity—is starting to show signs of life. While the GSMA has pegged the total carrier enterprise opportunity at $400B[2], the majority of current enterprise demand has been for Private (dedicated) 5G networks, which is already a $2.5-3B global market and growing at 21% annually[3].
Private 5G networks essentially fall into two broad categories – virtualized and dedicated, operating in licensed, unlicensed or shared radio spectrum as depicted in Figure 1.
Figure 1: Private 5G Approaches
With each flavor, whether a slice of a carrier’s network, a vRAN instance in the cloud or dedicated hardware on-site, there are tradeoffs from an enterprise’s perspective. The primary tradeoff is between the capital and operational models. Carrier-provided services offer capital offload. Dedicated networks enable more end-to-end control. More control, however, also means taking on more management of the network, but the ever-shrinking supply of uncongested radio spectrum poses new challenges and operational complexity.
Historically, private networks have relied on Wi-Fi systems operating in the unlicensed spectrum; however, new, autonomous mobile applications are pushing wireless networks to their limits. This is most evident in autonomous warehouses, factories, ports, and other industrial operations where wireless connectivity has become mission-critical. These massive facilities—often spanning up to 10 million square meters in size or more—drive millions of dollars in daily commerce and, as such, have migrated to 5G in search of more reliable and secure broadband connectivity.
The Future of Spectrum is Shared
Spectrum is the lifeblood of the wireless industry, where the “mid-band” represents a finite and valuable resource akin to beachfront property due to its large bandwidth and attractive propagation characteristics. Approximately 80% of the sub-7 GHz spectrum is already allocated or used for commercial purposes in various forms throughout the world[4]. If you consider all the spectrum between 600 MHz and 8.4 GHz, an estimated 98% of it is encumbered in many regions around the world[5]. As a result, spectrum-sharing is emerging as the final frontier in sub-8 GHz licensing, with the potential to unlock nearly 2,500 MHz of new bands for 5G and 6G, as shown in Figure 2.
Figure 2: Global Spectrum Landscape
The trend towards shared spectrum is just one of many challenges that must be addressed. One tenet of shared bands is incumbent user coexistence, in other words, protection of primary users, often military radar operators. Citizen’s Broadband Radio Service (CBRS) was the first mid-spectrum band that incorporated a sharing framework involving
- coexistence between incumbent radar users and secondary 4G/5G users,
- spectrum sensing (called environmental sensing capability or ESC), and
- tiered access, enforced through a centralized spectrum access system, or SAS
Because ESC uses a limited number of sensing sites spread across the U.S. coastal regions where most of the radar signals are, its detection resolution is coarse and relies on conservative propagation and interference models. This has created wide exclusion zones, resulting in very limited, mostly indoor private 5G use and woefully underutilized spectrum. While new dynamic protection and clutter models (known as CBRS 2.0) should improve the utility of the band, other emerging shared bands, including the 3.1-3.45 GHz in the U.S., may have radar devices operating in the air, sea, land or even space. For such new bands, spectrum sensing needs to become smarter and more pervasive.
Industrial AI and 5G Innovations
Beyond radar signal sensing, industrial environments present extremes in signal path loss and interference that constantly change as equipment and products move about the factory or warehouse. Additionally, spurious impairments, such as electromagnetic interference (EMI), can emanate from numerous unidentified sources. Without the ability to quickly detect, locate and react to these signals, the operational downtime caused by wireless network outages can cost businesses millions of dollars daily. As a result, industrial use cases, such as those shown in Figure 3, are driving the need for technology innovation at the intersection of AI and 5G.
Figure 3: Examples of Industrial 5G Applications
Around the same time AI was enabling machine vision breakthroughs, taking image classification accuracy to new heights, the founders of DeepSig recognized the power of AI for wireless signal processing and began the industry’s most fundamental R&D in this new field. When you look at a radio signal on a spectrum analyzer, it’s essentially an image, though not one that the average human readily recognizes. In a cellular radio, the received signal is converted into a digital stream known as in-phase and quadrature (IQ) data. Rather than making inferences on constellations of pixels, a wireless AI model is trained to detect and classify any signal it “sees” in these IQ data. With an accuracy of up to 99%[6], this is what the OmniSIG® Engine does. Figure 4 contains a screenshot of the OmniSIG UI (read more about OmniSIG).
Figure 4: Screenshot of OmniSIG Wireless Signal Classification
DeepSig did not stop there. AI models trained on reference signals (i.e., the wireless “beacons” transmitted by a cellular network) and on how those signals are affected by scattering, fading, Doppler shift and self-interference can accurately identify the precise “channel conditions” in which over-the-air (OTA) communications are taking place. By replacing conventional channel estimation algorithms with such a pre-trained AI model in the 5G physical layer, a cellular network can deliver higher user data rates in some of the most challenging cell edge or interference conditions. This was the first commercial use case for OmniPHY® 5G software (read more about OmniPHY 5G).
Why it Matters
DeepSig’s AI-driven spectrum sensing and signal processing solutions are transforming Private 5G networks by addressing some of their largest challenges:
- OmniSIG 5G Spectrum Sensing: Maximizes utilization of new spectrum enabled by 24/7, 360-degree spectrum awareness and dynamic spectrum sharing (DSS)
- OmniPHY 5G Signal Processing: Enables unprecedented link resilience in complex industrial environments through deep learning and AI-driven 5G signal processing enhancements.
OmniSIG for Private 5G
DeepSig’s AI-powered sensing models have been widely deployed in security and defense use cases for more than three years. An even larger opportunity is emerging in the 5G market, driven by the need for dynamic spectrum sharing or DSS. This promises to open up large swaths of new mid- and upper-mid-band spectrum globally for both carrier and dedicated private 5G use. While DSS may be the killer application, once integrated into the network, the technology will transform all aspects of spectrum management—from fault detection and threat analysis to interference mitigation and predictive maintenance. As a starting point, let’s look at what a private 5G pilot deployment might look like, as shown in Figure 5.
Figure 5: OmniSIG Industrial 5G Pilot System
OmniSIG AI spectrum sensing can handle virtually any signal or emission type in the electromagnetic spectrum. Its models are trained on a wide variety of signal types, including cellular, Wi-Fi, Bluetooth, IoT, EMI, and radar. Starting from the left side of the diagram, AI spectrum sensing can assist with RF planning. To characterize the RF environment, you can scan up to thousands of MHz of spectrum with an OmniSIG-enabled spectrum analyzer (like the Anritsu Field Master Pro shown in the picture). As unknown signals or anomalies are detected, more data can be collected, analyzed, labeled, trained and used to train new signal models, as shown in the center of Figure 5. Once deployed (right side of Figure 5), a continuous, always-on sensing capability can capture and record intermittent issues like electromagnetic interference (EMI), PIM (passive intermodulation), and other RF anomalies.
The OmniSIG inference engine runs on embedded, mobile and server-class compute devices, which can also function as an element manager for a trial network. It processes many Gbps of raw IQ sample data from software-defined radios (SDRs) and summarizes concise metadata representing the signal activity it sees, which can be reduced to kbps, allowing easy transport, storage, and event-based triggers. This data can augment RAN monitoring with a whole new set of KPIs, alerts and analytics and be utilized in real-time to trigger interference mitigation and radio resource optimization. The advantages and benefits are as follows:
Advantages:
Key Performance Indicator (KPI) | OmniSIG vs. Conventional |
Typical band characterization speed per 100 MHz | Milliseconds vs. minutes |
Classification accuracy | 90-99% vs. 70-80%[7] |
Signal detection sensitivity | -10 to -20 dB SNR vs. 0 to +5 dB SNR |
Interference handling | Advanced/learned vs. limited |
Time to train/deploy for new signals | Minutes/Hours vs. weeks or months |
Benefits:
With improved signal coverage/comprehension, interference mitigation and diagnostic speed, Private 5G operations can realize the following economic benefits:
- Faster mean time to diagnosis and repair of RF outage: $ opex savings
- Improve operations up-time: $ loss reduction
- Higher network & industrial asset utilization: better return on capital
- More wireless threats detected: reduced financial risk from breaches
OmniPHY for Private 5G
While AI extends across nearly all network layers, the physical layer—where users’ signals are transmitted and received over the air—remains the most challenging frontier despite garnering the majority of the industry’s R&D and capital expenditures. DeepSig is leveraging deep learning to solve major performance challenges in 5G and next-generation networks. We are proud to have announced the world’s first AI native 5G processing application with HTC in 2023—a neural receiver—for their G Reins Private 5G network-in-a-box. Earlier in 2024, we also announced the deployment of the world’s first commercial AI-RAN 5G neural receiver in a carrier network with Viettel in Vietnam.
The OmniPHY 5G neural receiver is a software module that functionally replaces a portion of the Distributed Unit software—with a neural network, as shown in Figure 6.
Figure 6: OmniPHY Neural Receiver Integration
The solution addresses key challenges that still plague many public and private 5G networks, namely low user throughput at the cell edge or in the presence of interference. The OmniSIG software module replaces traditional physical layer functions such as channel estimation, equalization, and interference rejection combining (IRC) and supports a variety of antenna configurations, including 4T4R, 8T8R, 32T32R massive MIMO and distributed MIMO. In low signal-to-noise-and-interference (SINR) conditions, the neural network can improve uplink throughput by up to 2-3X.
By leveraging ORAN standards and Intel’s FlexRAN platform, DeepSig has developed a software solution that can easily integrate with multiple private and public 5G O-DU vendors. OmniPHY is processor agnostic, and DeepSig anticipates integration with other RAN silicon platforms going forward. Today, the OmniPHY AI model is trained offline to recognize a variety of channel conditions and deployed as a generalized model across networks. Over time, the product is transitioning to location-optimized models, where training is being calibrated with real network data.
The path to deeper 5G network integration for AI-based layer 1 signal processing (i.e., OmniPHY) is more complex than for AI spectrum sensing (i.e., OmniSIG) as discussed earlier. As a result, the initial OmniPHY 5G implementation was designed to be fully compatible with and transparent to existing 3GPP air interface standards. As such, there are no changes to the 5G protocol stack or interoperability issues; OmniPHY simply replaces some existing functions with a neural network pre-trained to perform those functions better than human-designed algorithms in challenging environments like a busy industrial factory floor.
DeepSig is collaborating with industry and academic partners in the three major industry bodies that are shaping the future of AI in cellular networks, as shown in Figure 7. 3GPP develops the foundational end-to-end specs for next G standards and is developing a roadmap for AI integration from Release 19 onward. The ORAN Alliance has been focused on open architecture, RAN automation, and is now developing Generative AI-based channel models for its Next Generation Research Group (nGRG), especially surrounding the consideration of future RAN Digital Twins. The recently formed AI-RAN Alliance has a complementary charter to 3GPP and the ORAN Alliance. It aims to accelerate AI innovation in the RAN in whole new dimensions, including the convergence of AI and RAN infrastructure.
Figure 7: Major Industry Bodies Shaping AI-Cellular Network Integration
3GPP standardization anticipates incremental AI enhancements to beamforming, channel state prediction and overhead reduction and location accuracy starting in Release 19. The first significant AI enhancements to L1 are expected in Releases 20/21, the first to define 6G, which will kick off in early 2025. Don’t expect wholesale changes to the physical layer in the first iteration of 6G. More likely, you will see things like new AI-driven channel models, new reference signal designs and learned air interface features like constellation shaping. When combined as part of an AI-first, end-to-end design, such features will take wireless link performance to whole new levels, including 4 or 5 nines of reliability. As end-to-end concepts, you’ll most likely see these show up first in standalone greenfield use cases like indoor industrial wireless and non-terrestrial (satellite) networks.
Looking Forward
The impact of AI on wireless will grow as we approach 6G. To unlock enough spectrum for the capacity demands of the next decade, the majority of new allocations in the mid-bands will be shared. The required advances in DSS will push the boundaries of RF sensing to speeds, sensitivity and accuracy levels that can only be achieved by leveraging deep learning and network-wide deployments. As industrial applications demand more from wireless networks, AI-driven spectrum sensing and signal processing will be critical to enabling the reliability, security and efficiency to scale mission-critical wireless applications worldwide and delivering on the original promise of 5G.
[1] Ericsson Mobility Report, June 2024
[2] GSMA Intelligence sees $400B operator enterprise opportunity, Mobile World Live, Oct 8, 2024
[3] Sources include IDC, CCS Insight, other public sources and DeepSig estimates
[4] Considering 600 MHz to 7.25 GHz, not accounting for regional differences in allocation and use globally.
[5] Considering 600 MHz to 8.4 GHz not accounting for regional differences in allocation and use globally.
[6] Use case and data dependent. Source: DeepSig testing and benchmarking.
[7] As measured in short detection windows, e.g., milliseconds.